2012
11th Jan 2012 Acton response
The response part of the lowest level actons represent something that is done. Paying attention is done when performing action habits. Should the Where part of the goal be incorporated into the response part of the lowest level acton? Or should it be the Where part of the trigger in the second acton of a sequence? And the second acton response part would not have a Where part? When being done subconsciously the only places where stimuli must be obtained are represented in the trigger stimuli. When a subconscious action habit completes it performs the last response and does not wait for or attend to any goal stimulus. This would imply the answer is No. The Where part of the goal stays in the goal stimulus and is only used when practicing.
Goal Interest
The interest that drives the action habits evolves as follows. When the goal is novel we want to do the action habit to get the goal. We give the trigger expected interesting. When the goal is neutral or uninteresting we have no interest in doing the habit. If it is uninteresting does the trigger get expected uninteresting? When the goal is expected interesting we think about it. We then recall the next action habit. We do it or not depending on its most recent goal interest. But with a neutral goal the action habit can still have a redo interest because its sequence was novel. This also can cause us to do / practice the habit.
12th Jan 2012 Distraction versus Unexpected
We could be practicing an A-Habit and get an unexpected stimulus but at the same where location. For example the goal expected is a red apple but we get a green one instead. Or we can be distracted by a stimulus on a different sense, e.g. a buzzer. Is there a difference when processing these two situations? The unexpected just gets a different what value for the goal while the distraction provides a whole new where and what. Both of them cause you to update the action habit. But if the distraction has an interest level above the concentration level it is not used as the goal of the action habit being practiced. The action being practiced is interrupted and not updated. It is still waiting to be practiced.
However, if we are performing a reflexive response, trying something for the first time because we are trying to avoid a repeat, then the next most interesting stimulus will become the expected goal of the action habit formed.
13th Jan 2012 Practice versus doing
I've realized that practice is done when there is a redo-interest in the action habit while a habit is started for doing (consciously or subconsciously) when the goal is interesting or expected to be interesting. If there is both a redo-interest and a goal interest then the practice must be done first to learn the habit.
When practicing a habit one does not feel success or failure because it is not the goal that is wanted. On failure one just has a new action habit to practice. On success the action habit is ready for use based on the desirability of the goal. One only gets success or failure when a learnt action habit is being done and the goal is achieved or not.
It is also possible that thinking is only based on the goal's interest, not on the redo-interest.
14th Jan 2012 Redo-Interest
I realized that the action habit's redo-interest for practicing purposes is an interest property of the A-Habit even for A-Habits that are orienting / attention paying only. These are the habits that correspond to perceiving the S-Habits for a dynamic [?]. STMs continue to recognize and change the interest of the S-Habits they perceive. This solves the problem that I was having in which the conscious STM was recreating the sequence that a sense's STM was creating and thus its interest becoming neutral immediately.
25th Jan 2012 Practice and Doing
Is it possible that if one is concentrating on what one is doing then one is practicing it even though it may already be learnt?
Reaching Goals
From the 31st of July 2011 we have an acton structure in which each acton has a left and right sub-acton. When the trigger of an active acton (master) is recognized it enables / activates its two sub-actons (slaves) simultaneously. Every active acton is always looking to see if its trigger stimulus has occurred. If the left slave returns the master reactivates it so it will repeat if its trigger repeats. If the right slave returns and the left slave has returned at least once then the master returns. If the right slave returns and the left slave has not returned then the right slave has recognized its 2nd part prematurely. Does this mean that each active acton needs to know if it is the 1st or 2nd and if it should operate in parallel or sequentially? No we can stop it performing actions prematurely by making its trigger contain the entire sequence it needs before doing its actons. This eliminates the need to check if the 1st slave has returned before the second.
So the acton structure for recognizing the sequence A, B on sense #1 would look like:
A + a1 + a2
/ \
A + _ A, B + _ where the _ stands for an orient response
Even though the two slaves are enabled in parallel the second waits to recognize A, B before reacting. If they were to execute in parallel then the seconds trigger would not contain the first's trigger as part. It would be recognizing A ^ B where A and B come from different sources and look like:
A + a1 + a2
/ \
A + _ B + _
The sequence A, B, A on the same sense would look like:
A + a1 + a2
/ \
A + a3 + a4 A, B + a5 + a6
/ \ / \
A + _ A, B + _ A, B + _ A, B, A + _
Disabling active actons
From the 14th Nov 2011 I still need to come up with a practical strategy for disabling an active habit when it does not get its expected trigger. Could it occur when the action habit hierarchy fails to progress at least one step each cycle? That is, none of its slave actons get done in a cycle. This should work because S-Habit performance is also being represented in the actons.
Parallel configurations
Even though I am performing an action habit to do something the rest of my muscles are locked in position because they are not involved in what I am doing. Thus at any one time I am performing many action habits in parallel that are freezing my muscles in position to maintain their status. These freezing / stationary action habits are not part of the active one that I am practicing. E.g. holding my head erect carries on independently in parallel while I type with my fingers. Surely holding my head erect is not part of my typing habit. But then again maybe it is because I have to position my eyes appropriately. How about where my legs and feet are positioned are independent of my typing habit and are not parts of it but must be done in parallel.
27th Jan 2012 Practice and Doing
The answer to the 25th Jan 2012 is No! The test for this is to determine if practice mode is any different from just concentrating on what one is doing. One is concentrating on what one is doing when one is repeating a learnt action habit with a desirable goal. A characteristic of practice is that one repeats the exact same sequence to learn the habit. Nothing else is tried during the practice. A characteristic of just doing is that in the middle of repeating the sequence if one thinks of a better approach one tries it. Thus there is a different level of concentration between practice and doing. Doing is all about reaching a desirable goal and the action is started and performed sub-consciously. The conscious concentration level is then set at neutral (not at the goal interest level) so that thinking can take place and cause variations in the action sequence to be tried.
28th Jan 2012 Practicing
When we try a new action sequence we are not only practicing the sequence but also expecting the goal. The goal is also practiced. Or in other words we stay in practice / concentration mode until the expected goal is reached, or not. So a reflexive response includes the practice of the unwanted goal that is a repeat of the trigger.
31st Jan 2012 Goal Recognition
I'm still trying to clarify the correct strategy for goal recognition at the end of practicing an action when the goal is a sequence. It was discussed on 30th April 2011 and 17th Sept. 2011. Also check out 19th Dec 2011.
If we consider the situation in which the goal is not a sequence the strategy is clear. Given an action habit with trigger T, Action h and Goal G that needs to be redone it will be practiced when T is experienced. After the h has been done we will be concentrating on the Where of G hoping to get the G. If we get the G then the T h G action habit is learnt. If we get B instead we form a new action habit T h B that needs practicing. In the case where h is a reflexive response to T then we are practicing and the expected goal is a repeat of T that we do not want. If T reoccurs then we have learnt about the repeat (the T h T action habit will be formed and be waiting for verification through practice) and T will not be an interesting goal to achieve. If B occurs instead of the repeated T then we form a new action habit T h B that needs practicing. In all cases we have an action habit, either we create a new one or update the existing one.
The same should apply when the goal is a sequence. If the sequence matches in length and content we have the expected goal and we use it. If the sequence fails to match we must detect this before the end of the sequence. This is first noticed as soon as the first stimulus does not match. We could be expecting ABCD and get ABEF instead. The shortest stimulus that does not match is ABE. The execution of the attention action fails at this point. The attention action is not just following the Where sequence but also matching the stimulus values. Thus ABE would seem to be the new goal stimulus for the action habit.
1st Feb 2012 Goal Recognition
I can't use the acton structure to perform the S-Habit recognition required after a response has been made because at any one time a particular S-Habit may be in the process of being recognized by many different action habits. The one being practiced could be waiting for its S-Habit goal to complete while a subconscious action habit is also waiting for a subset of the same S-Habit to be recognized. Though I still need a mechanism which is being done that is tracking the S-Habit recognition, returns completed or failed with the recognized S-Habit.
A possible mechanism is that each active acton that has a trigger that it is trying to recognize keeps track of its own progress. It would have to have a 'waiting for S-Habit' value which it updates each cycle. If it does not get what it is waiting for it fails and we can find what occurred instead. If it does get what it is waiting for it either updates it to the next longer S-habit or it matches the acton's trigger and the acton executes.
3rd Feb 2012 Sequence Formation
I need to reanalyse the process of combining novel and familiar stimuli in sequences. We have the principle that novel stimuli are combined into the longest novel sequence. Also familiar sequences are recognized as objects. Or more exactly a sequence of NNNN where N represents a novel stimulus forms a novel S-Level = 4 S-Habit. And as long as we continue to get additional novel stimuli we form a bigger sequential object. Similarly as long as the stimuli in sequence are familiar (represented as F) and their combinations are familiar we recognize known S-Habits. FFFF forms an S-Level = 4 familiar S-Habit. And this has resulted in the idea that novel stimuli get combined with novel and familiar with familiar stimuli. Read 3rd Sept 2011. In tree structure these form:
N F
NN FF
NNN FFF
NNNN FFFF
Now consider what to do with the sequence FFNN. Assuming the two Fs form a familiar pair we get the separation of the sequence at the edge between the familiar and novel. This is:
F N
FFNN
And similarly the novel sequence terminates as soon as a familiar level 1 stimulus occurs.
N F
NNFF
But what if the novelty occurs at a higher S-level based on familiar ones below.
F
FF
FFF N
FFFF F
FFFFF F
In this situation the novelty is found at level 3 after combining the 6th stimulus with previous ones. The pair of the 5th and 6th stimulus is familiar. Should the strategy be to separate out the last stimulus and use it to start the next sequence as in?
F
FF
FFF N
FFFF F
FFFFF F + +
Or should the novel one be split out leaving a reduced previous stimulus as in?
F N
FF FF
FFF FFF
And what happens if the next stimulus is familiar but the combination at S-Level = 2 is novel as in?
F N
FF FF N
FFF FFF F N
Obviously the last novel stimulus at level 1 separates from whatever came before. The STM could use a mode variable that is set such that it is processing a novel sequence or a familiar one and the mode is switched as it is processing the stimuli. Given the sequence:
N
NN
NNN
FFFFF
FFFFFF
It would start in the familiar mode and switch to the novel mode and stay in it after the 3rd stimulus.
What happens at a level higher than 1 when we are processing a novel sequence and we get a familiar stimulus?
N F
F F F
Should the level 2 familiar be formed or recognized or should it be discarded and continue as?
N
F F F + +
The transition from N to F or F to N at any level signals an edge in time between two sequences. And similar to P-Habits the edge should be used to signal the boundary between two familiar objects or a novel one and a familiar one. In this pattern:
F F
FF FF
FFF N FFF
FFFF FFFF
The novel combo of the 4th and 5th stimulus signals the edge between two familiar sequences. And in this pattern:
N
N N F
NN N FF
NNN N FFF
F F F F FFFF
We see an edge between a novel sequence and a known one.
So I think the strategy of only combining novel with novel and familiar with familiar at the same level will give me what I want.
Goal Recognition
The above approach changes the goal recognition strategy because the goal expected might start out matching the goal sequence obtained but then the sequence terminates before the end of the expected sequence because a novel stimulus occurs. We need the concept of the goal obtained so far that is used as the goal when this happens. It would not contain the stimulus that started the novel sequence and deviated from the goal expected. And if there were no goal obtained so far, would the goal be the next longest novel or familiar stimulus?
4th Feb 2012 Reflexive Response
When doing a reflexive response the strategy for the expected goal should be whatever comes next that is a complete familiar sequence (longest) or the longest novel sequence. This replaces the strategy of expecting not to get the trigger repeated. This approach would use the goal so far idea that ends when the sequence is uniquely identifiable. Thus if the goal is the Happy Birthday song then as soon as it is recognized it becomes the goal stimulus. Do we have to wait to the end of the song before we update the action habit if we can uniquely identify the goal after only a few of its stimuli? If we only have one sense then maybe the answer is yes.
I have implemented a STM mode that is recognizing a familiar sequence or novel one and a flag for when there is an edge in time caused by the change from novel to familiar or vice versa at the same S-Level. I believe I also need a mode in which we are waiting for a complete stimulus sequence before we use a stimulus as a trigger or goal. In this mode the stimulus sequence is accumulated as the stimulus so far and does not get used consciously until an edge or a repeat occurs. But it causes a problem for triggers because the 1st stimulus of the goal has to be obtained and have caused an edge for us to recognize the trigger. It works for recognizing a goal and updating an action habit being practiced because we can wait for the edge-causing stimulus to terminate the goal. But the goal cannot be used as the next trigger because the edge-causing stimulus has occurred. Maybe it might work if we also keep a most recent response and form the action habit using it.
5th Feb 2012 Conscious / Actionable Stimuli
In this morning's test run of AB repeating, I have the RecentTrigger accumulating a sequence of novel stimuli until an edge or repeat occurs. Then the trigger and the reflexive orient response are made into a practice acton and we start accumulating the goal in GoalSoFar. The goal ends when an edge or repeat occurs. Then the whole action habit is created. If it's an edge the goal is not available as another trigger because the edge-creating stimulus has occurred. However if a learnt action habit gets its expected goal then this is available as trigger in the same cycle. What I need STM to do is to maintain the stimulus so far and then flag attention that it is complete. Only a complete stimulus would then be used as a goal or trigger.
STM would be either:
- Building up a novel sequence (no next stimulus yet)
- Building up a familiar sequence (no next stimulus yet)
- Have a complete novel sequence & a next familiar stimulus
- Have a complete familiar sequence & a next novel stimulus
- Have a complete familiar repeated sequence (no next stimulus yet)
However, large familiar sequences may contain shorter sub-sequences that have associated action habits. These would not be permanent. These shorter sequences must be consciously attended to in case they trigger useful action habits. Thus a 6th possibility is:
- Have an actionable familiar sequence (no next stimulus yet)
11th Feb 2012 Conscious STM
If I create a conscious STM that contains the conscious stimuli it can be used to detect repetitions. To do this, combinations of conscious stimuli must be produced. These combinations could be triggers when they are seen to repeat. They could be goals and they would be recognized based on the same sequence of conscious stimuli. Conscious STM would only be processed each time there is an attended to conscious stimulus. Would I still only combine novel ones with novel ones and familiar with familiar? Would the edge between novel and familiar signal the end of a trigger or goal? A sequence of two novels would be from different senses because they would have been already combined if from the same sense. The same applies to two familiars.
13th Feb 2012 STMs
I've decided that the per-sense STMs just combine sub-conscious stimuli until they recognize a conscious stimulus sequence or a repeat occurs. They do not combine two conscious stimuli. This is the job of the conscious STM or may only happen after they have become permanent. But what is it that recognizes an unexpected sequence of conscious stimuli. This must be the subconscious S-Habits (action habits with orient responses). So of the list of what the per-sense STMs do that is above we have just 1, 3, 5, and 6. It might make sense to make all sub-conscious stimuli novel until they become conscious. Only then can they become familiar.
14th Feb 2012 Sequential edges
No. I should not make all subconscious stimuli novel until they become conscious. In this case they would only become conscious upon repeat. There would be no recognizable edges between novel and familiar stimuli. A sequence would have to repeat first to become conscious.
17th Feb 2012 Device Types
Back on 23rd July 2011 I concluded there were two fundamental types of devices, discrete and graduated; based on the nature of the responses they were sent. For discrete devices they are given symbolic values that represent a command for an action that the device is to do. Examples are rollover, jump, sit, move left, move right etc. The device must have some built in intelligence that converts the command into action. For graduated devices the value provided could be interpreted by the device as either an absolute goal to be achieved or an amount of change to make. In the former case the device must have some built in intelligence that converts the graduated value into an appropriate action to reach the goal setting. For example, a motor could be given the new angle position to which it must move. In the latter case the device just makes the change requested not knowing where it currently is.
18th Feb 2012 Permanent stimuli
I have a problem when a high-level stimulus turns permanent and no longer attracts attention. Attention is then paid to lower level combinations that are part of the permanent stimulus. I still need to be able to pay directed attention to these lower level parts but not have attention attracted by them. I think the strategy is to mark all the subparts of a permanent stimulus also permanent. Permanent means that we no longer react to them, which more accurately means, they never attract attention again. However directed attention can still find them.
19th Feb 2012 Parallel Configurations
I've started to use the word "configuration" more often recently when talking about stimulus patterns that contain stimuli that occur simultaneously but are independent of each other. A P-Habit stimulus is a good example because the senses are independent of each other. A configuration consists of one or more stimuli each from one or more sources (independent sensors) / senses. Thus there are no edges involved in forming the configuration patterns. The same concept should be used to describe the configuration of a contrast pattern, a relative size (shape) pattern, separation pattern (which is the relative size pattern including gaps), orientation pattern and negative pattern for an object found from a graduated sense.
Yesterday my problem with P-Habit configurations was that A^B^C became permanent, A^B and B^C which are part of it were flagged as used in this configuration but A^C was not so flagged. A^C then became available to attract attention as the largest familiar stimulus. One solution is to continue making A^B^C available for attention even though it has become permanent. Another solution is to mark as permanent all the possible configurations of the lowest level parts in a higher level permanent configuration. But this is a non-trivial algorithm. There must be a simpler way. I could mark all the lowest level parts as needing to become permanent and then mark them permanent as they are combined the next time they are input. This also is fairly complex. Another approach is to not generate the A^C in the first place because of its use in the other two. But the A^C must be generated if I am to recognize it as the attended to part in AXC. At the same time I must consider that only stimuli with the same familiarity get combined. [Read 14th March 2011, 20th March 2011 and 14th Sept 2011]. I think the best solution is to make the A^B^C available for attention.
20th Feb 2012 Parallel Independent Configurations
The note of 14th March 2011 gives the reasons for wanting to not be conscious of configurations of parallel stimuli that are interdependent. The idea being that a stimulus or configuration of stimuli that are only found in one higher level configuration should not be independently available for attention. I've been trying to keep track of dependence between two or more sensors at the sensor and value level. May be I should keep track only at the sensor level and not include the value. This would mean the sensors involved in a configuration were locked together reading their values from the same object. This idea would mean the objective is to find out if the sensors are dependent or independent. Once they are found to be independent, even if it just for a single configuration out of many they would then remain independent. No, I think this is too general. Two sensors may be locked together just for certain value combinations but not others. For example I can feel things and hear them independently but when I sneeze I never hear it without feeling it or vice-a-versa. This is an example of a two-way lock. What about a one-way lock? This is the case with the ABX example on the 15th March 2011. AB always comes with the X but not vice-a-versa. AB is only found together in the ABX configuration.
I've been working on a possible strategy based on the ABX example. In the following analysis, c stands for conscious, s stands for sub-conscious, n stands for novel and f stands for familiar. On the left is the series of stimuli from 3 independent symbolic sensors. On the right are the single, pairs and triplets of stimulus configurations recognized. The stimuli are the upper case letters.
Stimuli singles pairs triplets
1 nA nB nX nAs nBs nXs nAXs nABs nBXs nABXc
2 fA nC fX fAs nCs fXs fAXc
3 nD fB fX nDs fBs fXs fBXc
4 fD fC fX fDs fCs fXs nDCs nCXs nDXs nDCXc
5 fA fC nY fAc fCc nYs nACc
6 fD fB fY fDc fBc fYc nDBs nBYs nDYs nDBYc
7 fD fC fY fDc fCc fYc fDCc fDYc nCYs nDCYc
8 fA fB fX fAc fBc fXs fAXc fABs fBXc fABXc
9 fA fB fY fAc fBc fYc nAYs fABc fBYc nABYc
If rows 6 and 7 were reversed
6 fD fC fY fDCc nDYs nCYs nDCYc
7 fD fB fY nDBs nBYs fDYc nDBYc
It appears that when the top level / largest combination is novel and set to conscious then the ones below that were familiar are changed to conscious. When the top level is familiar and changed to conscious there is no change to the subconscious ones below. Note at 5 AC is novel because it was not formed at 2 because A was familiar and C was novel.
21st Feb 2012 Parallel Independence
The objective in keeping certain configurations subconscious is that they are not available for sequential recognition because they are always part of one larger configuration. This means that my thinking about the above steps 1 to 2 is wrong. The B has changed to C. C is novel. It should be the part that attracts attention. Thus C should be conscious. This gets back to the idea that it is the biggest novel configuration that attracts attention. This would imply that the D in step 3 attracts attention. But the C has also changed to a B. In fact the AC has been replaced with the DB. Shouldn't the DB attract attention? When going from step 3 to 4 only the B changes to a C. Should just the C attract attention? Note that in 3 the DX pair was not formed because of the combine same familiarity rule and thus in step 4 DX is novel. Does it make sense to use the combine same familiarity rule with independent sensors / senses? The rule tends to advance the recognition of known / familiar combinations and delay the recognition of novel ones. Maybe it only works when there are edges to deal with. What we really want from independent stimuli is to be able to take advantage of anywhere there is a dependency and not have to explore the sub-parts of a configuration when they are always found in the same combination. But we also want the biggest novel combination to attract attention.
22nd Feb 2012 Parallel configurations
Stimulus signals in the brain travel from neuron to neuron through existing synapses faster than neurons have a chance to form new synapses. This results in the familiar being recognized before the novel. Maybe what I should do at each level is test the existence of a combination but not form it if it is novel. Then after forming all the familiar configurations at that level then form any new ones from the unused configurations. Would this delay the novelty formation a little more than just combining novel with novel and familiar with familiar? I think it would stop me from combining two familiars to form a novel. This is what is happening when I come across an edge in time. The novel combination stops the formation of any higher level combinations. But at some point two familiars that combine to form a novel must happen. If they start occurring more frequently together then they start being recognized as a single object. Maybe what I need is the same concept as in the action habits. The first time they occur they form a pair and are flagged as novel but not to attract attention but to represent something that needs to be re-tested.
Edges
Millions of edges occur per second, all the time between objects in our visual perception but we never form pairs of the two stimuli that are on either side of the edge. We recognize it as a novel pair and thus an edge. We use this knowledge of the edge to recognize the boundary of the objects on either side but do not retain the edge combination. If the edge combination continues to occur we still don't recognize it as an object. The familiar combination of parts forms first and as long as it continues to stay together in the same configuration it is never subdivided into its parts. We then have to consciously recognize when two familiar parts are occurring together more frequently than by chance.
Solution to P-Habit recognition
The algorithm needs to go through the level it has just combined and mark any stimuli that were not used in any combinations as conscious for attention purposes. If there are any novel ones then the biggest attracts attention. Two familiar stimuli that would form a novel pair are not combined and no pair is created. This is because the process starts by assuming the biggest combination of novel stimuli are all interdependent. It then subdivides this into smaller conscious configurations when these smaller configurations occur with some other smaller configuration. Thus larger familiar configurations are only broken up into parts when one part exists without another part.
24th Feb 2012 Parallel Configurations
I have solved the problem from 19th Feb 2012 by first assuming some fixed order to the sensors / senses and then producing binons whose two source parts are separated by a shorter distance before producing those with parts that are separated by longer distances. Thus at any one level in the P-habit creation I first combine adjacent pairs, then binons that are two apart and then three apart etc. Also the two source binons must either be unused or be used at the current level. This approach results in A^C, which is two apart, not being created because the A and C are used up in the 1 apart separation process at the 1st level. However they will be used at the 1st level when AXC occurs because the AX and XC will not be formed leaving the A and C unused when combining 2 apart source binons. These two pairs will not be formed either because X is novel while A and C are familiar or because they are all familiar but the AX and XC do not already exist.
25th Feb 2012 Conscious stimuli
I have realized I need to form P-Habits and S-Habits of conscious stimuli. If I get two conscious stimuli at the same time they should be combined like they are subconsciously. The conscious STM will then join conscious stimuli into sequences.
Repetition
I've been assuming that when the J repeats in the sequence AJJ that I have an A,J sequence and a J. Maybe it should be an A and a J that repeats. That is the edge in time should be between the A and the J once it is known that J is a repeater. Similarly if I had ABXYXY I should not get ABXY followed by an XY but get an AB followed by an XY that also repeats. The ABXY is actually recognized as a separate part when the second X occurs and not when the XY repeats. By this time the A-Habit with ABXY as trigger and no goal expected has already started.
28th Feb 2012 Width and Separation
On the 14th April 2010 I concluded that the width pattern and separation pattern for lines was exactly the same. I need to create gaps and use them between lines that are not adjacent. When the widths of the gaps are included in the width pattern then I do not need a separation pattern as well. However for uniformity when the lines with gaps between them come so close together that the lines are adjacent I should place gaps of width zero between adjacent lines. This would maintain the pattern for an object at the same level of complexity.
This leads to the fact that an object recognized on a sense with graduated readings and dependent / graduated sensors is a P-Habit combination of; contrast pattern, width pattern (shape), contrast orientation, negative and shape orientation (rotation). All of these object properties are independent of each other. This idea was previously mentioned on 14th Feb 2011.
Levels of Complexity
Contrast and width patterns are based on relative values. But how does a pattern get re-used at a higher level of complexity? For example we have intensity reading of 3 lines of 1,2,5. This forms two binons at level-2 that are 1,2 and 2,5. At level-3 the contrast pattern binon is (1,2),(2,5). Now we observe 6 lines with intensities of 1,3,2,6,5,15. Note that each pair (1st and 2nd lines, 3rd and 4th lines, 5th and 6th lines) is the relative contrast binon 1,3. And also notice that each pair increases in relative intensity of the 1,2,5-contrast pattern. If we create the tree of binons using all adjacent lines we will not recognize the 1,2,5 pattern. To do this the adjacent binons at level-2 that are not sharing an overlapping level-1 line (these are the pairs) must be combined. This in effect is adding a gap of one binon between the level-2 binons. Extrapolating this principle it would seem that at every level combinations must be created using all possible gap / separations. This is the same sort of thing that happens when combining the stimuli in P-Habits as described on 24th Feb 2012.
Another example to illustrate this point is based on the 9 lines of intensities 1,3,1,2,6,2,5,15,5. The triples are based on the 1,3,1 pattern and these are found at complexity level-3. When these level-3 binons are combined the 1,2,5 pattern is found. But at level-3 the triples are found at every 3rd binon (a gap / separation of 2 binons).
29th Feb 2012 Where is it
However the 1,2,5 pattern describes what has been recognized not where it was found. The ‘where’ information that must be put with the two patterns for an object must provide the level of complexity at which it was found.
1st March 2012 Parallel Configurations
On the 24th Feb I didn’t quite solve the problem of 19th Feb. What the solution does is stops the A^C combination from being formed on the first ABC occurrence. But AXC will not create the A^C because both A and C are familiar but the pair does not exist. So it is not a solution. So I need to create the A^C when they appear together and both are novel. A possible solution is to flag the biggest combination as independent and all the sub-parts at all level below it that went into its composition as dependent. When combining parts to form higher levels never combine any parts that are independent. To be combined the parts must both be novel or familiar dependent and the combination must already exist. This will not solve the case in which ABC becomes permanent and A^C is not.
I’ve been looking at how many subpart combinations at different levels belong to higher level combinations. For example with 3 independent sensors there are 3 stimuli at level-1, 3 pairs at level-2 and 1 level-3 combination.
# Sensors Stimuli @ Lev-1Lev-2 Lev-3 Lev-4 Lev-5 Lev-6 Lev-7 Lev-8 Total
3 3 3 1 7
4 4 6 4 1 15
5 5 10 10 5 1 31
6 6 15 20 15 6 1 63
7 7 21 35 35 21 7 1 127
8 8 28 56 70 56 28 8 1 255
Maybe the parallel independence concept is over thinking the problem and I should really go back to my thinking of 28th July 2010 and use the change that takes place in sequence to determine what P-Habits to use.
3rd March 2012 Habituation
Habituation occurs when a sensor detects a repeated stimulus. But it does send the stimulus the first time it detects it. In order that repeated stimuli become habituated when conscious we must first explore them. They then become permanent. Maybe rather than become permanent I should automatically start the repeater A-Habit that does the orient response and set it up to be a repeated acton done subconsciously.
11th March 2012 Conscious STM
Do I start sequentially combining two permanent stimuli in the sense's STM or do I combine them in a conscious STM? When I get two permanent stimuli in sequence from different senses I will have to combine them in a conscious STM. But when I am recognizing a goal stimulus in a directed way the two stimuli may not be in the conscious STM. This directed attention must go to the appropriate sense's STM based on the stimulus expected. STM's are used to build up stimuli that attract attention and not to create them for directed attention matching.
13th March 2012 Conscious STM
I think conscious STM is only necessary to recognize a repeat. It is not trying to recognize new sequences of conscious stimuli. New sequences will show up as A-habits with an orienting response and an interest in being redone. And all goal stimuli originate from stimuli that attract attention so it would seem that there is no need to have directed attention ever recognize a sequence formed in conscious STM. This is where a problem would occur because the sequences produced by the conscious STM could be comprised of stimuli from various senses. This means that the directed attention need only match up with the STMs that are per sense or per P-Habit. This also leads me to the conclusion that a sequence of permanent stimuli from the same sense should be combined in that sense's STM.
16th March 2012 Levels of Complexity
For objects to be recognized at any level of complexity, every level that has been produced has to be reprocessed as though it was level 1. At level 1 the objects are adjacent, there is no overlap. In fact to be independent of level of complexity any two adjacent objects of different or the same complexity must be combined. Each object must carry with it its total size and intensity (presumably of its left most part). Gaps are then recognized objects made up of any number of parts serving as a single adjacent object. The information that identifies what is an object consists of its shape (width pattern); its contrast pattern based on the same parts used in the shape pattern plus sense pattern (highlighted). The object's values are then its size, intensity, contrast negative, contrast reflection, shape rotation and position.
|
Sensors are: |
Dependent |
Dependent |
Independent |
Independent |
| Sense pattern | Yes | Yes | Yes | Yes |
| Sensor pattern | - | - | Yes | Yes |
| Sensor position | Yes | Yes | - | - |
| Shape pattern | Yes | Yes | - | - |
| Shape rotation | Yes | Yes | - | - |
| Size | Yes | Yes | - | - |
| Symbol pattern | - | Yes | - | Yes |
| Contrast pattern | Yes | - | Yes | - |
| Contrast negative | Yes | - | Yes ? | - |
| Contrast reflection | Values | Symbols | No ? | - |
| Intensity | Yes | - | Yes | - |
As soon as the sensors are independent we lose the shape pattern, size, shape rotation and contrast reflection and instead have a sensor pattern (configuration). As soon as the readings are symbolic we lose the contrast pattern and contrast reflection. With symbolic readings the contrast pattern is replaced by a unique identifying symbol and the contrast negative and contrast reflection are not relevant.
Where and What
Therefore to identify an object type you need a shape or sensor pattern, a contrast or symbol pattern and the sense pattern. Then to give it values the other items are combined. Which of these two collections is the 'where' information for paying attention and which is the 'what' information? [Read 14th June 2010 and 20th Dec 2010] Sense and sensor patterns are used as where information, whereas symbol patterns are what it is. Does that mean that shape patterns describe where and contrast patterns describe what? Or is the where information just the structure of the composition? That is just the configuration of the parts at the different levels of complexity. This could be obtained from the sense and sensor patterns or the sense and shape pattern. Would the what information then be the entire collection of patterns and values? Let's consider black and white vision and list somewhere information that one can use to pay attention. One can pay attention to:
- The brightest or darkest object
- The object in a certain location relative to the frame (top left, bottom middle etc.)
- The largest or smallest object
- The object of a certain shape
- The upside down object
- The negative object
- The most or least complex object
These examples seem to imply that the where information is any combination of the values plus the object type as identified by the pattern combination. This makes me wonder why do I need to separate where from what? Directed attention looks for the closest object to the expected one. The where / what separation helps one think of this idea more systematically. But is that the right approach? What do I mean by 'closest object'?
When adding time / sequence to this a sequence is dependent. Thus a sequence has a time shape and edges occur between adjacent objects in time.
17th March 2012 Levels of Complexity
I've realized that if I combine stimuli for dependent sensors at all levels as I do for P-habits then I will be producing combinations that are adjacent at all levels as necessary to recognize contrast and shape patterns independent of level of complexity. But then I must apply a restrictive principle to minimize the combinatorial explosion. And that principle is the one that says two familiar stimuli are not combined to produce a novel one (one that does not exist). This is currently used for S-Habits. Two novel stimuli in sequence get combined to form a novel stimulus but not two familiar ones to form a novel one. The two familiar ones will be ultimately combined when they both become permanent and adjacent sequentially. This is all about edges. Edges consisting of two adjacent familiar stimuli that continuously change from one frame to another don't get an opportunity to become a familiar pair. But at some point two objects stick together and form a single new larger object. At this point the edge pair becomes familiar. It's the same rule that I use for learning A-Habits. They have to occur twice, once to create them in novel mode and then again in practice mode to become familiar. Maybe an edge needs to occur twice in a row before becoming a novel stimulus. This would certainly be consistent with the idea of two objects sticking together. But if a particular edge occurs for the first time in frame 1 it does not exist and will not be created so it will not exist when frame 2 is experienced. How am I to recognize that it now exists and is familiar? Do I create them in a transitory state and delete them if they do not reoccur?
19th March 2012 Edges
Edges are not remembered. Edges are used to recognize objects. But an object may contain an edge within it. This edge is represented as the contrast between two adjacent parts. But the familiar parts inside the two adjacent parts are also familiar across the edge. This means there is internal familiarity across the contained edge.
Levels of Complexity
To address the level of complexity problem I need a pattern structure that can incorporate gaps and still recognize a pattern no matter what the complexity of its parts. For the general-purpose gaps solution what I need is a gap width pattern that is combined with the Shape/Sensor and Contrast/Symbol patterns in object type identification. For example if 5 dependent sensors produce ABCDE,
Then at Object-Level (O-Level) 2 I need to produce:
AB, BC, CD, DE Gap width 0
AC, BD, CE Gap width 1
AD, BE Gap width 2
AE Gap width 3
Then at O-Level 3 I would produce:
ABC, BCD, CDE Gap width pattern 00
ABD, BCE Gap width pattern 01
ACD, BDE Gap width pattern 10
ACE Gap width pattern 11
ABE Gap width pattern 02
ADE Gap width pattern 20
At O-Level 4 I would produce:
ABCD, BCDE 000
ACDE 100
ABDE 010
ABCE 001
The same rules for combining pairs from the level below apply for the Contrast/Symbol and the Gap patterns. That is the overlap rule of a common part. This results in the object type being identified by the combination of the 4 patterns - Sense, Shape/Sensor, Contrast/Symbol and Gap.
Object Values
Values will be needed for each object for the seven properties as found in this table.
|
Sensors are: Readings are: |
Dependent |
Dependent |
Independent |
Independent |
| Shape rotation | Y/N | Y/N | N | N |
| Contrast negative | Y/N | N | Y/N | N |
| Contrast reflection | Y/N | Y/N | N | N |
| Size | # of Sensors | # of Sensors | # of Sensors | # of Sensors |
| Intensity | Value | Symbol | Value | Symbol |
| Sensor position | Sensor # | Sensor # | Sensor # | Sensor # |
| Perception / Thought | Y/N | Y/N | Y/N | Y/N |
Binon Structure
To apply binons to these concepts I would use binons to combine the object type patterns as independent objects. This means only 3 combinations at level 2, 2 combinations at level 3 and 1 unique combination at level 4. Similarly the seven properties combined in a tree to uniquely identify the combination of values. Finally the object type and values would be combined to capture the experience.
Level of Complexity
But this still does not allow me to recognize an O-Level-2 object made up of O-Level-3 objects. What I need to do is take all the adjacent level 2 objects (each 2nd one) and treat them as a set of level-1 objects and create all the higher levels including appropriate gapped ones. Then I need to take all the adjacent level 3 objects (each 3rd one) and treat them as level-1 objects and create all the higher levels.
22nd March 2012 Reflections versus Negative images
On 7th April 2010 until 21st April 2010 I was working on reflections and negatives of images. I have now been rethinking the algorithm to make it simpler while I am also rearranging the binon structure for pattern recognition. Facts known so far include; all graduated reading objects have an intensity value, a reflection (left to right order) value and a negative (increasing or decreasing intensity) value. Given two parts X and Y that form an object their order can be X,Y or Y,X. And X and Y can be individually positive or negative (intensity value direction). When they are combined you end up with two different objects XY#1 and XY#2 that are not reflections or negative of each other. However each has a reflection, a negative and a negative reflection. The 1st is the X,Y in which the direction of the values increase / decrease in opposite directions. The possible combinations are:
XY#1 Reflection Negative Real Numbers
(order) (value direction) Difference Difference
X+ Y- no no 3,4 1 4,2 -2
X- Y+ no yes 3,2 -1 2,4 2
Y+ X- yes no 2,4 2 4,3 -1
Y- X+ yes yes 4,2 -2 2,3 1
XY#2 Reflection Negative
(order) (value direction)
X+ Y+ no no 2,3 1 3,5 2
X- Y- no yes 5,4 -1 4,2 -2
Y+ X+ yes no 2,4 2 4,5 1
Y- X- yes yes 5,3 -2 3,2 -1
If we now have another two objects Y,Z#1 and Y,Z#2 as follows:
When they have a common part with the X,Ys and we combine them we get the following:
YZ#1 Reflection Negative Real Numbers
(order) (value direction) Difference Difference
Y+ Z- no no 4,6 2 6,3 -3
Y- Z+ no yes 5,3 -2 3,6 3
Z+ Y- yes no 3,6 3 6,4 -2
Z- Y+ yes yes 6,3 -3 3,5 2
YZ#2 Reflection Negative
(order) (value direction)
Y+ Z+ no no 3,5 2 5,8 3
Y- Z- no yes 8,6 -2 6,3 -3
Z+ Y+ yes no 3,6 3 6,8 2
Z- Y- yes yes 8,5 -3 5,3 -2
XYZ#1 Reflection Negative on the left on the right
(order) (value direction)
X Y- Z 3425 no no XY#1 no no YZ#1 no yes
X- Y Z- 5463 no yes XY#1 no yes YZ#1 no no
Z- Y X- 5243 yes no YZ#1 yes yes XY#1 yes no
Z Y- X 3645 yes yes YZ#1 yes no XY#1 yes yes
XYZ#2 Reflection Negative
(order) (value direction)
X Y Z 2358 no no XY#2 no no YZ#2 no no
X- Y- Z- 8752 no yes XY#2 no yes YZ#2 no yes
Z Y X 2578 yes no YZ#2 yes no XY#2 yes no
Z- Y- X- 8532 yes yes YZ#2 yes yes XY#2 yes yes
ZYX#1 Reflection Negative
(order) (value direction)
X Y Z- 2352 no no XY#2 no no YZ#1 no no
X- Y- Z 5425 no yes XY#2 no yes YZ#1 no yes
Z Y- X- 2532 yes no YZ#1 yes no XY#2 yes yes
Z- Y X 5245 yes yes YZ#1 yes yes XY#2 yes no
ZYX#2 Reflection Negative
(order) (value direction)
X Y- Z- 5641 no no XY#1 no no YZ#2 no yes
X- Y Z 1025 no yes XY#1 no yes YZ#2 no no
Z Y X- 1465 yes no YZ#2 yes no XY#1 yes no
Z- Y- X 5201 yes yes YZ#2 yes yes XY#1 yes yes
From this analysis I have determined the algorithm should act as follows: First the two parts should be combined based on a common part to form the new object.
Then the new object's #1 or #2 is the same as for the YZ part that was used.
Its reflection property is based on the order in which the two parts were joined.
If its reflection property is 'no' then
its negative property is that of the left part that was used.
or if its reflection property is 'yes' then
its two object #s are the same then
its negative property is that of the right part that was used.
or its two object #s are different then
its negative property is that of the left part that was used.
Gap problems
Having started to implement this algorithm I've realized that I have to solve the problem with gaps that expand in size along with the other parts. Also on a circular sense with only 2 lines there could be two possible gaps.
23rd March 2012 No Reflection / Negative
I also have to deal with the situation when using independent sensors two adjacent sensors can have the same value. This produces a contrast pattern with no reflection and no negative. I also have to figure out if this also would happen when doing contrast pattern creation with independence of level of complexity.
24th March 2012 No Reflection / Negative
I have realized that with independent sensors and graduated readings that contrast reflections and negatives are not possible. It is only when sensors are adjacent in a linear relationship that this can happen. Independent sensors are not adjacent. Thus, on the table from the 16th March 2012 the 'Yes?' and 'No?' need to be blank.
Reflections versus Negative images
I need to see if the logic for O-Levels higher than 3 are the same at O Level 3. There is only one Level-2 object pattern. And when the values are reversed it is flagged as a negative. The two values are A and B in which B>A.
AB Reflection Negative Real Numbers
(order) (value direction) Difference = 1
AB no no 4, 5
BA no yes 5, 4
If we now have another level 2 object BC (C>B).
BC Reflection Negative Real Numbers
(order) (value direction) Difference = 3
BC no no 5, 8
CB no yes 8, 5
When these are combined the following patterns are possible.
ABC#1 Reflection Negative on the left on the right
(order) (value direction)
AB CB = 452 no no AB no no BC no yes
BA BC = 436 no yes AB no yes BC no no
BC BA = 254 yes no BC no no AB no yes
CB AB = 634 yes yes BC no yes AB no no
ABC#2 Reflection Negative
(order) (value direction)
AB BC = 458 no no AB no no BC no no
BA CB = 874 no yes AB no yes BC no yes
CB BA = 854 yes no BC no yes AB no yes
BC AB = 478 yes yes BC no no AB no no
From this analysis I have determined the algorithm should act as follows: First the two parts should be combined based on a common part to form the new object.
Then the new object's ABC#1 or ABC#2 is based on whether the negative signs of the two parts are the same or different.
Its reflection property is based on the order in which the two parts were joined.
If its reflection property is no then
its negative property is that of the left part that was used.
If its reflection property is yes then
if the two part's negative properties are different then
its negative property is that of the left part that was used.
If the two part's negative properties are the same then
its negative property is opposite that of the left part that was used.
25th March 2012 No Reflection / Negative
I have revised my thinking for independent sensors and graduated readings. Reflections are not possible because this needs some form of order of sensors. However I need negatives because the contrast pattern of independent sensors that read 3444 must match 4344, 4434 and 4443. It must also match any combination with 3 sensors of value x and a forth sensor of value x-1.
26th March 2012 Edges
Continuing thoughts from 19th March 2012. For an edge to become the separator of two parts of the same object the two parts have to somehow be seen as belonging together. This means they have to move together or change their intensity together. Said another way the edge cannot be changing or the edge has to exist for more than one frame in a row. This approach would mean that given two sensors on a graduated sense with symbolic readings A and B the 1st frame might contain AB. The individual stimuli A and B would be created but not the combination because that represents the edge. Then when the same AB occurs a second time the combination would form. If instead the 2nd frame contains AC and the 3rd frame AB, the 3rd frame would not cause the combination to form because it was not immediately after the first AB. Extrapolating this idea to multiple sensors would mean that on each additional frame of an unchanging pattern the combinations would be formed at higher levels of complexity. This approach would slow down the combination of any two simultaneous familiar stimuli into a new one and would work at any level of complexity. It could also be used to slow down the combination of two simultaneous novel stimuli. The combination of a familiar and novel stimulus would not take place until the third occurrence. On the 1st frame the novel ones would become familiar. On the 2nd frame the combination would not exist and would wait for the 3rd frame to be combined. Thus known familiar combinations would be recognized much faster than any other combinations.
If however it were applied to sequences a pattern such as ABCD would take 4 occurrences before it was formed. Is it possible that the rule only applies to the sensory level forming combinations? At least then the familiar combinations would have sufficient lead on the unfamiliar combinations and it would not take 4 cycles to recognize an O Level 4 complexity object.
28th March 2012 Edges - Sensors
In order to address the edges and combination of familiar stimuli problem I have decided that the following logic must take place at the different levels. At level 1 (sensor detected readings) novel stimuli are generated, they become familiar the 2nd time they occur and otherwise they are already familiar and stay such. At level 2 novel stimuli are not combined. Familiar and novel stimuli are not combined. Familiar ones are combined if the combination already exists. But two familiar stimuli are combined if the same pattern is found in the previous frame. At level 3 familiar ones are combined if the combination already exists. Novel and familiar are not combined but two novel ones are combined.
This started me thinking about graduated reading sensors. They are not just measuring devices providing a reading. They have memory and that allows them to detect change. But this does not allow them to identify objects. But as soon as two sensors detect different readings we have an edge and therefore a pattern. And this can be used for identity. So maybe what I need is a level 0 object that is what each sensor detects. Then level-1 objects are the pairs of readings with an edge between them which is the 1st symbolic / identity level. Can the same be said for size / width / shape patterns? Level 0 would be the width reading and level 1 would be the shape pattern formed from a pair of widths.
However gaps as patterns (identifiable / symbolic objects) don't really come into existence until what to date I am calling level 3. 3 lines are needed to give 2 gaps that combined give a gap width ratio. Shape patterns come into existence at level 2 because there are two widths. Contrast patterns also come into existence at level 2 from two readings. Symbolic objects and sensor numbers exist at level 1 because they are identifiable at this level. So rather than introduce a level 0 maybe I should say that level 1, and 2 are processed by the sensors and levels 3 and higher are all symbolic levels and come after sensory processing. Does this mean that at level 1 and 2 I don't produce experiences? I just hold the information that is created at those levels and use sensory rules for combining information rather than symbolic rules for combinations.
Or if I keep the current levels and combinations it means that the combination of novel and familiar rules differ at each level depending on the graduated / discrete nature of the information obtained. For example with symbolic readings and independent sensors at level 1 shape and contrast patterns are uniquely identified already. Thus level 2 combinations are formed if the two source parts are novel or the two are familiar and the combination exists. They are also formed if the two source parts repeat on the same sensor pair. This last rule applies for P-Habits because each part is on a different sense and therefore represents a different type of information. But if the parts are on the same sense they represent the same type of information. This means that a configuration of symbolic values should be sensor position independent. A tiger on sensor #1, plus a monkey on sensor #2 and a parrot on sensor #3 should be the same tiger, monkey, and parrot combination on any other sensor combination. Or TMP is the same as TPM, MPT, MTP PMT and PTM. The sensor combination must then have an ordering to represent where these symbolic values were obtained.
Symbolic independent sensors
If there are many sensors only a subset will change from one frame to another. The ones that do not change should be combined into a familiar combination because they don't change together which is the same as saying they all change together. Of the ones that have changed many symbols may be familiar. Any combinations of these that exist should be recognized. No new combinations should be formed from them. This rule solves the problem of when static happens. This is when lots of things are changing from one frame to the next and all sorts of novel combinations of familiar stimuli occur. We don't want to form these combinations unless they have some degree of permanence. Of the ones that have changed many symbols may be novel. Are these all combined to form their biggest combination? I suspect yes just as I do for sequences of novel stimuli. The concept is they all came together as novel and therefore all changed together. If a subset of this novel collection repeats then it will be conscious independent of the whole. Is it consciousness that these combinations have that their parts do not have?
31st March 2012 Sorted stimuli
I have successfully implemented an algorithm that generates the right patterns for independent sensors. For 4 independent sensors with symbolic readings I need a unique ‘where’ that is the sense # plus the series of sensors 1,2,3,4. I then need a unique configuration independent of which sensors the symbols occur such as E,M,P,Y and the sensor pattern on which they were found such as 3,1,4,2. This then describes the pattern MYEP if the sensors are in number order. The pattern PYME would be the same configuration E,M,P,Y (sorted) but the sensor pattern would be 4,3,1,2. For 4 independent sensors with graduated readings I need the same 'where' sense and series of sensors 1,2,3,4. However the contrast configuration is a sorted series of contrasts, an initial reading and the sensor pattern. Thus if the sensors based on sensor # produce 7241 the initial reading is 7, the contrast series is 1,2,3 and the sensor pattern is 4,2,3,1. The sorted intensity pattern is 1,2,4,7. The contrast between these 4 numbers gives the contrast series. The sensor pattern positions them and the initial reading determines the overall intensity. If the initial reading were 9 the following operations would generate the original experience. Generate a series starting at 0 with the given contrast values = 0,1,3,6. Now reorder it according to the sensor pattern and you get 6,1,3,0. Take the difference of the 1st value in this pattern and the initial reading = 9 - 6 = 3. Add this to the reordered pattern and you get 9,4,6,3. I just realized it's the contrast pattern that needs to be in sorted order - not the intensity reading pattern.
Edges - Sensors
I had thought that I needed to separate the changed from non-changed stimuli before combining them so that the non-changed get combined into new patterns because they did not change together. However a better strategy is to 1st combine any pair of novel ones at the same level that have change flags and thus form the complete tree of them. Then at any given level I go through the list of familiar stimuli and combine all the ones for which combinations exist. This leaves the unused familiar ones at this level. Of these I combine any that are unchanged. This will produce novel combinations that have an unchanged flag. At the next level higher a pair of these novel unchanged ones would be combined to form a novel unchanged combination. And then the unused familiar ones that have changed don't get combined. They have to stay unchanged before they get combined.
In table form the sequence is:
| Order done | Both sources | Sources Changed | Pair exists | Result will be | Result Changed | Order done at next level |
| 1 | Novel | Yes | - | Novel | Yes | 1 |
| 2 | Familiar | Yes or No | Yes | Familiar | Yes or No | 2, 3 or 5 |
| 3 | Familiar | No | - | Novel | No | 4 |
| 4 | Novel | No | - | Novel | No | 4 |
| 5 | Familiar | Yes | - | No | - | - |
The order of processing could be 1, 4, 2, 3 and 5 at each level. Step 2 must always be done before 3 to mark the used ones. Only at step 2 can a changed and unchanged familiar pair be combined.
1st April 2012 Parts of the experience
The following table describes what the parts are that make up an instantaneous experience, not including sequences or P-Habits.
|
Sensors are: |
Dependent |
Dependent |
Independent |
Independent |
| Sense (where) | Y | Y | Y | Y |
| Position (sensors or location - where) | Y | Y | Y | Y |
| Shape (sensor or width pattern) | Y | Y | Y | Y |
| Shape size | Y | Y | ||
| Shape rotation | Y | Y | ||
| Gap pattern | Y | Y | ||
| Gap size | Y | Y | ||
| Gap rotation | Y | Y | ||
| Image (symbol or contrast pattern) | Y | Y | Y | Y |
| Intensity | Y | Y | ||
| Contrast negative | Y | |||
| Contrast reflection | Y | Y | ||
| Perception / Thought | Y | Y | Y | Y |
Interesting / Familiar
I'm wondering if I need to use the interesting / familiar information at the part level rather than the experience that is made up of the parts. If we consider the symbolic independent sensors the 'where' information is the sense and sensor pattern and the 'what' information is the symbol pattern and its sensor location pattern. There is no novel / familiar aspect to the 'where' information but both parts of the 'what' information can change independently. I can have AB on sensors 2 and 4 and then have MN on these two sensors. This is a change in the symbol pattern. Or I can have AB on sensors 4 and 2. This is a change in the sensor location pattern. The two parts of what it is can change independently. It is hard to think about the sensor location pattern without slipping into the mode of thinking about it as the sensor pattern part of the where information. It might be easier to ask the question about graduated dependent sensors in which the where information is the sense and a position while the 'what' information is the shape (and gaps), contrast and properties. The shape and contrast parts can vary independently. And it seems realistic to have interest in a shape and / or its contrast. But the lowest level 'what' is the shape. The contrast is a property of the shape. Contrast patterns provide a texture to the shape. The shape is the most general type of thing while a contrast pattern makes it more specific.
Where versus What
The where and what concepts overlap. At the extreme / simplest ‘what’ end there is the shape or symbol pattern. At the extreme ‘where’ end there is the sense. As more properties are added to the shape (size, rotation, position) or symbol pattern (reflection, position) the object becomes a more specific what. As more properties are added to the sense (position, width, intensity, contrast) the where becomes more specific. A subset of the where information can change while the remainder what is the same or vice versa. But it is the 'what' information that is interesting. The where- information is used when directing attention to a goal. So each more complicated combination of what properties can have interest provided there is an object type (shape or symbol pattern) at its core. Properties don't have interest in themselves. Width, intensity, location, reflection, rotation etc. do not have interest. But when added to a specific type of object the combination has interest. For symbolic independent sensors the AB could be interesting. Both AB on 2 and 4 and on 4 and 2 could have separate interests. But sensors 2 and 4 would not have interests. Another implication is that AB on one sense is not the same as AB on another sense since each sense measures a different type of information. Thus the sense must be added to the symbol or intensity at its lowest level to identify the type of information. However the properties of reflection, rotation, negative and others can be sense independent. Thus they are often called concepts rather than objects.
5th April 2012 A-habits
The action habits that are easily delegated to subconscious execution tend to be those that have a repetition / cycle in them such as walking or produce a holding pattern such a keeping your head erect or a limb in a given position. However there are one shot action habits that are started consciously but performed subconsciously such as stand up from sitting position.
Edges - Sensors
The algorithm for object recognition combines any two novel parts provided they have both changed or not changed - producing a changed or unchanged novel combination. It finds any two familiar parts that already exist as a combination (whether the parts changed or not) - producing a familiar one and it is marked changed if one part changed and the other didn't. Else it is marked with the same change as its two parts. It also combines any two familiar parts that did not change and whose combo does not exist. This will produce a novel combination that did not change (handled at next level up). Any two familiar parts whose combination does not exist in which one or both parts changed are not combined. The only parts left unused will be familiar ones that changed but their combinations do not exist and possibly a single familiar one that is unchanged and possibly a single novel unchanged and possibly a single novel changed one.
Sequences of P-Habits
As a result of the changed strategy of combining parallel stimuli and my current S-habit formation approach the following sequence of P-Habits should be analysed as follows.
| Sequence | Sense pattern |
Comments - Each letter position represents a different sense. |
| 1 | A B C D | Form all the P-Habits up to A^B^C^D by combining novel stimuli. The largest combination does not become conscious because it is the 1st novel stimulus in a sequence. |
| 2 | A B E D | A^B^D has not changed, it repeats and it already exists. It becomes a conscious stimulus. Attention is attracted to this repeat and it is expected to repeat again. E is novel so C,E is a novel sequence and remains subconscious. |
| 3 | X Z L N | Since all the senses change and all the stimuli are novel all the P-Habits up to X^Z^L^N are formed. The active action habit was expecting another A^B^D on sense #1, 2 and 4 but gets an X^Z^N instead. This forms the A-Habit A^B^D _ X^Z^N in which _ is the orient response. Thus X^Z^N becomes conscious. L is novel so C,E,L is a novel sequence forming on sense #3. |
| 4 | X H L P | X^L has not changed, it repeats and it already exists. H^P is novel and forms this P-Habit. The novel sequence Z,H N,P and Z^N,H^P are formed and remains subconscious. The repeat of L will cause C,E,L to become conscious as well as the L. The C,E,L attracts attention before the repeating X^L. Thus X^L remains subconscious. The A-Habit C,E,L _ L is formed. L becomes the focus of attention waiting for whatever stimulus occurs next. |
| 5 | X Z C D | All the stimuli are familiar. X^Z and C^D are already existing P-Habits. Both will be flagged as changed but not combined because they don't have a common part. The Z,H sequence is terminated because the familiar Z repeats. The N,P sequence is terminated because the familiar D occurs. The novel N,P attracts attention and the L _ N,P action habit is formed. It is now waiting for whatever stimulus occurs next after the N,P. |
| 6 | X B L D | All the stimuli are familiar. X^L and B^D already exist. Both will be flagged as changed but not combined because they don't have a common part. X^D is formed because it repeats with no change. But it is the conscious L that causes the C to become conscious and it attracts attention. The N,P _ C A-habit is formed. The L then has attention and it is expecting a N,P because of the L _ N,P action habit. |
In step 4 the X^L remains subconscious because C,E,L ends because the L repeats. May be X^L should not have formed at all.
If we use a strategy where a sensor does not produce a stimulus a second time in a row because it habituates, then only changed stimuli would be available for attracting attention. (Read 26th June 2009, 6th March 2010). It may be useful to distinguish between the stimuli that are formed and combined to attract attention and those that can be formed to satisfy directed attention. The repeat of a stimulus on a sensor should not be used to attract attention but should be available for directed attention. But then if there is only one sense with one sensor on the 1st frame the stimulus is novel and attracts attention. On the second cycle to get a repeat attention must be directed at that sensor.
Entire tree
On the subject of the whole entire tree being formed this is facilitated by the rule that two familiar stimuli are always combined if the combination already exists. This takes place whether or not a change has taken place. But if sensors habituate and don't produce the stimulus when it repeats this entire tree formation would have to be a result of directed rather than attractive pattern matching.
6th April 2012 Edges - Sensors
A diagrammatic representation of the rules for combining parts helps to illustrate the algorithm.
F = Familiar, N = Novel, u = unchanged, c = changed, e = exists already
Nc Nu Fec Fec Feu Nu
/ \ / \ / \ / \ / \ / \
Nc Nc Nu Nu Fc Fc Fc Fu Fu Fu Fu Fu
Wherever there is an unchanged stimulus we have a repeat. For the last one of these situations we are creating a novel combination but immediately saying it is a repeat. This happens because in the previous cycle the parts were familiar or novel but were not combined. But isn't a repeat equivalent to uninteresting. If then we reinterpret the u to mean uninteresting we have a novel stimulus that is uninteresting. This seems like a contradiction in terms because novel is usually associated with interesting. But now we must understand novel to mean newly created. Now let us look at these situations using F = Familiar, N = New, i = interesting, u = uninteresting, and e = exists already.
Ni Nu Fei Fei Feu Nu
/ \ / \ / \ / \ / \ / \
Ni Ni Nu Nu Fi Fi Fi Fu Fu Fu Fu Fu
I now realize that familiar implies that it exists already. Now using E = Exists already and thus is familiar, N = New, i = interesting, and u = uninteresting.
Ni Nu Ei Ei Eu Nu
/ \ / \ / \ / \ / \ / \
Ni Ni Nu Nu Ei Ei Ei Eu Eu Eu Eu Eu
The interest or no interest originates at the sensor level in the form of change or no change. But the next problem is the new and existing information. When creating the combinations the parts always exist. I have been using the interest information setting of 'interesting' to indicate it is new. So now I replace the N with E for existing and look at the situations.
Ei Eu Ei Ei Eu Eu
/ \ / \ / \ / \ / \ / \
Ei Ei Eu Eu Ei Ei Ei Eu Eu Eu Eu Eu
Merging the 1st and 3rd rule causes no problem because if the combination does not exist we create it because both parts are interesting. The 6th rule would create an existing uninteresting combination. These are only meant to be combined using rule 2 that now looks like rule 5 that happens to be the same as rule 6. However they are not meant to be used as one of the parts in rule 4. So rule 4 says two parts with different interest can only be combined provided the combination already exists. In summary, 1/ two parts are combined if they have the same interest to produce a combination with this interest and 2/ two parts of different interest are only combined if their combination already exists. In the second case the combination is interesting. In the first case if the combination exists already it is not created but just found.
7th April 2012 S-Habit Recognition
When I examine the rules for STM sequential pattern recognition based on the above rules for P-Habit recognition I come up with the following observations. When two parts occur in sequence and the second is the same as the 1st it will be uninteresting. This is happening in the processing of the 2nd frame. STM should use this and immediately flag it as a repeat. It should also be accumulating a history of stimuli to recognize longer repeats, as it currently does. The pattern that repeats becomes conscious / actionable. For two parts that occur in sequence that do not repeat STM starts building the S-Habit. An edge occurs when the tree it is building changes from creating new combinations to finding existing ones or vice versa. And when it is building either tree and it comes across a conscious / actionable stimulus or combination then the tree and the conscious / actionable stimulus become conscious.
8th April 2012 Edges - Sensors
My analysis of the 6th April 2012 was wrong because a change does not mean there is an interesting stimulus. Interesting stimuli are the new ones. A changed one may be new or existing / familiar. But an unchanged one is uninteresting. Before this mistake I had F = Familiar, N = Novel, u = unchanged, c = changed, e = exists already
Nc Nu Fec Fec Feu Nu
/ \ / \ / \ / \ / \ / \
Nc Nc Nu Nu Fc Fc Fc Fu Fu Fu Fu Fu
I now replace Familiar with Existing and u = uninteresting.
Nc Nu Ec Ec Eu Nu
/ \ / \ / \ / \ / \ / \
Nc Nc Nu Nu Ec Ec Ec Eu Eu Eu Eu Eu
I still have this contradiction of a new stimulus that is boring in rule 6. I think the problem is that change should not be propagated up the tree. It should be interest that goes up the tree. So at the lowest level u = unchanged. Then interest will be created at level 2 and higher as a result of change. Now using i = interesting, o = neutral interest and b = uninteresting / boring.
Level 2 - Ni XX Eo Eo Eb Ni
/ \ / \ / \ / \ / \ / \
Level 1 - Nc Nc Nu Nu Ec Ec Ec Eu Eu Eu Eu Eu
XX is not possible because Nu is not possible. Eo means the combination already exists and combining the two parts changes its interest to neutral as a result of it being experienced a subsequent time. Now, at levels 2 and higher the possibilities are combined based on the parts being new or familiar. Familiar means existing with either neutral or boring interest. Parts are always combined if the combination already exists.
Ni Eo Eo Eb Ni
/ \ / \ / \ / \ / \
Ni Ni Eo Eo Eo Eb Eb Eb Eb Eb
But I am starting to question whether I need the neutral interest state. Let's get back to basics. If a combination of parts already exists it is first to be recognized because the paths are already established. Is the interest in this combination neutral or uninteresting? Does the existing combination become uninteresting if its two parts are uninteresting? The two parts became uninteresting because of no sequential change i.e. a repeat so it would seem that the combination also has no sequential change and is uninteresting also. If two parts are new / interesting then the combination is new / interesting and this gives us a future combination that can be recognized as already existing. Two parts are assumed to be independent until a dependency occurs. For two new parts the dependency is they came into existence at the same time. Dependency occurs generally when a relationship between two parts does not change. The two parts change position, size or brightness etc. in unison. This requires two frames. One frame to establish the relationship and the second to notice that it has not changed.
What if we say that all stimuli that a sensor can detect are known and exist, i.e. are familiar and can never be novel? This might work at the sensor level for graduated readings but not symbolic ones. The challenge for graduated readings is to not form combinations out of random static. What do we do with random static of symbolic readings? Symbolic readings imply that the parts are uniquely identified. If all the lowest level symbolic parts are known / familiar when do we form combinations? The relationship between two parts must not change. However the relationship is the combination. In a first frame the combination is not created. Then in the second frame the combination is created when the two parts are the same symbol on the same sensors. If we were dealing with graduated readings and dependent sensors then both of the contrast and shape patterns would have to be the same from one frame to the next.
12th April 2012 Edges - Sensors
I have re-read my notes and I think that the idea from 26th March 2012 of combining only one additional O Level of novel parts per cycle / frame might be the answer. Thus a 100 sensor sense would have to repeat the same readings 100 times before the O Level = 100 combination was formed if the 1st frame of readings were all novel. However, existing (familiar / known) combinations would be formed from familiar readings on just the 1st frame. If one considers sequential recognition of 4 novel stimuli it takes 4 cycles / frames to form the S Level 4 S-habit. Using this approach basically says relationships between 2 relationships are only formed when the 2 relationships repeat. So for P-habits and multiple sensors (independent or dependent) O Level 1 processing would create binons for novel stimuli. O Level 2 and higher processing would never combine novel parts but create novel combinations from familiar / existing ones and only then if the relationship between the two parts has not changed.
13th April 2012 Interest propagation
I believe interest should propagate up the binon tree based on the interest of the parts. At level 1 the change or lack of it should allow for exiting / familiar stimuli to be assigned interest or not (uninteresting). Then as familiar parts get combined the interest in the combination should be based on that of the two parts. Hopefully this will allow the biggest change to attract attention.
This change resulted in the need to flag stimuli and combinations of stimuli as novel or familiar independent of their interest level. The novel/familiar information controls the combination process, even in the STM formation of sequences. Interest level is then only used for attention attraction.
16th April 2012 Conscious STM
I need to have a conscious STM to detect a repetition of the attended to stimuli which includes the thoughts we are conscious of. This must operate on the stimuli used in practice mode and the stimuli that attract attention when not practicing. However I also need to combine sequences of inter-sense stimuli that might attract attention or be disregarded if concentrating while in practice mode.
18th April 2012 Short Term Memory
I have been rethinking STM. Based on statements from the 1st paragraph of 26th April 2010 I want STM to identify unexpected stimuli. This would mean that as soon as it detects a novel stimulus or novel combination of stimuli it makes this available for conscious attention. The next time this stimulus is experienced it will be familiar and not be provided. However it will be used in combination with a next familiar one to form a novel combination because familiar ones will always be combined to either form a novel or familiar combination. Familiar combinations would not be provided for attention unless they were to repeat. This approach would mean that an S-Level 10 stimulus will need to be experienced 10 times before it is familiar. Familiar / expected sequences are effectively habitualized. Then conscious STM is doing the same thing but with stimuli sequenced across different senses. This approach may also rid me of the problem I currently have in Adaptron of two sequential stimuli being provided at the same time for conscious processing.
20th Apr 2012 Uninteresting
I've been trying to clarify exactly how to handle uninteresting stimuli. These are uninteresting because they repeat. How do we go about ignoring them? If we are concentrating on practicing an A-Habit they do not interrupt us. When they become conscious we are bored and react. When we are not practicing but doing some things subconsciously we usually revert to the thinking level of concentration and this is not interrupted by a repeating stimulus unless it happens to be so intense it overcomes our sensitivity threshold which must be set as part of concentrating. Do we set up some kind of subconscious habit to ignore uninteresting habits? But a subconscious habit does not have a goal. We have started it consciously to achieve a goal but the goal is not part of its execution. But intermediate sub-goals are. But even when these actions are done they don't change the interest level of the stimulus, they just recognize it and continue executing. So ignoring uninteresting stimuli just comes down to paying attention to something else.
I just had an idea that maybe, since neurons habituate when a stimulus repeats which means that it basically stops firing, I should not combine uninteresting stimuli with other familiar ones. Since they are not firing they cannot be part of a combination. The result is that only familiar neutral stimuli get combined sequentially.
22nd April 2012 Combining parts
The challenge with two senses and a repeating pattern X^Y is that on the 1st frame they are both novel and do not get combined. On the second frame the X and the Y repeat and are therefore both uninteresting. Again they do not get combined. The same happens on the 3rd frame etc. Thus all the exploration is done on the 2nd sense. It does not seem right. Why would one sense have precedence over another? Unless that sense happens to be muscle tension and we are learning how to control our muscles. There does not seem to be an opportunity to combine the two senses into a P-Habit given my new rules for combining stimuli. But maybe the clue is that each sense provides a different type of information (property) about the object. Thus they can be combined. It's the combination of two novel parts to form a bigger part that is limited until the two parts are familiar. For recognizing an object across dependent sensors with graduated readings I will combine the shape pattern with the contrast pattern even though each may be novel because they are different types of information / properties about the same object.
23rd April 2012 Properties versus objects
So when combining properties there is no restriction as to the two sub-properties having the same familiarity / novelty / interest. The properties of the object are all aggregated. But as soon as you have two readings for the same property such as two heights or two colours either in parallel or sequentially you form a pattern of these values and this is what identifies an object. This is where the rules for combining parts apply. So when forming P-Habits across senses all combinations / configurations of properties / senses are formed.
Object oriented thinking would say that all objects of a class have the same properties but may have different values for them. However the object class is based on all the objects of that type having a combination of properties that have the same value. These properties could be the existence of sub-parts such as "has teeth" or the values of attributes such as "colour is red". It is the non-class identifying attributes that have different values. Unless one uses the attribute to further discriminate / specialize the class of object. Any attribute can be used as a typing attribute to identify sub-classes. Each attribute value gives a different sub-class. The more general the type of object the fewer identifying properties are used. Until we are down to one property such as volume. Based on its value you have loud, quiet and some medium volume objects in between. It's a sound object. We have not included direction or frequency properties with values. That would be more discriminating.
We pay attention to the differences between two experiences and base our behaviour on those differences. Change (sequential) is a source of difference in time. We either generalize our behaviour by selecting fewer properties with specific values to act upon or we discriminate and base our behaviour on a larger set of properties with specific values. Because specific values are being used we are basing behaviour on object types.
This has led me to experiment with using interest levels to combine senses / properties. The best rule so far is one in which the two properties must have the same interest level to be combined. This still produces highest level P-Habits that are uninteresting so that the largest combination of senses / properties can attract attention if there is nothing novel or neutral. Also this produces the highest level P-Habit that is interesting so we pay attention to the largest combination of senses / properties to start with.
24th April 2012 First Novel
A lack of symmetry has shown up today in the recognition, attention to and reaction to the first stimuli when multiple senses are involved. If the combination rule is that two novel parts are not combined until they become familiar when they occur again later then the 1st occurrence of two simultaneous novel stimuli does not form a combination. Thus we cannot pay attention to their combination. If however we pay attention to one and not the other then we start to form orienting action habits based on it and end up with an asymmetric behaviour across senses. The other solution is to not form the orienting action habit until the stimuli are familiar. Thus, when a novel stimulus occurs we "sit back and enjoy it". This would mean that orienting action habits would only start when no novel stimuli or combinations were being experienced. S-habits certainly follow this rule. Another approach would make all lowest level stimulus values already known and familiar. But this would still end up with the same problem because level 2 stimuli would then be novel simultaneously. A significant result is that the single sense, single sensor scenario now requires A to repeat 4 times before a reaction occurs!
Have you ever done an action to get a result that you know was wrong just to see if it still gave the same result? Does this say that a novel action habit is repeated even though the goal object may be uninteresting? That is currently how Adaptron works. What if the action the first time caused no change and that is why the result is uninteresting? Would you still be interested in trying the action again? Exploration and what one intuitively feels about exploration would seem to say you would explore a no change action a second time to verify it. Apply this to the situation in which you are distracted by a totally novel noise - "Klunkityclang". Your attention is attracted to it and you listen to see if it repeats. You then hear it again (2nd occurrence). Does it become uninteresting at this point because it repeated or does it just become familiar? You now purposefully listen for it again. It then occurs for the 3rd time and thus it definitely loses its interest because it consciously repeated. But do you now have an orienting response that has only been done once and you wish to explore it to see if the result repeats? Thus you wait for the 4th occurrence before you pay attention to something else or perform a reflexive response.
25th April 2012 Properties versus objects
At a very abstract level a property is a type of object while a value is an object. We aggregate / combine values to more specifically identify the object and a type of object is an aggregation of its property types. Thus it would seem reasonable that the rules for combining parts into objects (slower novel combinations, immediate familiar combinations, ignore repetition) also apply for combining property values.
27th April 2012 Gap Patterns
I don't think I need a gap pattern as well as a shape pattern. The idea of the gap was to represent a width of the dependent sensor array that acted like a window in which 2 or more lines appeared which were not part of the object of interest. For example the object of interest could be a wire-frame through which any number of lines could be seen in the background. As the wire-frame expands or moves so do the gaps. But since the lines inside a gap are a shape at a level of complexity one or higher I could just create all shapes from all line combinations. This would mean, for example, a line combination of 2 lines could be combined with a 4-line combination to create a 2-part shape object. To do this my level-1 shapes would consist of all possible shape combinations except the total of all lines. Then the level 2 shapes would be any two of the level-1 shapes that were adjacent. Level 3 would then be ones with a common overlapping piece.
Levels of complexity
This idea may also solve the need to recognize a pattern independent of the level of complexity. Both contrast and shape patterns are level of complexity independent if they are applied where their lowest level parts are not necessarily at the absolute lowest level of the sensors. Contrast and shape patterns are level independent if they can be applied where their lowest level parts are adjacent but not at the absolute lowest level of the sensors. This is achieved by creating logical level 1 parts that are all possible combinations of adjacent sensor level 1 parts (lines). Then logical level 2 parts are all possible pairs of the logical level 1 parts not necessarily adjacent. Then level 3 and higher combinations require a common overlapping part.
24th May 2012 Generalization
I have cleaned up the P-Habit recognition algorithm such that P-Habits are formed from the object level stimuli (O-Habits - combinations of simultaneous stimuli (parts) from a single sense) before the object level stimuli are used in S-Habit formation in the STMs. This way the O-Habits do not become uninteresting, they are either novel or familiar for P-Habit formation.
Independent stimuli
I have previously been breaking down the pattern of symbolic stimuli from independent sensors into the independent properties of Position pattern (sensor numbers), Position order and symbol pattern. These 3 properties can change independently of each other. This would result in the following properties for the given patterns:
Source Pattern Position Pattern Sensor Order Symbol Pattern
At 1 2 3 4 5 6 7 8
A B C 1 5 6 1 5 6 A B C
A C B 1 5 6 1 6 5 A B C
B A C 1 5 6 5 1 6 A B C
B C A 1 5 6 6 1 5 A B C
C A B 1 5 6 5 6 1 A B C
C B A 1 5 6 6 5 1 A B C
However the sensor order is not appropriate because sensors are independent of each other and ordering them does not make sense. What is needed for the order pattern is the symbolic order. This is what really changes. Thus the new 3 properties are Position pattern, Symbolic order and Symbol pattern. This results in:
Source Pattern Position Pattern Symbol Order Symbol Pattern
1 2 3 4 5 6 7 8
A B C 1 5 6 A B C A B C
A C B 1 5 6 A C B A B C
B A C 1 5 6 B A C A B C
B C A 1 5 6 B C A A B C
C A B 1 5 6 C A B A B C
C B A 1 5 6 C B A A B C
Similarly for graduated readings from independent sensors the Shape/Sensor pattern should be replaced with the Shape/Order pattern. The order for the three sensor readings of 2 5 9 should be 0 3 7 and for 5 2 9 it should be 3 0 7. Both have a Contrast/Symbol pattern of 3 4. This is obtained as the difference in readings once they are sorted in ascending order.
25th May 2012 Dynamic structure
I'm trying to devise a dynamic structure for the creation of the two frames of experienced stimuli such that they are easily created, searched and linked to the STMs. The experienced stimuli are created in the order; O-Habits - line or sensor properties, combination of these properties per line or sensor, combinations of these lines or sensors, then P-Habits - combinations of lines across senses followed by S-Habits of all the previous stimuli. But only familiar ones are combined and novel ones are interesting.
Long term memory consists of a sequential list of binons. Each binon is of a particular type and has pointers to its two parts. It also has a pointer to a linked list of action habits for which it is a trigger stimulus. As stimuli are recognized the operation that is most frequently done is the search of this list for the binon given the type, two parts and a possible value. If the binon is not found it is created. Binons / parts do not have pointers to all the binons of which they are part. This could be added such that each binon has a linked list of binons for which it is a part.
The two frames of experienced stimuli are sequential lists of the binons found / recognized or created. Each entry contains pointers to the two entries used to create it, the experienced binon and an additional information binon pointer used to create the experienced binon. Each entry also has a STM list of past experienced stimuli for forming S-Habits and detecting repetition. These STM lists could be dynamically allocated and pointed to by the objects on the experienced stimuli lists. The STM lists are traversed once in order of smallest to biggest O-Habits and then smallest to largest P-Habits in order to find the most interesting stimulus. Longer S-Habits have precedence over shorter ones, larger P-Habits have precedence over smaller ones and larger O-Habits have precedence over smaller ones.
The most challenging problem is that of paying attention to the goal stimulus. Given an expected goal the experienced stimuli must be searched for this goal or the "nearest" stimulus at the same "location". This happens for the action habit being practiced as well as the subconscious / learnt action habits being done. Although in the latter case only an exact match is acceptable for the action habit to continue. This search is done on the STMs based on the "where" information of the expected goal. The "where"-information is the combination of the object / property types that are in the goal stimulus.
A possible solution is to use the master binon list for access to all other lists. Each binon would have pointers to the following linked lists:
- The action habits for which it is a trigger (already exist)
- The binons of which it is a part for faster finding or creation of parent binons
- The STM it is added to, based on its property / object / P-Habit type
But what would distinguish those binons that form the experience from those that don't? Ones that have been experienced would have an active STM. The STM would contain the latest binon experienced for the given property / object / P-Habit type. Ones that have not experienced a second stimulus in a row would be empty. An STM might not experience a second stimulus in a row because the two source stimuli were not familiar.
23rd June 2012 Where versus What
I have used a bitmap successfully to identify the type of objects. I am currently calling it the Where information but it isn't really. It identifies the kind / class / type of object. I need to create where information which is a combination of Sense, Position, Order, Reading (Intensity) and Size. These can all change while what information does not. What-information is a combination of relative things such as Sense (for symbolic readings), Symbol, Contrast, Shape, Negative, Reflection, and Rotation.
Motion
Motion is the change of where but not of what. It would seem to be captured in a STM for each what. The current STMs are for changes in what at a given where.
Paying Attention
When directing attention we can ask either "what is at a given where" or "where is a given what". Our attention is attracted to a given what at a given where.
Combinations of properties
Right now I create all possible combinations of the up to 8 properties for an object. These 8 properties are Shape/Order, Position, Size, Rotation, Contrast/Symbols, Intensity, Reflection and Negative. This means that combinations include strange pairs such as Size and Reflection or Rotation and Intensity. I think I should be combining the up to 4 properties determined from Position and then the up to 4 properties determined by Intensity and then combining these two results. Then I should add the independent sensor position and the sense. Is this similar to the division of visual stimuli into a where (position) and what (intensity) stream?
26th June 2012 Reflections of symbols
I need to correct the algorithm for determining the reflection status for symbolic readings from dependent sensors. This forms a symbolic pattern kept as the Contrast/Symbol pattern.
If I have 3 source symbol patterns A, B, and C each of which are at O-Level = 2 as follows:
Symbol pattern = A
Contains A Reflection
XY no
YX yes
Symbol pattern = B
Contains A Reflection
YZ no
ZY yes
Symbol pattern = C
Contains A Reflection
XZ no
ZX yes
And they are joined to form the following.
Pattern Left is a Right is a
No. Source Pattern Reflection? Reflection? Reflection?
1 AB XYZ no no no
1 BA ZYX yes yes yes
2 AC YXZ no yes no
2 CA ZXY yes yes no
3 BC YZX no no yes
3 CB XZY yes no yes
4 AA XYX no no yes
5 AA YXY no yes no
M 45 XYXY no no no
M 54 YXYX yes no no
If the two source parts are different objects then the reflection of the pattern is based on the order in which the two parts are combined. If the two source parts are the same object then when combined they form two different objects based on the reflection combination 'no'&'yes' or 'yes'&'no'.
29th June 2012 Where - Type - Source
The bitmap for the combination of Perception/Thought & Sense & Sensor & Property is currently being used to identify the type of the stimulus. But it might better be called source for the stimulus. The source properties are combined in the same order that the stimulus values are combined. Each stimulus comes from a particular combination of sources. This determines its type.
Complexity
Possible combinations of stimuli need to be far more flexible than they currently are. When forming P-Habits I should be able to combine a combination of lines at any complexity level and any property combination from one sense with any combination of lines at any complexity level and any property combination from another sense. Then S-Habits should be able to contain a series of any two stimuli at any complexity level, property combination and sense combination.
30th June 2012
See the 2012 Scientific Research and Experimental Development (SR&ED) tax credit claim.
Details about the Canadian Revenue Agency’s SR&ED tax incentive program are here.
3rd July 2012 Action Habits
Is it possible that when practicing an action habit and paying attention to what one is practicing that we can only think one step ahead in order to modify the action habit "on the fly"? We often do this when concentrating on what we are doing. We think of an alternative way of accomplishing the result and try some alternate "move". Note that thinking one step ahead can mean a large / long duration action habit because the acton in an action habit can be the top of a very large tree of actions. Does it also mean that continuous thinking cannot be done while practicing? To think ahead more than one step we need to stop concentrating on practicing an action habit because this uses up our attention cycles.
Parts and Interest
I have a situation in Adaptron in which the whole object has become expecting interest but the parts are uninteresting. Is this possible? At its creation the whole is a novel combination made up of the parts when they are both neutral. The next time the parts occur we recognize the whole and it becomes neutral. If the parts repeat they become uninteresting. Then if the whole gains an expectation of interest shouldn't the parts also gain this expectation of interest such that a partial match of them as a possible goal will cause the action to be taken?
4th July 2012 First Novel
I have to implement the idea of 24th April 2012 for non-orienting action habits as well. When practicing an A-Habit and a novel goal stimulus is perceived then the action habit is not created. The interesting nature of the goal stimulus is distracting and interrupts the practicing. And a novel action habit has to be practiced and succeed a second time before it is finally learnt. This minimizes the possibility of learning actions that result in random or often changing goal stimuli. These are situations in which there is less likelihood of a causal relationship between the action and the result (goal).
Thus creation and update of action habits must follow these rules.
Goal expected Interest in Goal Stimulus perceived Action-Habit
None (a reflex) Neutral or uninteresting (familiar) Create- interesting redo
None Interesting (novel) Not created
Exact match Neutral or uninteresting Update - neutral redo
Exact match Interesting - Impossible to be novel and match
Different goal Neutral or uninteresting Create - interesting redo
Different goal Interesting Not created
Can't find None Not created
This brings up the question: Does it really make sense to both create and then repeat an A-habit that gets an uninteresting goal? If the goal were unpleasant one would not repeat it.
If you get part of the expected goal but the other part perceived is different then you pay attention to the part that was different. That is the property of the perceived goal that was different from the expected goal. If the property is familiar and the whole is familiar a new action habit should not be formed with the different part as its goal. Instead a new S-Habit is formed with the different part in the sequence. The original A-Habit remains. A new A-Habit gets formed with the new goal when the sequence becomes familiar. For example, when I tap my brother on the shoulder he always gives me a red apple (colour property = red, shape property = apple). When I tap my sister on the shoulder she always gives me a green apple. Both types of apples are familiar to me. Then when I tap my brother on the shoulder he gives me a green apple. I pay attention to the colour property because it is different from what I expect. My action habit does not change; it continues to expect a red apple. But because the perceived (S-Habit) sequence: tap brother, green apple is novel this is what I remember. Then the next time I tap my brother on the shoulder and I get a green apple I remember this as a new action habit and I will want to practice it again to learn it. So the S-Habit caused by an action must be familiar before the action habit gets created. This would seem reasonable considering the cortex is continuously monitoring the sequences of stimuli caused by subconscious action habits being performed by the cerebellum. The cortex is executing all the familiar / known S-Habits looking for unexpected events just in case a subconscious action habit fails.
5th July 2012 Distraction / Interrupts
I'm trying to sort out distractions and interruptions of practice caused by a change. We may not notice a change because we are concentrating on what we are doing or thinking. This occurs even though the change may be some completely novel stimulus. But usually if the change is a value above the range of expectations (loud, bright etc.) then it attracts attention. So are there two cases here? Read 12th Jan 2012. We could have a stimulus of the type expected and it has a different value. We could have a stimulus of the type expected but it is novel. We could have a stimulus of an unexpected type. Or does every habit expect values from all sources? Or do we have many simultaneously subconscious habits, each of which is expecting something from a subset of all sources that are not overlapping. A very loud stimulus would interrupt one of them. These would have to be orienting responses (S-Habits) and they would have to be looping so that they are continuous. Do all A-Habits automatically start by looping and then it is learnt to terminate them / stop them and only do them once? Babbling seems to be a good example. They would stop when no change occurs, that is they do not cause a change.
On the 31st July 2010 I wrote: "If you are practising an action habit to learn it then an unexpected stimulus will not distract you unless it is part of the expected stimulus of the action habit." This is called "change blindness". However the second case is where the action habit is interrupted because attention is drawn to a totally unrelated stimulus. I think the idea that there are many subconscious habits executing in parallel each with different non-overlapping source expectations is the right approach. But an active action habit does not change the interest level in a stimulus. It just stops executing when it does not get its expected stimulus. What we really are describing is habituation, which is built in to every binon. It effectively stops firing when it detects a repeat of its stimulus. And if we have S-Habit binons for all known / familiar sequences then novel stimuli create novel S-Habit binons which attract attention. But we are blind to them if we are practicing unless they are part of the expected stimulus. I believe this practicing also includes the thinking that is being done to solve a problem, make a decision or devise a course of action.
6th July 2012 Incentive - Redo Interest
I've stopped recording S-Habits in which the second stimulus is novel and I have stopped recording A-habits in which the goal is novel. This leaves only recording reflexive actions in which the goal is familiar and repeating / practicing these in order to reduce their redo interest level. So now how is incentive meant to work? How is thinking going to work? When a new binon / habit is formed do the two parts gain interest because they were involved in the creation of an interesting combination or does the first one attract attention when it occurs because it is part of an interesting combination? For A-habits this is the later approach right now because the redo interest is the interest in the newly created A-Habit. This means the only incentive is to create new sequences through action. It is not the pursuit of interesting goals. This also means that thinking must traverse A-Habits using the redo interest as the criteria rather than goal interest. This would be true for interest as a motivation but for the pleasant / unpleasant motivation it may be the goal stimulus that is used. Interest as a motivator is all about creating new experiences and this is what the redo interest does. It is the interest in new sequences. Then comes the question, when pleasant / unpleasant goal stimuli exist and thinking takes place, is a pleasant goal stimulus thought about before or after an interesting redo level on the same A-Habit?
Distractions
So, while practicing a novel stimulus will not interrupt the practice because it is the same interest level as the practice, an emotional - pleasant /unpleasant stimulus will interrupt it.
7th July 2012 Action Habits
From today's test runs I think that the trigger stimuli for all actons and goal stimuli for all A-Habits are atomic (S-Level=1). This means they must be formed before (or in parallel with) the recognition of S-Habits in STMs. But what criteria are used to identify the appropriate atomic goal stimulus. It must be that a novel S-Habit can be formed from the trigger and the goal. I create A-habits for orient responses that correspond to S-Habits. Their actons must be able to habituate on the trigger stimulus. Thus it would seem reasonable to make the "do-nothing" act part of the acton a repeater act and these should be in Act1 of the acton.
10th July 2012 Incentive - Redo Interest
I now have A-Habits being created whenever a new S-Habit is created in any STM. This includes when a known S-Habit is created but a different acton was used to achieve the goal. But I still have this feeling that the pursuit of a novel goal stimulus can be the objective and not necessarily the creation of a novel sequence. A novel goal stimulus never creates a novel sequence because S-Habits are currently only formed when both stimuli are familiar. The reason is to avoid the premature creation of sequences that are coincidental, not causally related and therefore very unlikely to (will rarely) occur again.
A-Habit execution
A device needs an off state such that when it is receiving no instruction it is off. But it also needs an instruction to not move, i.e. a change of zero. At any one instance all the devices are receiving no instruction, a "don't move" instruction or a change amount signal. Having many A-Habits executing in parallel does this. The "don't move" instruction is the do nothing response and it needs to be sent every cycle when just paying attention is required. But given a trigger stimulus there may be many possible S-Habits each with a different goal stimulus. We don't want to be executing an A-Habit for every possible one. This is the job of the STMs. But if we have performed an action we only want its goal. Anything else in the goal stimulus's STM is unexpected for the given action. If the stimulus perceived is familiar a new action habit is formed that needs to be practiced to make it learnt.
11th July 2012 Pruning - forgetting
I could incorporate a function to delete a binon that was created from two familiar source stimuli and is novel but has not become familiar after so many X cycles. This would free up the space taken by randomly occurring, coincidental combinations of familiar stimuli.
Partial Matching
When attention is paid due to a distraction it is the longest / largest S-Habit / P-Habit with the highest interest level that is used. Then if this stimulus is neutral its redo interest is checked. Its two parts' redo interest is checked next. So the practice of the specialized whole's action habit is done before the practice of the more general parts' action habits. Also when doing reflexive responses they are done using the most specialized trigger object that has just become consciously boring. Thus action habits are first attached and practiced for higher level stimuli that are the more specialized objects.
Generalized versus specialized A-Habits
But maybe A-Habits should first be explored on the most general objects / concepts - lowest level stimuli. This would explore the existence of any actions that cause a consistent change in general before exploring the specific. Then there are more possible things that can be tried for the more specialized triggers because they are a combination of the more general stimuli. But it violates the fundamental understanding that you can only learn things of which you are conscious. And it is the specific stimuli that you are conscious of. Note however that you learn to recognize simultaneous combinations of stimuli without being conscious of it. But just as the idea of two source stimuli need to be familiar to combine into a novel one slows down the formation of higher level stimuli maybe there is a similar principle at work for action habits. Action habits must be tried on the lowest level stimuli first. These must be explored first. But what are the criteria for this process? Currently I have creation of the first actons as reflexive responses to boring conscious stimuli. Maybe it should be reflexive responses to boring stimuli but not necessarily conscious since lower level stimuli become boring before higher level ones. But then when to stop exploring the lower level ones, that is what criteria are used to determine when do they become permanent?
Exploration
With this new approach it should explore its body's internal senses (such as kinesthetic and proprioception senses) before the external senses.
12th July 2012 Deduction and Induction
Deduction is the process of assigning knowledge to a specific object based on knowledge about one of its properties. One starts with knowledge about an initial known fact or value of a type or property and is then presented with a specialized situation that contains this type or property with the known value. One then deduces the knowledge also applies to the specialized situation. For example one knows all types of objects that are swans are white. One is then told about a specific situation that involves Daisy and Daisy is a swan. One knows that Daisy will be white. Induction works the other way. Given many specialized situations in which something rewarding happens, one needs to figure out if there is a common fact or value of a type or property that always is associated with the reward. This is also called concept learning. If one always feels warm when seeing a blue animal one induces that it is the colour and not the type of animal that is the common factor.
Exploration
I am trying to attach A-Habits to the most general stimuli before the specific. But I have this notion that it is the most interesting highest level stimulus that attracts attention. And then we learn about these stimuli and explore them. Somehow I must explore the lowest level ones first. Does this mean changing the strategy for what attracts attention? Maybe a stimulus needs to become permanent (explored all possible A-Habits) before it is combined into P-Habits or S-Habits. This would slow down the creation of these higher level stimuli and attention would remain at the lowest levels until the A-Habit possibilities have been exhausted. Maybe that is why we have very little to no recollection of what it was like when we were 2 years or younger.
13th July 2012 Sequential familiarity
I believe the best way to slow down the creation of combinations of stimuli is to not only insist that the two parts be familiar but also insist they have sequential familiarity. That is the stimulus must be a trigger in a familiar S-Habit or A-Habit.
Uninteresting
I'm wondering if I need to use the concept of uninteresting. Right now, my STMs set a stimulus's interest to uninteresting when it repeats. The attention algorithm decides things are boring when the most interesting stimulus is uninteresting. Then a reflexive action is done. But wouldn't the same be accomplished if there were no interesting stimuli, all are neutral and thinking or partial matching produces only neutral results. Do we have to have the repetition of the conscious stimulus to be bored? Can't we just have no unexpected / novel / interesting stimuli and no ones to redo due to thinking to be bored?
14th July 2012 Biggest combination
I've been using the principle that the biggest combination of stimuli that are novel / interesting attracts attention. This includes S-Habits that are new. But maybe this rule needs modifying. The STMs create the sequences and it is unexpected sequences that they provide to attract attention. The stimulus that causes the unexpected sequence must be familiar but the sequence is novel. Maybe the rule is that as long as there are new combinations of simultaneous stimuli (O-Habits and P-Habits) the novel S-Habits don't have a chance to attract attention. Also note that when change occurs from one frame to the next attention is drawn to the largest simultaneous combination of those stimuli that were unexpected. So S-Habits are being used to highlight the second stimulus which was unexpected that caused the new S-Habit to be formed rather than the novel S-Habit formed.
Repeating Stimuli
When the same stimulus repeats simultaneously adjacent to itself on an array of sensors it produces the same stimulus but with a width reading. The same should happen for sequences of the same stimulus repeating. It should produce the same stimulus but with a quantity / value representing how many times it has repeated. The same should take place for O-Level stimuli above level 1 that are adjacent to each other.
15th July 2012 Unexpected
When bored we perform a reflexive response. STM recognizes the sequence or not and should set the unexpectedness interest of the goal stimulus. The unexpected stimulus attracts attention and the unexpectedness is reset for this stimulus. The A-habit is created for the reflexive response. If there is any A-Habit with a redo interest for the attended to stimulus then this is done in practice mode. When the next frame occurs practicing finds its goal and the redo interest is reduced. The STM recognizes the sequence and should set the unexpectedness interest. Do-nothing / orient responses are acts that request zero change be done by the device. Because you can get a variety of goals when you do nothing there are going to be many such A-Habits each with a different goal. Also since in the past you may have experienced several different goals for the same act there needs to be A-Habits for all of these with the same act. However when it comes to redo interest only the most recent is done.
S-Habits
If S-Habits are to be used to detect the unexpectedness or expectedness of stimuli that are caused by action habits and action habits are only to be created using conscious / attended to stimuli then S-Habits should also only be created using the attended to stimuli. Thus rather than having an STM per stimulus type I should have just one STM that is forming sequences of the attended to stimuli. When I start doing A-habits in parallel then I will need to be doing S-Habits in parallel.
16th July 2012 S-Habits and STMs
I have been considering strategies for recognizing unexpected stimuli using S-Habits and STMs.
For each stimulus on the Path() that has a list of A-Habits I could start an STM because every familiar stimulus on the Path() with at least one A-Habit has an expected next stimulus and if it occurs next we want it's unexpected interest to be neutral. Then go through all the active STMs and see which ones find one of their next expected stimuli.
Do I need a list of all S-Habits that have been consciously created for each trigger stimulus? Like a list of A-Habits - but disregarding the acton part. Or should I be executing S-Habits for all the possible outcomes of the A-Habit that is being done. Many outcomes if a do-nothing / no change acton and maybe only one if a do something acton. I need this list without the actons in it so that sequences caused by subconsciously executing actons are recognized and are not unexpected (are expected).
I could mark all the stimuli in a frame as unexpected and then each S-Habit recognized could mark their goal stimulus expected. This is the alternative to what I do now which is to mark the goal stimulus unexpected when a new S-Habit is generated in the STMs. The STMs know what the source is for forming their S-Habits because they look for sequences where the source of the goal is the same as the trigger. But S-Habits formed by paying attention to unexpected stimuli do not follow this rule. Thus the trigger - goal sequences in the A-Habit list must be used to recognize expected and unexpected sequences.
17th July 2012 Learning
Learning is the process of finding out what happens, G (goal stimulus) when action X is done in situation T (trigger stimulus). An unexpected G can be caused by a novel stimulus or an unexpected sequence of a familiar T and then a familiar G. The biggest combination of instantaneous stimuli that are unexpected forms the unexpected G. We pay attention to the unexpected G and it becomes the expected G for any current T and X. But what if the G is not unexpected (it's expected) and not novel (familiar)? Then we start thinking and checking parts of the biggest G for any A-Habits that need to be redone / practiced. If we decide to start an A-Habit that needs redoing using part of G as the trigger then our conscious STM must use that part as the conscious stimulus and not the whole G.
19th July 2012 Expectations
To determine the most interesting stimulus I need to combine the unexpected parts. I should not be determining object unexpectedness after forming them from their parts. Thus as each part occurs I need to know whether it was expected or not and then combine them appropriately. To know whether a part was expected it must be marked as such before it occurs. If it occurs and it was marked expected then mark it as found. Then it can be combined with other expected and found parts. Those stimuli that were marked as expected but did not occur are of no concern and can be unmarked before the next cycle. Those stimuli that occur but were not marked expected are then unexpected and these can be combined together. This process of marking stimuli as expected is the first time I have come across the equivalent mechanism to inhibition.
20th July 2012 Attracting Attention
I now have 3 interests that are added together to determine the most interesting stimulus.
- The novelty of the stimulus
- The unexpectedness of the stimulus
- The redo interest of an A-Habit for the stimulus as a trigger (primitive thinking)
But it brings up questions of priority. All other things being equal such as level of complexity what combinations of interest should be most interesting? One constraint is that if a novel stimulus occurs it will also be unexpected. But it will not be combined with any familiar stimuli that were unexpected. Thus there will be a combination of familiar unexpected stimuli and then zero or more novel unexpected ones that will not have been combined. I believe the largest novel unexpected stimulus should be more interesting than the largest familiar unexpected stimulus. Therefore adding novelty and unexpectedness seems reasonable to determine overall interest.
However adding in the redo interest makes it more complicated. If there is a stimulus with an A-Habit that is worth practicing then it has redo interest. It is not possible for it to be novel but it maybe unexpected. The question then is a novel unexpected stimulus more or less interesting than an unexpected redo interest stimulus?
22nd July 2012 Parallel Action Habits
I have been thinking about how to create the tree of action-habits and actons for subconscious execution. I have realized that when the stimuli involved are from internal senses there is usually only one action to be taken in response to a given situation and usually only one result / goal stimulus. Thus these actions are worth learning first. Consistent results occur until a traumatic accident causes you to have to go through physiotherapy and relearn them. However when the stimuli come from external senses there can be many possible results from any action. The external world is less consistent and predictable. Thus all the learnt action habits that contain the same trigger and action must be started in parallel so that all the possible result / goal stimuli are expected. The actons are saying one or the other stimulus may occur.
Also there is a do nothing action required. It is the relax action. It is the turn off device action. And it must be distinguished from the do no change action which is being done in order to hold the device in a steady position. The do no change action habit is the one that will continuously restore the device to the given position whenever it is perturbed by an external influence.
All the action habits that have been started and are running subconsciously must also have started their possible streams of expected result stimuli. We cannot have any expected stimuli from non-executing action habits. The question is; are all these streams being expected by S-Habits or A-Habits? I feel that S-Habits should be set-up for this purpose. This is because we recognize a tune like Happy Birthday whether we are playing it ourselves or someone else is playing it.
Practice speed
When we are trying to learn a new action sequence we do it first at a very slow speed combining less complicated steps in series. We then practice the series again at a little faster speed. This establishes it as a learnt action habit with a specific identity. We then can start it and perform it subconsciously.
23rd July 2012 Parallel S-Habits
When we activate an acton it starts many possible actons in parallel so that all manner of disturbances can be handled. We also want to disassociate the subconscious acton performance from the subconscious sequence recognition. Therefore we must also start many possible S-Habits in parallel to set the expectedness of the subsequent stimuli. This means I need to rethink the representation of the S-Habit structure to handle the 'OR' between multiple S-Habits. I need to aggregate S-Habits if they occur in parallel or if one trigger & response combination can produce multiple result/goal stimuli.
31st July 2012 Levels of Complexity
This is a follow on from the 27th April 2012. I feel I need to separate shape information from contrast information and only bring them together at the last minute. The solution to recognizing the shape of objects independent of their level of complexity is definitely to group adjacent objects and treat them as logical level-1 objects as though they had no content or edges within them. However the fact they do have edges within them gives them a contrast pattern that is not at level 1. This should also solve the Gestalt perceptual grouping problem. The logical level 1 objects that contain edges should only be grouped at logical level 2 if they are adjacent. They should not have gaps between them because gaps are handled at level 3 by the overlapping logical level 1 part.
I've done some more experimenting with Perceptron2 only recognizing shapes and shape combinations and I realized I need to recognize the patterns of contrast objects for adjacent contrast objects and these then are treated as the logical level-1 shape objects. These logical level-1 shape objects can then be combined even if they are not adjacent. This is because relative contrast values within a shape change in unison or don't change together. Or from another perspective things move more often than the lighting changes.
Binon content
I've just realized that the binons should just contain the combination of Shape/Order and Contrast/Symbol. They should not contain sensor position, size, reflection, intensity/reading, rotation or negative/positive. These are all absolutes. Objects are all independent of these values. Only the current experience holds these in the Path(). However if one of these absolute values changes for a given object then we have something worth keeping in an S-Habit binon.
1st August 2012 Binon structure
For independent sensors and senses the combinations of parts needs to be all possible combinations. This ends up with all binary numbers where each bit represents a sensor or sense. My algorithm with the 2nd layer containing all possible pairs and the 3rd and higher combining pairs with common overlaps achieves this. However for dependent sensors the combination of adjacent parts must be used at level 2 and higher. Non-adjacent parts are only associated / dependent on each other if they change in unison. But even then the part between them changes in unison with them. Independence of level of complexity is obtained by creating at level 1 all possible objects between all possible edges. Edges are wherever the intensity is different between adjacent sensors.
Shape rotation
I also have to sort out what unique shapes to produce based on the rotation of the parts. For example with shapes of width 1, 2 and 5 the following are the different shapes. A zero rotation means it is symmetric; the sizes of the two parts are equal.
No rotation rotated parts' rotations shapes' rotations
Parts are the same shape
1st shape 1,1,1 = 1,1,1 0, 0 = 0, 0 0 and 0
2nd shape 1,2,1 = 1,2,1 +, - = +, - + and +
3rd shape 2,1,2 = 2,1,2 -, + = -, + + and +
Parts are different shapes
4th shape 1,2,5 = 5,2,1 +, + = -, - + and -
5th shape 2,2,5 = 5,2,2 0, + = -, 0 + and -
Or = 5th 2,2,1 = 1,2,2 0, - = +, 0 + and -
6th shape 1,5,2 = 2,5,1 +, - = +, - + and -
7th shape 2,1,5 = 5,1,2 -, + = -, + + and -
Alternating widths at level 4 produce the
4th shape 1,2,1,2 = 2,1,2,1 +, + = +, + + and -
6th August 2012 Reuse of Binons
I realized yesterday that any binons learnt for a particular sense with a fixed number of sensors could easily be attached to any other sense and sensors and still carry on learning. Even if it was the same sense but just a larger or smaller number of sensors added or taken off the same binons would still be useful. This is because everything they record is relative.
7th August 2012 Levels of Complexity
I've been devising a new algorithm such that levels of complexity processing only does perceptual grouping when it gets to the higher level rather than producing all possible groupings at level 1. The idea is as follows:
When we have a level of objects either from the sensors (one object per sensor) or from just combining the previous levels objects start by treating them as though they are level 1 of complexity. This means taking each adjacent pair (not sensor adjacent but next to each other) of objects in the list and combining them if they are the same object and size. At sensor level they must also have the same value because this is where edges are used. Rotation is not important if they are the same objects because they will be symmetric in shape. The pair of objects may be sensor adjacent or have an overlap. If they overlap the overlap part is the same as the parts of both objects because they must be symmetric. This is done repeatedly replacing the pairs with the new objects. Then this list must be traversed to find all the objects at this level that are sensor adjacent and these adjacent combinations must be produced as though they were sensor level 1 objects. However this will produce shapes in which the two parts are between edges separated by 2, 3 or 4 etc. objects. It does not combine shapes for example where the first object contains 2 parts and the second object contains 3 parts.
The next thing is to produce the next higher level of complexity. Here adjacent in list objects are combined. From level 1 to level 2 these will also be adjacent in sensors. These will always have an overlap from level 2 up.
I don't believe I need to combine parts that have a separation until I have motion, i.e. two frames in which a change in location can be detected for some parts. Then the separated parts that change in unison can be detected.
8th August 2012 Coincidence / Synchronicity
The principle of coincidence is fundamental to the combination of binons. The rule "Cells that fire together, wire together" is based on it. So is the rule "All the parts of an object change in unison" or "All the parts of an object maintain constant relationships". Another rule is "Closer parts (adjacent / overlapping) are more likely to be part of the same object"
10th August 2012 Levels of Complexity
I've been trying to figure out the combination tree in more detail. At the start of each level it is assumed that we have a list of overlapping and adjacent objects with position, sizes and rotations. The 1st thing to do is to go through the list and repetitively combine the same shapes. They must be the same object (value if at the sensor level), the same size and the same rotation. They can overlap but they must be symmetric to do this. Otherwise non-symmetric objects will be adjacent. If adjacent non-symmetric objects are combined in this way any objects made up of their parts between them (it will overlap with their parts) must be removed from the list. ( Inhibited! ) This is done repetitively until no two adjacent objects are the same. For example: given the following list of shape objects the resulting shape object results. The object format is rotation sign Object Id [Size].
Objects given Objects resulting
-23[7], +15[2], +15[2], +15[2], +15[2], +24[5] -23[7], +15[8], +24[5]
-23[7], +12[5], -12[5], +12[5], -12[5], +12[5], +24[5] -23[7], +12[15], +24[5]
The 2nd thing to do is to add two level-1 shapes from all objects that contain a common shape. This common shape's size plus the left part's size must be used to form a level-1 shape of the combined size. The common shape's size plus the right part's size must be used to form a level-1 shape of the combined size. These new level-1 shapes include one or more edges that were being used for combining shapes but are now to be disregarded. This process should produce shapes that will provide for perceptual grouping. These new level-1 shapes and all previous level-1 shapes are combined provided they are sensor adjacent in step 3 below. Effectively these higher order groupings being treated as level-1 shapes are delayed in their creation until their internal shape has been recognized at the higher level.
The 3rd thing to do is to go from level 2 up and combine the new shapes from the level below with whatever shapes were already at the level below using adjacency between level-1 objects and overlap (having a common part) of level 2 and higher objects.
11th Aug 2012 Levels of Complexity
The 2nd thing to do could be simplified and the level-1 shapes could be produced from the shapes at the level produced before the 1st thing to do. Thus after level X objects have been created in step 3, step 2 would use the level X objects position and size to produce the new level-1 shapes. This would produce them one level earlier than the previous approach.
I've decided that although binons have a level of complexity based on how many parts they are composed of they are not identified based on their level. However experiences need to be identified based on their level of complexity. Thus xxoxxoxxo is 3 lots of xxo while xxxxxxooo is the same shape but only occurs once. It's the same binon / shape object but is experienced at different levels of complexity.
15th Aug. 2012 Repeater
I've been using bucketed sizes and have added a repeat property for the patterns that repeat in order to represent this information. However I think what I should be doing is using the bucketed value of the repeat count at each level as the size for whenever a shape repeats at that level. I have also realized that I need only store the difference between the two bucketed sizes since the bucketed size is the Log based 2 of the actual size and the difference of two logs is the ratio of the values before the logarithm.
20th Aug 2012 Repeated Patterns
At level 1 if the sensor values repeat then they are adjacent and can be combined immediately. At level 2 if the shapes alternate between two widths every second pair is the same and adjacent. At level 3 if the shape ratios repeat then every group of 3 is the same and adjacent. The pattern of repetition can always be detected at level n by comparing each nth entry provided they have been created using overlaps. For example if ABC repeats then at level 1 the pattern is ABCABCABC etc. At level 2 the pattern is AB BC CA AB BC CA AB BC CA AB etc. At level three it is ABC BCA CAB ABC BCA CAB etc. At this level every 3rd entry is a repeat and adjacent and this criterion can be used to detect strings of repeated patterns. Notice that at level 2 every 3rd entry is also a repeat; however they are not adjacent to each other.
25th Aug 2012 Repeated Size
The number of times a pattern repeats is its size. Thus the size property of the binon should not be kept as the O-Level but as one of the Value attributes of the binon as is currently done at O-Level 2 binons.
31st Aug 2012 Motor control
The common principles that keep-on re-occurring in motor control appear to be: The difference between the goal and actual stimulus is used to determine the next output response to reach the goal. The device being controlled is identified - is this the same concept as the desired action being identified? The device ID and a signal / response value is required. This is the same as an action ID and a goal value pair. A Central Pattern Generator (CPG) is an identified action that could be sent a response value to give it a faster or slower frequency or strength response. If an action is a repeater such as a CPG then it can be sent at least two values to control it. One is frequency and the other is extent / amplitude. Is a repeating action comprised of two more primitive actions? Is each waiting for the stimulus/cause by reaching the end of the travel? If so an increase in frequency translates into a speed of travel for the primitive actions. The amplitude of the repeating action translates into the goal of the primitives, the stimulus that determines the end of one primitive and the trigger for the second. This implies that even the primitive action needs two values sent to it to describe the action required. A goal and speed / strength. But if the action does not check if it has reached its goal only the strength is needed. In fact the control logic at the next level up is the one that is checking the goal. The action just has to perform to the required strength. The control remains active even when the goal is reached because it is asked to obtain the goal. Even when external perturbations occur it must react and stabilise the device at the goal position.
20th Sept 2012 Intensity measures
At the sensor input level we might have intensities of 333222221177. The widths are all 1111111111111. At the next level the repeating adjacent intensity & width P-Habits are produced and the width values are 3422. 4 because log-base 2 of width 5 is 4. The intensity of these 4 binons is 3217. At the next level the differences in the width values are used for the width binons and these are 4-3,4-2,2-2. The intensity binons at this level need to be given intensity values that are the log-base 2 of the differences in intensity. These are Log (3-2), Log (2-1), and log (7-1). I think from here on the intensity binons are combined to form unique combinations representing contrast / texture patterns but the intensity value does not need to be kept like the size / width value needs to be obtained due to repeating shapes.
21st Sept 2012 Intensity measures
The solution is to use addition at level 1. The three 3s should be added together, the 5 2s added together etc. so that the total luminosity intensity is obtained. Then take the log based 2 value of this sum. So at level 2 the intensity readings are log(9) = 5, log(10)=5, log(2)=2 and log(14)=5. The 3 3s and 5 2s look about the same brightness because of the sum of their intensities. At the next level the differences between these log(intensity) values can be used just as the differences in width values. A binon must only and always maintain a relative value. It must always represent a relationship. It must never contain or represent an absolute value else it is not independent of things such as size, position, reflection, orientation, negative, complexity, and intensity.
At the 1st level of binons no relationships / differences are calculated because the raw values from the sensors are being used. Width (value 1) & Intensity binons are being created and combined into level 1 P-habits. These P-Habits are then combined if adjacent ones are the same. This creates level 2 binons for widths greater than 1 using the log of the total width and intensity binons with the log of the sum of intensities of these adjacent sames. Then these level 2s are combined into a level 2 P-Habit. Level 2 P-Habits that are adjacent sames are now combined. Level 3 and higher now use the differences / relationships between the level 2 width and intensity binons. Level 3 width relationship binons and level 3 intensity relationship binons are produced. These 2 types get combined into level 3 P-Habits. When adjacent same P-Habits are combined the new width is based on the log of the count of the P-habits involved and the intensity is based on the log of the count times the intensity which is the log of the sum of the intensities of the P-Habits involved. So Weber's law (logs) gets applied at every level when adjacent sames are combined. Adjacent same criteria must be not only the same binon adjacent but also the same orientation, size, reflection, intensity and negative. Complexity is the same because the binons are the same.
22nd Sept 2012 Level of complexity independent
Complexity level is a measure of the number of parts that make up the object. As such it is an absolute value. But binons are not meant to contain absolute values so binons should not specify their absolute level of complexity just their relative level of complexity. This happens by default since each says it is a combination of its two parts. It also means that each binon once created or recognized should also be used as a sensor which is providing a width and intensity value. The absolute values of this information should be used from the active path. This will result in recognition of a gap without the parts that comprise it.
Perceptual grouping
Making binons independent of level of complexity solves the perceptual grouping problem. Examples of shape perceptual grouping are that we recognize elevator floor number displays even though the numbers are comprised of separate big dots. This is shape perceptual grouping. An example of objects that are only intensity / contrast objects are textures and patterns such as wood grains. We can also experience perceptual grouping in the intensity domain by seeing wood grain patterns comprised of smaller granular patterns. Seeing shapes in clouds is an example. Also a line that traces the edge of a shape as in cartoons requires perceptual grouping. In nearly randomly positioned dots we can pick out shapes such as in the spotted dog on the spotty background and the dots that they put on peoples joints and follow with video cameras to produce stick figure choreography displays. Independence of level of complexity solves the problem because it uses all combinations of edges, not just adjacent ones, to determine the shapes. At higher levels of complexity more intermediate edges end up being ignored / skipped. Thus as more dots are involved and the closer they become to pure random scattering the less likely perceptual groups are recognized. But once a particular group has been found it is found faster the next time. This implies that a binon representing this pattern is formed. An important fact is that the object / pattern to be recognized must already be known / familiar for it to be found in a near random pattern.
3rd Oct 2012 Perceptual Grouping
To address this from a design perspective I need to rethink the combinations formed. The level becomes a property of the experience and should not be used to group the combinations for processing purposes. If there are two different values A and B from two adjacent sensors then there is only one combination binon. If there are three different values A, B and C there are eight combination binons. The adjacent ones, AB and BC and the overlap one (AB)(BC). Then there is the perceptual grouping ones (AB) C and A (BC). Using differences this can be extended for a larger numbers of sensors.
# Sensors # Binons Minus Sensors Diff Diffs of Diffs
1 1 0 0 0
2 3 1 1 1 1
3 8 5 4 3 2
4 20 16 11 7 4
5 47 42 26 14 7
6 99 93 51 25 11
If I use two lists of experiences (Paths()) and flip between them as each pass is done at a higher level of complexity one list will be the source and the other the new one produced. All the experiences that start at position 1 will be 1st in the list. Then will come all the experiences starting at position 2 etc. Each one will have its complexity level as a property. Given each one in the source list once the repeaters have been removed I will search for the next in the list at the same level to produce overlap binons. Then I will search for the next adjacent one at each of the lower and current levels. Each time I will produce the new combined experience and place it in the other non-source experience list.
The possible combinations for four sensors with values of A, B, C and D, whose values are such that they are all different produces the following 21 experiences. Brackets indicate the source pattern has been used as though it was a level-1 object. Only its position and size have been used. Dashes indicate where the two parts overlap.
Entry# Seq Src1# Src2# Level Pos Size Adjacent PG=Perceptual Group
1 A -1 -1 1 1 1 Overlap A
2 B -1 -1 1 2 1 B
3 C -1 -1 1 3 1 C
4 D -1 -1 1 4 1 D
5 AB 1 2 2 1 2 A A B
6 BC 2 3 2 2 2 A B C
7 CD 3 4 2 3 2 A C D
9 ABC 1 6 2 1 3 A PG A (BC)
8 ABC 5 3 2 1 3 A PG (AB) C
11 BCD 2 7 2 2 3 A PG B (CD)
10 BCD 6 4 2 2 3 A PG (BC) D
17 ABCD 5 7 2 1 4 A PG (AB) (CD)
12 ABC 5 6 3 1 3 O AB-BC
13 BCD 6 7 3 2 3 O BC-CD
15 ABCD 1 10 2 1 4 A PG A ((BC) D)
18 ABCD 1 11 2 1 4 A PG A (B (CD))
16 ABCD 8 4 2 1 4 A PG ((AB) C) D
19 ABCD 9 4 2 1 4 A PG (A (BC)) D
22 ABCD 9 10 3 1 4 O A (BC)-(BC) D
20 ABCD 12 4 2 1 4 A PG (AB-BC) D
21 ABCD 1 13 2 1 4 A PG A (BC-CD)
14 ABCD 12 13 4 1 4 O ABC-BCD
23 ABCD 5 11 3 1 4 O A B-B (CD)
24 ABCD 8 7 3 1 4 O (AB) C-C D
Note that 15 and 18 are the same because when (BC) D and B (CD) are treated as level 1 objects they are the same thing. Similarly 16 and 19 are the same thing.
Note that 20 is also the same as 16 because the AB-BC is used as a level 1 object. Similarly 21 is the same as 15. I think it will be easier to control the creation of these by always combining a part that contains an overlap such as 12 with an adjacent part from the level below.
If these changes are made and the list is reordered for production from top to bottom based on the level and what is available in the previous pass/list then the result is as follows.
Entry# Seq Src1# Src2# Level Pos Size Adjacent Perceptual Group
1 A -1 -1 1 1 1 Overlap A
2 B -1 -1 1 2 1 B
3 C -1 -1 1 3 1 C
4 D -1 -1 1 4 1 D
The 1st list will contain: 1 2 3 4
5 AB 1 2 2 1 2 A A B
6 BC 2 3 2 2 2 A B C
7 CD 3 4 2 3 2 A C D
The 2nd list will contain: 1 5 2 6 3 7 4
9 ABC 1 6 2 1 3 A PG A (BC)
12 ABC 5 6 3 1 3 O AB-BC = ABC
8 ABC 5 3 2 1 3 A PG (AB) C
17 ABCD 5 7 2 1 4 A PG (AB) (CD)
11 BCD 2 7 2 2 3 A PG B (CD)
13 BCD 6 7 3 2 3 O BC-CD = BCD
10 BCD 6 4 2 2 3 A PG (BC) D
The 3rd list will contain: 1 9 5 12 8 17 2 11 6 13 10 3 7 4
21 ABCD 1 13 2 1 4 A PG A (BC-CD)
22 ABCD 9 10 3 1 4 O A (BC)-(BC) D
23 ABCD 5 11 3 1 4 O A B-B (CD)
14 ABCD 12 13 4 1 4 O ABC-BCD = ABCD
20 ABCD 12 4 2 1 4 A PG (AB-BC) D
24 ABCD 8 7 3 1 4 O (AB) C-C D
The 4th list will contain: 1 21 9 22 5 23 12 14 20 8 24 17 2 11 6 13 10 3 7 4
The rules for combining parts are:
Level-1 and level-2 list entries are considered to be overlaps or adjacents in these rules.
A pure overlap is one that contains no adjacent combinations anywhere lower down.
- Always combine parts that have an overlapping piece and they are at the same level of complexity, creates the Os. One of the two parts needs to be an O. One of the two parts must come from the previous pass.
- Combine parts that are adjacent, provided that one of them was created as a pure overlap in the previous pass, creates the As. The part created is always a level-2 object.
- When a pure overlap part is used in an adjacent combination it is used as a level-1 part.
I need to check this with 5 sensors
Entry# Seq Src1# Src2# Level Pos Size Adjacent Perceptual Group
1 A -1 -1 1 1 1 Overlap A
2 B -1 -1 1 2 1 B
3 C -1 -1 1 3 1 C
4 D -1 -1 1 4 1 D
5 E -1 -1 1 5 1 E
The 1st list will contain: 1 2 3 4 5
6 AB 1 2 2 1 2 A/O A B
7 BC 2 3 2 2 2 A/O B C
8 CD 3 4 2 3 2 A/O C D
9 DE 4 5 2 4 2 A/O D E
The 2nd list will contain: 1 6 2 7 3 8 4 9 5
10 ABC 1 7 2 1 3 A PG A (BC)
11 ABC 6 7 3 1 3 O AB-BC = ABC
12 ABC 6 3 2 1 3 A PG (AB) C
13 ABCD 6 8 2 1 4 A PG (AB) (CD)
14 BCD 2 8 2 2 3 A PG B (CD)
15 BCD 7 8 3 2 3 O BC-CD = BCD
16 BCD 7 4 2 2 3 A PG (BC) D
17 BCDE 7 9 2 2 4 A PG (BC) (DE)
18 CDE 3 9 2 3 3 A PG C (DE)
19 CDE 8 9 3 3 3 O CD-DE = CDE
20 CDE 8 5 2 3 3 A PG (CD) E
The 3rd list will contain: 1 10 6 11 12 13 2 14 7 15 16 17 3 18 8 19 20 4 9 5
21 ABCD 1 15 2 1 4 A PG A (BC-CD)
22 ABCD 10 16 3 1 4 O A (BC)-(BC) D
23 ABCDE 10 17 3 1 5 O A (BC)-(BC) (DE)
24 ABCD 6 14 3 1 4 O A B-B (CD)
25 ABCDE 6 19 2 1 5 A PG (AB) (CD-DE)
26 ABCD 11 15 4 1 4 O ABC-BCD = ABCD
27 ABCD 11 4 2 1 4 A PG (AB-BC) D
28 ABCDE 11 9 2 1 5 A PG (AB-BC) (DE)
29 ABCDE 12 18 3 1 5 O (AB) C-C (DE)
30 ABCD 12 8 3 1 4 O (AB) C-C D
31 ABCDE 13 20 3 1 5 O (AB) (CD)-(CD) E
32 BCDE 2 19 2 2 4 A PG B (CD-DE)
33 BCDE 14 20 3 2 4 O B (CD)-(CD) E
34 BCDE 7 18 3 2 4 O B C-C (DE)
35 BCDE 15 19 4 2 4 O BCD-DCE = BCDE
36 BCDE 15 5 2 2 4 A PG (BC-CD) E
37 BCDE 16 9 3 2 4 O (BC) D-D E
The 4th list will contain: 1 21 10 22 23 6 24 25 11 26 27 28 12 29 30 13 31 2 32 14 33 7 34 15 35 36 16 37 17 3 18 8 19 20 4 9 5
38 ABCDE 1 35 2 1 5 A A (BCDE)
39 ABCDE 21 36 3 1 5 O A (BC-CD)-(BC-CD) E
40 ABCDE 22 37 4 1 5 O A (BC)-((BC) D-(BC) D)-D E
41 ABCDE 6 32 3 1 5 O A B-B (CDE)
42 ABCDE 24 33 4 1 5 O A B-(B(CD)-B(CD))-(CD) E
43 ABCDE 30 19 4 1 5 O (AB) C-(CD-CD)-DE
44 ABCDE 11 34 4 1 5 O AB-(BC-BC)-C (DE)
45 ABCDE 26 35 5 1 5 O ABCD-BCDE = ABCDE
46 ABCDE 26 5 2 1 5 A (ABCD) E
47 ABCDE 27 9 3 1 5 O (ABC) D-D E
4th Oct 2012 Levels of Binons
Level-1 binons have two null source binons and retain the identity of the object measured/detected by the sensors. Level-2 binons point to two level-1 binons and retain the relative ratio of the values the level-1 binons represent. Level-3 and higher binons point to level-2 binons and represent the possible combinations of 3 values / binons 2 levels below. There are seven types of these binons. These are documented in the 1st August 2012 notes. They result from rotations / reflections of the 3 values that produce the same pattern.
5th Oct 2012 Levels of complexity
As soon as two objects that are adjacent are combined without overlap as they are in the scheme above it seems to me that they must become level-2 objects with a ratio of sizes and intensities. The absolute size/width is the total size. The absolute intensity is the sum of the intensities. However when two objects are combined because they repeat and are adjacent do I create a level-1 object?
11th Oct 2012 Generalization / Specialization
When a frame is processed if all binons are created in which the two source binons are novel one ends up with the highest level binon capturing the most accurate representation of the object / pattern in the frame. This is the most specialized pattern. The most general is at the smallest / lowest end of the tree where there are level-1 and level-2 binons. The most specialized binons are over-fitted for purposes of identifying a class of objects. They identify specific instances of objects. The most general are too vague because they are used as parts of many different classes of objects. They are also the levels in which noise is being processed. The best level is the one that is the lowest level in which the binon is used in only one class of object. It is the pattern that is used in all objects of the class but not in any other class. It is the combination of properties that allows one to uniquely identify the class.
12th Oct 2012 Perceptual Grouping
I've been modifying the rules I started on the 3rd Oct.
13th Oct 2012 Perceptual Grouping
To make the rules even more specific, because I am trying to deal with the situation when the two adjacent or overlapping objects are the same. The rules for combining parts are:
Definition: A pure overlap is one that contains no adjacent combinations anywhere lower down or is a level-1 or level-2 object. They also have the property that their level of complexity equals the pass that created them.
- Combine parts that have an overlapping sub-part. This means they must be at level complexity 2 or higher. The created object will be at the next level of complexity. The two parts must be at the same level of complexity. One of the two parts must come from the previous pass. The other part may come from any pass except the 1st. They do not have to be pure overlaps.
- Combine parts that are adjacent. They must both be pure overlaps. One of them must have been created in the previous pass. The parts are used/treated as though they are level-1 objects. This means the level-1 pattern they are equivalent to is found and used as the part. The one created is always a new level-2 object composed of the two level-1 parts and is a pure overlap based on these parts.
To deal with repeated objects that are adjacent we need to enhance these rules. When an object has an internal structure/pattern and it is repeated adjacently several times we want to recognize all of the repeats as a single object earlier than having to wait for the pure overlap version of it for using/treating it as a level-1 for adjacency purposes. We also want to stop the higher level pure overlap from forming out of all of the parts. To do this at the start of each pass we need to go through the objects created in the previous pass and look for adjacent and overlapping versions of the same object. To be the same it needs to be the same type of object such as a width, intensity or combinations object. Two combination objects that are the same will have the same width and intensity pattern. When two adjacent same objects are found they are combined using the rules above for adjacent combinations. The level-2 object created will be symmetric in its parts and be a pure overlap like all level-2 objects. When two overlapping same objects are combined the result will be the next level up pure overlap symmetric object. The one thing that has to be done when the adjacent same parts are combined is the object contained inside the combination of the same type must be removed from the experience so it is not used to form higher level pure overlaps. Mind you this also removes them from being used in perceptual grouping.
14th Oct 2012 Repeating patterns
I have been wondering whether repeaters of a shape should be combined even if the contrast does not repeat or vice versa. I have concluded that the shape and contrast whole object must repeat to be combined as the same object.
16th Oct 2012 Perceptual grouping
I have the repeater objects being compressed appropriately but am having some conceptual difficulty with combining two non-similar combine/experience objects together. When calculating the total intensity of two parts that are being combined because they are adjacent and I want to use the combination as a level-1 object for perceptual grouping I need to add the readings and then take the Log of the total. I also have to add the two sizes and take the Log of that. I can do this with the absolute values but it would be nice to be able to do it with the Logs
Levels
When I take a pure overlap and treat it as a level-1 for perceptual grouping maybe what I should do is create the level-1 equivalent of the pure overlap at the same time as creating the overlap. Then it would be available as a level-1 object. This would mean two level-1 objects A and B would form a level-2 difference object A-B and a level-1 sum object A+B. And when A=B the two objects are combined into A+A if they are adjacent of if they overlap properly.
So if I have A, B, and C then I create level-2 A-B and B-C and level-1 A+B and B+C. At the next pass I create level-3 A-B-B-C = ABC and the two level-2s A-B+C, and A+B-C. I also need to create A+B+C at level-1. I have the rule that only the ABC should be used for this purpose. Maybe it should be that the A+B+C is created at the same time as the ABC but using the A+B and B+C created at the same time as the A-B and B-C.
Level-1s are always indivisible. They have a width and reading but have no parts. A, B, C, A+B, B+C and A+B+C are examples. Level-2 binons always are formed from two level-1s and contain the ratios. Level-3+ are always patterns formed from level-2s or other level-3+s at equal levels.
17th Oct 2012 Perceptual Grouping
Perceptual grouping is a phenomenon of shape and not of reading / intensity / contrast. Adding intensities together for adjacent objects does not make sense but adding widths together does.
20th Oct 2012 Summary of Facts
This is a summary of the pattern classification facts I currently understand.
Level-1 binons are sensory level and take a graduated or symbolic stimulus as input. They convert graduated linear scaled values into logarithms. If the sensors are dependent then adjacent same objects are combined into a single experience which means it could have a width greater than 1. Level-2 binons capture the relationship between the outputs of two level-1 binons. For graduated stimuli this relationship is the ratio and is represented by subtracting one input value from the other. This results in shape for width values and contrast for reading values. If there is a dependency between the sensors then a negative result means we have a reflection or a rotation and only adjacent sensors / level-1 binons are processed. But also the two adjacent level-1 width experiences are added to produce a new level-1 shape experience. With independent sensors no position or width exists and rotations and reflections cannot be determined. Instead of shape there is order and instead of contrast there is a symbol pattern. At level-3 inversion of contrasts are produced and they are different from reflections of contrast patterns. If values are symbolic then there is no reflection or inversion (negative). Combining two level-2 binons that share a common subpart produces all level-3 binons. Higher levels must also follow this rule. A shape object is the combination of a shape binon and rotation binon. A contrast object is the combination of a contrast binon, a reflection and inversion binon. An object is the combination of a shape and contrast object. Adjacent or overlapping objects are collapsed at any level into a single level-1 object before being combined with any other adjacent objects.
Experiences consist of an object plus the values for position, width, and reading. For the two dimensional image perceived by the human retina we would have to have 4 2D planes; one for black and white and 3 for each of the colours detected by the cones. Then each 2D plane needs two independent sensor properties, one for horizontal and the other vertical. So given any position there are 8 dependent sensor properties each with a graduated reading.
Recognition confidence reaches 100% when a binon is found in the recognition process that is only part of one higher level binon. As soon as it is part of two or more binons then it is shared between objects. Both objects contain the same part.
Perceptual grouping is only done on shape objects. There are rules for combining level-2 shape binons with rotations to form level-3 binons with rotations. There are rules for combining level-2 contrasts with reflections and inversions to form higher level binons.
23rd Oct 2012 Reflections versus Negative images
On the 24th March, 2012 I came up with the rules for determining the reflection and inversion properties for combinations of contrast patterns for dependent sensors and graduated readings. However I failed to include the possibility that two of the parts were the same contrast pattern. So here I reanalyse the possibilities.
At level-2 the values are A and B in which B>A.
AB Example Reflection Negative Ref. Inv.
Log differ = 1 (order) (value direction)
AB 1, 2 +1 no no + +
BA 2, 1 -1 no yes + -
If we now have another level 2 object BC (C>B).
BC Example Reflection Negative Ref. Inv.
Log differ = 2 (order) (value direction)
BC 2, 8 +2 no no + +
CB 8, 2 -2 no yes + -
When these are combined, these two patterns are possible.
ABC#1 (#5) on the left on the right Ref. Inv.
AB CB = 482 +1-2 AB + + BC + - + +
BA BC = 428 -1+2 AB + - BC + + + -
BC BA = 284 +2-1 BC + + AB + - - +
CB AB = 824 -2+1 BC + - AB + + - -
ABC#2 (#6) Ref. Inv.
AB BC = 128 +1+2 AB + + BC + + + +
BA CB = 841 -1 -2 AB + - BC + - + -
CB BA = 821 -2 -1 BC + - AB + - - +
BC AB = 148 +2+1 BC + + AB + + - -
Now let us consider when A = C. There two ABA patterns.
ABA#1 (#3) Ref. Inv.
AB BA = 121 +1-1 AB + + BA + - + +
BA AB = 212 -1+1 BA + - AB + + + -
ABA#2 (#4)
AB AB = 124 +1+1 AB + + AB + + + +
BA BA = 421 -1 -1 BA + - BA + - + -
Now let us consider when B = C. Now there is only one ABB pattern.
ABB (#2) on the left on the right Ref. Inv.
AB BB = 122 +1 0 AB + + BB o o + +
BA BB = 211 -1 0 AB + - BB o o + -
BB BA = 221 0 -1 BB o o AB + - - +
BB AB = 112 0 +1 BB o o AB + + - -
And when A = B we have the BBB pattern
BBB (#1)
BB BB = 555 0 0 BB o o BB o o o o
From this analysis I have determined the algorithm should act as follows: First the two parts should be combined based on a common part to form the new object.
If one or the other parts are symmetric then there is a BBB or ABB pattern. If it is BBB then both the reflection and inverse are neutral. If it is the ABB pattern then it is a reflection if the BB is on the left. The Inverse is based on the same rule as below.
If the two parts are the same object then there are two objects ABA#1 and ABA#2 based on the parts' inverse signs equal or not. The result is always a reflection. The inverse is the same as the left part.
If the parts are not symmetric and not equal then based on their inverse signs equal or not the new object is either ABC#1 or ABC#2.
Its reflection property is based on the order in which the two parts were joined. AB on the left produces a positive reflection and AB on the right produces a negative reflection. Now based on this reflection property,
If its reflection property is + then
Its inverse property is the same as the inverse of the left part that was used.
If its reflection property is - then
Its inverse property is the opposite to the inverse of the right part that was used.
25th Oct 2012 Phenomena
I am making it far too complex. I am taking the phenomena that I am familiar with such as inversion, reflections and rotations and trying to implement the solution to them in the structure. This is a common error, not making it general enough. These things are learnt. For size I should use logarithms and have ratio values that range from the negatives to the positives. The rotation is captured in the sign of the ratio. So what if a 3/1 ratio is the same object as a 1/3 ratio. This is learnt by associating them both to the same other object whatever that is. For intensity readings I should add 1 so there are no zeros and then take the ratio by subtracting logarithms. Who cares if the ratio is not the correct value - the ratio is still unique given the values of the readings.
29th Oct 2012 Perceptual grouping
I've found a simple algorithm for generating the ratios between adjacent and all further away stimuli by using addition and subtraction. We have Log(A/B) = Log(A)-Lob(B) giving us level-2 ratios which are aggregated to form patterns of relationships. But we also have the rule that Log(AxB) = Log(A) + Log(B) that can be used. First subtraction of logs is used between adjacent values to produce a list of ratios. Then addition of the results in this provided a common sensor is shared will produce all the other ratios. For example, if we have stimuli values A, B, C, and D.
A B C D
\ / \ / \ /
ratios = A/B B/C C/D using subtraction of the log values
\ / \ / |
ratios = A/C B/D | using addition of the values gives ratios of stimuli 2 apart
\ |
ratio = A/D -------- using addition of the values gives ratios of stimuli 3 apart
The resulting level-2 ratios in a list then look like A/B, B/C, C/D, A/C, B/D, and A/D. Then aggregation of these pairs which share a common value / sensor will produce the threesomes (A/B)(B/C), (A/B)(B/D), (B/C)(C/D), and (A/C)(C/D). The final foursome will be formed between the 1st and 3rd entries = (A/B)[(B/C)(B/C)](C/D). But why don't I just produce all the level-2 ratios directly from the level-1 stimuli using subtraction? I believe the idea was to create the further separated ratios (breadth) at the same rate as producing the aggregations from adjacent stimuli (depth). But does it really produce perceptual grouping? The idea of perceptual grouping is that two non-adjacent parts produce a gap that has the same properties of a known object. What I have devised above is just a way of avoiding the use of explicit gaps in the objects shape pattern. Perceptual grouping recognizes the gap as an object but as though it had no internal structure.
3rd Nov 2012 Contrast is local only
I have decided that the contrast pattern does not go past level-2 relationship ratios unless it is also attached to a level-2 shape ratio. That is to say, at level-2 a size and reading ratio property binon are combined to form a level-2 object binon. At level-3 only level-2 object binons get combined. Thus contrast is attached to shape patterns at level-2 only. Another observation is that you can produce phantom shape illusions but you can't produce phantom contrast illusions.
Quantity / Numerosity
Also I recognize that the size of a level-1 object is the quantity of sensors it spans. But at level-2 the object binon has a size and this is its quantity of occurrence (numerosity) rather than the sum of the sizes of its parts.
Gaps
To recognize shapes in which the two parts are not adjacent I need to recognize gaps between the parts and these gaps need to expand and contract in unison with the parts. These gaps may be filled with many parts. The solution is to create all combinations of adjacent level-1 objects with the appropriate size but a reading of unknown. Any level-2 combination with these gap parts also forms an object with unknown reading. I also only want to produce the gaps at the same rate as I produce the pure overlaps. On each pass that produces the next level of pure overlaps I also need to revisit the lower levels combinations and create any new ones that incorporate the gaps.
5th Nov 2012 Re-identifiable Objects
An object needs to be re-identifiable. This is why a single sensor with any graduated reading is not an object. The absolute value of the reading is not remembered so everything the sensor detects is the same as anything else it detects when time (change) is not taken into account. Even multiple adjacent sensors all reading the same graduated value is the same thing because you can't re-identify it because the absolute width and absolute reading are not remembered. Only relative values are remembered. This makes the level-2 relationship of width and graduated reading the smallest parts from which to form objects. A relationship is the smallest pattern. But when one has symbolic (nominal or ordinal) values one has a uniquely re-identifiable value at level-1.
Independent (unordered) sensors are effectively all adjacent to each other.
6th Nov 2012 Continuous Differentiation
I've been thinking about how I could make each level only dependent on the previous one. I believe the solution to make each level differentiate and integrate. When combining binons from the previous layer integration should always use overlaps and differentiation should always use adjacent ones. My current level-2s are differentiating and level-3 pluses are integrating.
But it is not really the differential between two adjacent objects. It is the ratio. And then it is the ratio between adjacent pairs of objects. This means the pairs must contain the values of which the ratios are going to be taken. If this is the width then each pair must contain the sum of its part's widths. And if logs are only available then this can't be calculated.
12th Nov 2012 Just Noticeable Difference (JND)
I realized today that the formula I should be using for the just noticeable difference of the ratio between two values should be Integer(Log(X) - Log(Y)) rather than Integer(Log(X)) - Integer(Log(Y)). And the base of the Log is 1 + fraction (percentage) difference you want to notice. Thus if you want to notice 10 percent differences then use a Log base of 1.1. A Log base of 1.5 gives a 50% noticeable difference - far less discriminatory. And a 5% very discriminatory just noticeable difference is a base of 1.05. Experiments on hand-written number seem to work best with about a 20% JND.
22nd Nov. 2012 Two Dimensional images
For recognition of 2-D images the shape pattern comes into existence before the contrast pattern, i.e. at different levels of complexity. The blobs which are all the same intensity have shapes. They are formed by the size relationship between the horizontal and vertical extents with the same intensity. If one was to use horizontal scans the shapes could be formed from vertical relationships between each scan. The first 1-D scan might be 3 sensors wide and the next one down might be 2 sensors wide and offset from the first by one sensor to the right. E.G.
OOO
OO
4th Dec 2012 Numerosity / Quantity
I’ve been trying to figure out how to recognize shapes independent of the levels of complexity. For this to happen I need the sizes of the parts at all levels of complexity. I think I can use the Ln(A*B) = Ln(A) + Ln(B) to help with this for repeating parts. I need to multiply the quantity of a repeating part by its size to produce the overall size. Using the logs I can get the log of the overall size by adding the log of the size of a part with the log of the quantity of repeats. This log of the overall size can then be used at the next higher level if there is a repeat. If there is no repeat the quantity is one and the log value is zero which when added maintains the log of the overall size.
13th Dec 2012 Gaps = Background
A better name for gaps is background. It has a size but no contrast pattern. Or the contrast pattern is "don't care". Patterns formed from combinations of these background objects need to be identified / recognized based on shape alone. Adjacent background objects need to be combined to form larger background objects at level 1. Using the position and total size of pure overlap shape patterns does this.
14th Dec 2012 Level of Complexity and Gaps
Today I created all the level-1 gaps / backgrounds and their combinations to figure out how to recognize objects independent of their level of complexity and produce recognizable patterns where there is a gap between the parts. Obviously it produced a vast number of patterns. But I have now realized two fundamental principles to address the two situations. To recognize an object independent of its level of complexity first of all the higher level pattern must be familiar. Also the parts it is comprised of must be familiar. And there is no novel pattern created at the higher level of complexity, it is just recognized. So when pure overlaps are transformed into Level-1 widths for purpose of finding these higher complexity objects the gaps that are formed from them must already exist - be familiar. The gaps are not combined into novel shape patterns.
Gestalt - Common fate
The recognition of an object composed of two parts that have common fate - not adjacent with each other - have a gap / background between them is not possible with a single experience. There needs to be a sequence of two experiences to identify non-adjacent parts that change in unison. Only then can the two parts be associated and the gap formed as part of the combined whole.
19th Dec 2012 Quantity
Quantity is a property of a shape. It does not appear to form patterns. If I have two different repeating shapes that repeat alternately 2 then 7 then 3 then 1 I am unlikely to form the pattern 2/7/3/1. I am more likely to notice the pattern based on the shape size rather than the shape quantity. But I will group the repeating shape into a single object and not group its parts with adjacent shapes. Also if a shape appears multiple times in an image but they are separated I will probably not automatically count them with a quantity. But if they are separated by a gap then I will give them a quantity.
21st Dec 2012 Familiarity and Novelty
I have thought of a slightly different mechanism for creating objects from familiar parts. Currently if two source objects / binons are familiar and occur co-incidentally overlapped with each other at level-2 or higher they are combined to form a target binon that maybe familiar or novel. This ultimately results in creating all possible level-2 (edge) objects between level-3 and higher objects whenever they are adjacent to each other. The improved idea is to do two passes on each level of binons. The 1st pass goes through the binons and combines any two that overlap (or are adjacent if dealing with level-1) for which the target binon is familiar (already exists). The source binons used are marked as such. The second pass then combines any remaining unused overlapping familiar binons to produce novel target binons.
I've tried this but get more ambiguous parts. Maybe I need to now combine adjacent parts at higher levels than the first. And if these are the familiar combinations may be they will be recognized before the overlapping equivalents.
31st Dec 2012 Motor Control - Differentiation & Integration
To perform motor control using a PID control algorithm the binon network has to both differentiate and integrate. Integration takes place when two objects are combined with overlaps and differentiation when they are adjacent. My current level-2s are differentiating and level-3 pluses are integrating. For motor control the log of the sensor values is not used otherwise we end up with ratios. So raw sensor values must be used and then level-2 binons represent the differences between adjacent (in time) readings. But do we need level-2 binons that represent the addition of two adjacent values to get integration?