19th January 2009 Permanence and exploration
Currently Adaptron tries every possible response for each primitive stimulus and when it has done so for each stimulus the stimulus is changed to permanent and no more responses are done. Then permanent stimuli get grouped sequentially and all possible responses tried for the pair and so on up the tree of sequential stimulus recognition. But humans don’t try every response after all possible stimuli. Read 7th Oct 2007. Maybe a better approach to exhaustive exploration at all levels of S-Habit recognition is to always create S-Habits from sequential stimuli as though they were permanent to start with. When boredom occurs due to an expected stimulus or an expected sequence being repeated then perform a response. This would allow sequences to form first. Then responses and their resulting stimuli are that which is explored.
20th January 2009 Boredom
Adaptron generates random responses when it gets bored. For the sequence ABABAB etc. it gets bored at the 3rd A. The first AB and BA start out with interest levels = 1 and when the second AB occurs the interest is reduced to zero because it was expected. The second BA is similar. But if yesterday’s idea is to work rather than respond at the 3rd A, it should start to recognize the AB sequence as an object. Thus the 3rd AB should result in the creation of S-Habit #1. The next AB would also result in another S-Habit #1. Maybe at this point boredom should be experienced. The criterion being that the input stimulus (or sequence) is the same as the previous one. This would result in long delays before any reflexive responses are performed if the environment is always changing because the same sequence must be experienced twice without anything in between. But when the response is done and it generates an unexpected resulting goal stimulus this will be an interesting response/action habit. Then either of the two parts that comprise the trigger stimulus should be sufficient to trigger the action habit.
What is the interest level of the action habit based on? It is probably based on the interest level of the trigger + goal stimulus sequence. This matches with the idea that we pay attention to the sequence of musical notes as we play the piano (self-entertainment), the action habit taking place subconsciously / automatically. The sequential stimulus change is the objective. Does this mean that stimuli never become permanent? I believe so. (Read below, they just take longer to become permanent because the repeated sequences would become permanent first)
The two parts of an S-Habit that is a long sequence, say ABCDEF, which could trigger the action would be ABCDE and BCDEF. The second seems quite reasonable but the first would be premature action. Do we have a tendency to perform premature action? After getting the ABCDE and expecting the F for a certain period of time and not getting it we may be tempted to try the action. What about DEF or just F? Will these trigger the action? This is unlikely because not enough of the context is correct.
The interest level of the action habit based on the interest in the trigger + goal stimulus sequence means that the next time the trigger occurs, the action will be performed at a concentration level. If the same goal stimulus occurs the interest will go to zero for this sequence. Could this zero interest sequence still be used as a source of actions to try for a stimulus that only partially matches the trigger?
But does it really mean I don’t need the concept of permanent. What happens when all actions have been tried for a particular stimulus? Does it retry them all again? Let’s consider an unchanging stimulus source of A’s. After the second ‘A’ a reflexive response will be produced. This will be followed by a 3rd A, which will cause a second reflexive response, which will be followed by an 'A'. These will all be marked uninteresting because they get what was expected – another A. If it has tried all responses then we want to start recognizing the sequence AA and start responding to it. This can only happen if A becomes permanent. What if it has 2 or more output devices? Should it try all combinations of responses in parallel before declaring ‘A’ permanent? Yes.
If a particular action produces new sequences in many situations it will tend to be tried / chosen 1st as an action to try in a new situation if there is a partial match with an experienced trigger involving this action. When such actions are done the attention will be looking at the expected goal stimulus to see if the same result occurs or a new one does.
Once you have heard the happy birthday sequence and you have action habits that play all the needed notes – all the needed notes are goal stimuli of at least one action habit then what mechanism is required to play the happy birthday tune? If given a start note as a cue there is an expectation of the second note. But it does not occur because Adaptron is playing them. Can the second note expectation match with a remembered goal stimulus of an action habit that will produce the note and thus start the action habit? Yes, I believe that an expectation of a goal stimulus can trigger the execution of an action habit that produces this goal even if the trigger does not match the current situation. (See 21st Jan 2009) Also a partial match of the current situation to trigger stimuli of known action habits may produce a smaller set of known action habits to select from.
When change is graduated and takes on a range of values I believe that a successful action habit that produces a big change will be repeated and get no change more times than an action habit which produces a small change will be repeated and get no change. Thus there is some sort of decay in the expected change / interest in repeating an action habit. A 50% reduction decay rate would seem reasonable. Given only integer values would reproduce the decay of interest level of 1 to zero for the symbolic stimuli.
When experimenting with 2 output devices a simple model would be one device to move left or right in a maze and a second device to move up and down. Parallel execution of up and left would move diagonally. The results of this experiment can then be compared with a single device with 8 possible motions.
21st Jan 2009 Expectations as Triggers
An expectation of a goal stimulus could trigger an action habit that produces the goal stimulus if the expectation of the goal is stored prior to the execution of the action habit. This is what happens on the 2nd execution of an action habit but I currently don’t store the expectation. If I did it would form part of the trigger stimulus and thus partially match the trigger. The action habit would probably have a zero interest level because it got the expected stimulus. But it could be scored as an action habit that works and therefore one that is repeatable if the goal stimulus is wanted.
22nd Jan 2009 S-Habits and boredom
Further analysis of the sequence ABABAB confirms that boredom sets in after the 3rd A. The first AB has interest of 1. The second AB reduces this interest to zero. The third A then has an expectation of neutral interest and thus a reflexive response is tried to produce a new sequence.
Also the sequence ABACAD results in the interest level of 1 for AB being reduced to zero when AC occurs. And the interest level of 1 for AC is reduced to zero when AD occurs. The theory being the most recent A & whatever is used to determine the interest in performing the action at A. When an action X gets performed after A in order to get a different than expected (B) goal stimulus and it does not get a different one (it gets B) then the sequence AB must be recognized and not a new memory trace produced. The X action would be associated with the AB sequence. If action Y is tried later after an ‘A’ and a result C is obtained then a new memory trace (stimulus sequence) AC would be produced (or updated if it already existed).
Instead of marking repeated memory traces with forget / disregard symbols it might be useful to collapse them and store them as an S-Habit reference, even before the stimuli are permanent. For example, the sequence ABABAMABAW could be stored as ABA12AMA4AW. 1 is (AB), 2 is (BA) and 4 is (12) which is ((AB)(BA)). For any unexpected sequence the trigger and goal are stored. For any expected sequence if there is no further expectation (at the same level) it is stored such as when 1 is stored. That is, for any stimulus that is not found in memory (no expectation) it is stored. Relationships / associations between stimuli for expectations are only possible if the two stimuli are at the same S-level. This structure would never repeat in memory any sequence that already existed. Interest levels can be assigned to all sequences. Actions can be associated with the sequence they are in. The first occurrence of a stimulus could point to the most recent to quickly find the last action tried. The correct binary hierarchy can be maintained. However what if the action X after the 3rd A results in another B rather than the M. The memory trace would be ABA12AB with the action stored with the A at location 6. ABABCA would be stored as ABA1BCA. The 1 would not be of any use in this scenario because it has no expected stimulus at the same level.
23rd Jan 2009 Reflexive responses
The idea of 19th Jan of producing a reflexive response when a sequence gets repeated versus the definition of boredom from 22nd Jan differ similarly to reacting to pain and reacting to the expectation (fear) of pain. This would imply that if the reflexive response after the 3rd A results in a B then boredom is experienced and another reflexive response should be produced after the B.
When playing the happy birthday tune using subconscious action habits and listening to the resulting tune how do we know it is different from when somebody else is playing it? We are also receiving kinesthetic feedback that is being recognized by the action habit. We are paying conscious attention only to the tune stimuli.
Symbolic versus readings
When change amounts are quantified rather than being one or zero the following situation can occur. Consider a range finder measuring the distance to an object that moves away or towards you or a single pitch note that is played at different volumes. The object is the same and repeats, i.e. AAAAAA except the intensity changes. Intensity is a measurement made by the sensor when it compares one frame reading with the next. Given a symbolic stimulus a change of zero is measured if the symbol does not change and a one if it is a different symbol. For the graduated stimuli though the change from the 1st A (intensity 0 = no reading) to the 2nd A (intensity 4) is 4 for example. The sensor is then expecting another A at intensity 4. The interest level in the AA sequence is 4 = the amount of change. The 2nd ‘A’ matches the 1st ‘A’ object and starts a habit to expect another A. It does not perform a reflexive response because change occurred. The 3rd A comes in with an intensity of 2. The sensor measures a change of –2. Again no reflexive response because of the change experienced. If the 4th A has an intensity of 2 then no change is experienced and a reflexive response would be done. The intensity sequence is the experience and is what attention is paid to. All ‘A’ objects are the same and no attention is paid to them. But the changes in intensity form a sequence that should be intensity independent. The intensity sequence 042 should also match and be recognized the same as 375 because the changes are +4-2. Just as an object should be position, intensity, reflection and negative independent so should a sequence of changes be intensity independent.
In line recognition the edge objects are recognized first and then lines are recognized as composed of two edges. This is probably what I should be doing for sequential recognition. Recognizing the change between two sequential stimuli as a type of object which has a value / reading based on the change amount. The first symbolic stimulus A should probably consist of an object A with intensity 1. All symbolic objects would have this same intensity reading of 1. Each symbolic stimulus then creates a change in object type of 1 as well as a change in intensity stimulus of zero. Each type of change can attract attention. Attention is paid to a combination of the things that can change. One could pay attention to the object alone, the intensity alone or the combination. The initial A symbol should be accompanied by an object change object with reading 1 (the symbolic object type changed) and a reading change object with reading 0 because all symbolic objects have the same reading of 1. The next symbolic stimulus B would produce an object change object of 1 and an intensity of 1 and a reading change object of 0.
Extrapolating this idea to graduated stimuli gives a list of attention criteria and possible values for them.
Attention to: Values: Change:
Sense the sense number – only one sense at a time? Is a combination of senses possible? If only one sense can be paid attention to at a time the change is either yes or no. We don’t recognize a sequence of changes in sense attentions as a thing. Also sense numbers are not graduated. We tend to pole our senses one at a time, rather than form combinations of objects from 2 or more senses. We may have many habits executing in parallel, each of which uses up simultaneous stimuli from different senses.
Position the number of sensors – 1 graduated sensor results in one primitive object and multiple readings so no position change occurs. 1 symbolic sensor gives multiple objects but always the same reading. With multiple sensors one gets a position value and thus sequential position changes can occur. Are such position changes objects with values such that sequential patterns can be recognized? I think so because you can recognize the figure eight pattern of motion in 2 dimensions. Simultaneous position changes do exist; they are edge objects between lines.
Object Type Symbolic – object type identifiers. Change is either yes or no. A graduated object type does not exist. Objects though can be composed of two parts. Sequences of object types form S-habit objects but the changes do not form objects.
Reading Range of values if graduated else always 1 if symbolic. Change in readings is always zero for symbolic stimuli but for graduated stimuli a sequential reading change object (like an edge) exists and it has a value of the amount of change in reading. Since a change in reading is an object type with a reading it forms sequences.
Reflection Yes / No
Negative Yes / No
This would mean that every time you pay attention you perceive an object type on a particular sense at a particular position. It comes with a particular reading. But it also comes with a reading change object that has been generated by the sense. This is the amount of change from the previous stimulus perceived by this sense even if the previous stimulus was not attended to. The same would apply for the position value.
When you pay attention the following combinations of what to notice and what to disregard are possible.
- You must always have a particular sense selected. You may be concentrating at a very low level so that any change in stimulus on another sense will attract your attention from the one being attended to.
- You always have a particular position and width selected. The width may be as wide as the sensor range and then you will form a composite image of the entire sense. If the width is narrow any change occurring in another position will not attract your attention unless it changes more than the concentration level. If the sensors are discrete the positions being attended to forms a binary number from the selected sensors.
- You always have a value selected for the position change object based on the sense selected. This is how much motion to expect.
- You always have a particular object selected. It is either the one expected based on what habit is being done or it is the same one as last perceived on this sense at this position.
- You always have a reading and range selected. Readings are always 1 for symbolic values.
- You always have a value selected for the reading change object based on the sense selected. Each sense has its own reading change object type because each sense produces a certain set of objects with a certain range of readings. There is only one reading change object type per sense. The value selected for this reading change object is the concentration level. It is the expected interest. If the sensors are discrete then there must be a value selected for the reading change object for each sensor that is selected. This would seem to imply that for a graduated array of sensors there is a value selected for the reading objects for the two parts making up the object type.
Combined, these selected attention parameters describe the next expected stimulus.
S-Habits are also types of objects. They have their own sense. All S-Habit objects have the same position because there is only the one sensor - consciousness. Each S-Habit is symbolic. Each is a different object type designated by its location in LTM. S-Habit reading values are all the same and values for their reading change object are either yes or no.
24th Jan 2009 Change object
I don’t believe I need to make the reading or position change into a value of a reading or position change object. The sense and the object provide the equivalent to the reading and position change object. The reading and position changes are just other attributes of the object as experienced. Just like the reading and position are attributes of the object. However I have a need to store a series of changes and say the readings are not part of the items noticed. Above, at point 5, I have said that you always have a reading and range selected. So this must be all allowed readings are acceptable – none will attract attention. Or does this mean all reading values are to be disregarded?
What I need is a marker for each of the 2 through 6 items above that says a change in this thing is to attract attention or is to be disregarded. It would not apply for the selected sense because it is a change in the other things that might produce a change in sense attention. For the selected position it would say all positions are being attended to equally. For the change in position it would say all motion changes are acceptable, there is no upper threshold above which attention will be attracted. For the object it would say any object that is found would be attended to. This would be used to find the object in a particular position. For the selected intensity it would say all intensities are acceptable. And for the change in reading it would say any change would not attract attention.
I do need to create reading change objects for reading changes because as soon as there is more than one sensor involved there is more than one pattern of changes across any given subset of the sensors.
27th Jan 2009 Concentration & Change
Concentration implies a level of expectation. If we disregard position, an expectation is on a sense, for an object, with a reading and an amount of change. The amount of change gives an upper limit on the amount of surprise you are willing to disregard. For example you experience a reading change from 7 to 9 between two stimuli of the same sense and object. Later, you get the first stimulus as a trigger and are expecting the second with a reading of 9 and a change of 2. If you get a 10 instead the change is 1 and therefore not surprising. An 11 gets the expected change and a 12 is surprising because you got more than the expected change. If you get an 8 the change is minus 1 and still less within the surprise / change range. If the habit with the expectation is subconscious the surprise / change has to be greater than expected to interrupt / attract attention. Otherwise the subconscious habit recognizes the goal stimulus and there is no interrupt. If however the habit is the conscious one then the surprise is different from expected and a new habit formed replacing the surprise expected by the first experience.
There must be subconscious S-Habits always executing for all senses so that there are always expectation levels for all senses. Otherwise how will a sense recognize an unexpected stimulus? The conscious S-Habit is the one that is being learnt / proven such that its level of interest can be reduced and thus not repeated consciously.
28th Jan 2009 Expectations
Since S-Habits can only form between permanent stimuli and only ones at the same level associations between stimuli at different levels are not meaningful. Thus they should not be on the stimulus match list.
Also habits that are triggered by and expecting permanent stimuli should not be conscious if there is another S-habit expecting a non-permanent stimulus.
Having repeater S-Habits expecting the last stimulus only makes sense for S-Habit stimuli because the sensors detect sequential change of sensory stimuli. If sensory differences were to be implemented as repeater S-Habits then they would have to exist for all senses. This is not possible if LTM does not have it anywhere for a repeater S-Habit to point to. Essentially repeater habits are S-Habit sense equivalence to sensory senses.
6th Feb 2009 Recognizing S-Habits
I’ve been trying to solve a problem of which habits to start when the attended to stimulus is an S-Habit and it is stored in LTM. This is especially difficult when it is stored and then declared permanent. Breaking down the problem gives the following logic:
If S-Habit #27 (S#27) is at level 6, attracts attention and gets stored, what habits should be started?
If S#27 is NOT permanent then start any past experiences with S#27 as the trigger. This includes ones in which S#27 is followed by another level 6 S-Habit as a combo or followed by any other stimulus at any level. They are all possible. The newly started habits will all have flags that the trigger is stored in LTM. The level 6 combos will be flagged as combos and only be expecting the next primitive sensory stimulus. The level 5 stimulus (let us say it is S#15) that forms the second half of S#27 must be on the stimulus list and must be permanent if it is part of S#27. S#15 can be used as the trigger of any new habits in which it is the 1st part of an experienced combo provided its expected goal stimulus is also permanent. The goal must be permanent because the result can only be stored in LTM if it is level 6 and this can only happen if S#15 and its goal are both level 5 and permanent. No other stimuli on the S-List can be used to start any habits because they have not been stored in LTM. This includes any higher level S-Habit stimuli and any lower level permanent stimuli that might be in the second part of S#15.
If S#27 is permanent then start any past experiences with S#27 as the trigger except for ones in which S#27 is followed by another permanent level 6 S-Habit as a combo. This is because if the combo occurs again it would be recognized, but it can’t be stored in LTM because the first part is already in LTM. No higher or lower level stimuli on the S-List can be used to start new habits because they have not been stored as part of S#27. As above S#15 can be used as the trigger to start new habits provided the experience used is a combo and the expected goal stimulus is also level 5 and permanent.
If the attended to stimulus is permanent and no stimulus is stored then all permanent stimuli are used to begin new habits. All existing habits will continue but will be stopped as soon as a stimulus is stored. You cannot have habits continuing over a storage event because if any of them succeed and get stored they will contain duplicate stimuli to those already in LTM. However A-habits can continue over a storage event because they do not contribute to conscious stimuli. The longest continuing habit will provide the context should a novel stimulus, not containing this context, attract attention and get stored.
8th Feb, 2009 Starting S-Habits
The idea of starting S#15 above presupposes that the (S#15, S#?) combo so found will have the S#? stored in LTM and the combo (S#15, S#?) will not collapse. So an additional test must be that the combo (S#15, S#?) is not permanent.
Also a repeater of a store S-Habit stimulus should not be a combo repeat but a sequential repeat of the stored S-Habit.
9th Feb 2009 Similarity / differences
These are some notes I took after reading an APS article:
Use common relations in judging similarity.
Use common attributes in judging differences
Attribution matches (parts of the whole) are less important than relational matches for similarity judgements. Attribution mismatches are more important than relational mismatches for dissimilarity judgements.
Relational matches (properties between the parts match properties between another set of parts) e.g. 1st on top of second, both parts are same colour, 1st and 2nd part are same length.
Similarity used for generalization.
Differences used for ???
When interpreting analogies as good, people attend selectively to relational properties.
Differences are more analytical - need to dig into details, look at parts separately.
Similarities are more global / non-analytical - based on an already derived thing such as change amount.
7th March 2009 Sequences of S-Habits
Back on Jan 28th I wrote “Since S-Habits can only form between permanent stimuli and only ones at the same level associations between stimuli at different levels are not meaningful. Thus they should not be on the stimulus match list.”
However this seems a little radical. Is the only purpose of having an active recognition habit to form a new combo? No. A level 6 S-Habit in LTM that is followed by a level 2 S-Habit should be on the stimulus match list. And an active S-Habit with the level 6 S-Habit as trigger should be on the habit list. It’s expecting the level 2 S-Habit. There may or may not have been an action done after the level 6 S-Habit. If the level 6 S-Habit is permanent there will be no action. The level 6 S-Habit will not have been stored in LTM because it is permanent. This strategy only makes sense if there is the possibility of another level 6 S-Habit that is permanent and could form a combo with the first level 6 S-Habit because if the level 2 S-Habit occurs the level 6 S-Habit will have to be stored in LTM. This means my strategy for not storing a permanent trigger stimulus in LTM when it occurs should be restricted to only if there is an expected goal stimulus that is permanent, at the same level and could form a combo with the trigger.
Back on the 8th Feb I should have written the (S#15, S#?) combo must be non-permanent so that it gets stored if attended to. Both S#15 and S#? must be permanent and the combo can be stored in LTM provided it is not permanent. Then the S#27 goal (which is S#15) will match the (S#15, S#?) trigger in LTM. This applies whether S#27 is permanent or not.
8th March 2009 Episodic memory
While trying to solve the test if two adjacent locations in LTM form a known combo that is permanent for the item immediately above I thought I might get rid of episodic memory all together. But then I ask what purpose is it currently serving? Right now recall is done against it to get the experiences (matches and next expected stimuli in recency order) for any stimulus on the S-List. Habits have pointers into it for their next expected stimulus. Both of these could be built using pointers in the object tree. Each entry in the tree would have to have a list of pointers to “next” stimuli and the list would have to be in recency order. Each of these would need an interest value, which is currently in LTM. Also each of these would need the value of the tried response.
19th March 2009 Recognizing S-Habits
I’ve realized I also need to recognize any two sequential S-Habit stimuli that have any number of overlapping parts. That is, ABCDEFG could be ((AB)[(CDE))(FG)] where the (C[D)E] is the common part. Also it could be (ABC[(DE))(FG)] where (DE) is the common part. It all depends on which sequences have become permanent. Thus I have decided to store a count of the level-1 parts (primitive stimuli) that each stimulus in LTM shares with its previous stimulus. For example (AB) is S#6 and (BC) is S#9, then ABC would be S#6 followed by S#9 and have an overlap of 1 for S#9 since it shares the B. However ABBC would be S#6 followed by S#9 and have an overlap of 0 for S#9. Two S-Habits will still need to be at the same level to collapse but two S-Habits at different levels may occur sequentially with zero or more overlapping parts.
29th March 2009 Overlapping S-Habits
My current strategy for starting S-Habits with a certain amount of overlap is:
- If the attended to stimulus has been stored then
- For each recognized stimulus on the stimulus list, start a habit to expect its goal stimulus.
- If the trigger is permanent then
- If the expected goal stimulus is permanent then the s-habit overlap is the same as the level of the expected goal stimulus. This is because if the goal stimulus should occur the pair will collapse and the resulting stimulus will have an overlap of this amount with the last stored stimulus.
- Else if the expected goal stimulus is not permanent then the s-habit overlap is the overlap in LTM of the expected goal stimulus.
- End If
- Else if the trigger is not permanent then the s-habit overlap is the overlap in LTM of the expected goal stimulus.
- End If
- If the trigger is permanent then
- End For
- For each recognized stimulus on the stimulus list, start a habit to expect its goal stimulus.
- Else if the attended to stimulus has not been stored because it was a permanent stimulus then there should be no non-permanent stimuli on the stimulus list, so
- For each recognized stimulus on the stimulus list, start a habit to expect its goal stimulus.
- If the expected goal stimulus is permanent then the expected overlap is set to zero. This is because if the goal stimulus should occur the pair will collapse and the resulting stimulus will have an overlap of zero with the last stored stimulus.
- Else if the expected goal stimulus is not permanent then the s-habit overlap is the overlap in LTM of the expected goal stimulus. This is because, if the goal stimulus should occur the permanent trigger will be stored as context and this non-permanent goal will have the overlap with the context as found in LTM.
- End If
- End For
- For each recognized stimulus on the stimulus list, start a habit to expect its goal stimulus.
- End If
|Conditions and Actions||R||U||L||E||S|
|A. Attended to stimulus was stored?||Y||Y||Y||Y||N||N||N||N|
|B. S-List trigger stimulus permanent?||Y||Y||N||N||Y||Y||N||N|
|C. Expected goal stimulus permanent?||Y||N||Y||N||Y||N||Y||N|
|1. Use LTM overlap of goal||x||x||x||x|
|2. Use goal’s level as overlap||x|
|3. Overlap = 0||x|
|4. Not possible||x||x|
I have realized I am mixing up two concepts. One is the number of level-1 stimuli that the trigger stimulus of an S-Habit has in common with the last LTM stimulus and the number of level-1 stimuli that the trigger stimulus has with the S-Habit’s goal stimulus. When I store a stimulus in memory I should mark its second part stimulus as being part of the stored stimulus. This would be repeated for all sub-second parts. Any S-Habit started from these stimuli would have their Trigger overlap set to the stimulus’s S-Level. Any S-Habit’s started from any other stimulus (not part of the attended to one) would have the trigger overlap set to zero. This would apply to an unexpected stimulus.
Then the expected goal overlap from LTM is used when the S-Habit is set-up so that the parts in the overlap are not expected since they are in the trigger, which has already occurred.
3rd April 2009 Edges in time
An edge is a change in intensity on the linear sensor scale when it's a graduated sense. It is also an edge if the readings are discrete / symbolic; however there are many types of edges based on the adjacent symbols. If the sense is comprised of independent / discrete sensors such edges in position don't exist. Likewise using the time scale an edge is a change in intensity or symbol from one sequential reading to the next. These are the Level-1 S-Habits formed using symbolic stimuli and the Level-1 changes in intensity over time for graduated readings. Just like a line is the occurrence of two edges so is a Level-2 S-Habit a combination of two Level-1 S-habits. So what about changes in position over time or a change in size over time - are these the edges for these dimensions. If they are then two such edges give us an object in that dimension and thus position objects and size objects can be produced using two edges. Two changes in reflection will produce an orientation object and two negative changes will produce a negative/positive object. (Read 16th April 2008)
Then actions can be associated with objects (which could be combinations of objects) and not objects and experienced readings because experienced readings are also objects. But is this plausible? The same object is first at position 1 then 2. This is a change in position - a position edge - a speed measurement. Then it moves to position 3 producing a second position edge. This is over the time scale. Does this give us a position object? This represents an acceleration measurement. But just like the width of the line is the difference in position of the two edges and this width is part of the experience - not the object, the acceleration object also has an experienced amount. Thus as long as we are dealing with a linear graduated scale, edges /changes will border objects, which are based on relative difference in the edges, and they will have experienced values which are also graduated. Thus changes in these experiences can be used as higher level edges to form objects. Thus one cannot escape having objects along with their experienced values.
But maybe one should produce such higher level objects to identify the experience and use a criteria for creating the higher level objects based on whether such an object is "useful". By useful we could mean it correlates somehow with behaviour or reward. We certainly don't want to explore all such higher level objects.
ReadingsLinear needs to be set to true if readingsgraduated is false (symbolic) thus the test in CheckIntensity can be simplified.
Can the current S-Habit recognition process which combines each subsequent stimulus into ever increasing higher level S-Habits be used as the approach to recognizing a linear set of stimuli per sensor that all arrive simultaneously? That is, process them from left to right, combining them into higher level objects as we go. They would all be considered permanent because no time exists to respond to them. Overlap must exist. The form of the tree, which combines objects only if they are the same level and adjacent, is universal. Maybe also gaps are universal and should be considered for S-Habits. A gap is effectively a negative overlap.
Systems include the vestibular system (which codes the direction of balance, Stoffregen& Riccio, 1988; and gravitational acceleration), the muscular and articulatory endocaptors [PDF] (which regulate the balance of muscular tonicity), and the cutaneous exocaptors (which record the relative pressure across the soles of the feet).
16th April 2009 Object versus Experience
An object is recognized as a combination of parts that have the same relative properties to each other. And relative properties are edges / changes. Objects get recognized independent of their absolute property values. However the absolute property values are kept in experiences and are used for the basis of behaviour. Thus executing an S-Habit is the process of recognizing the same sequence of relative changes. Executing an A-Habit is the performance of a learnt behaviour based on experience. Thus an S-Habit is an object. If the stimuli are graduated the S-Habit is the series of changes in values. The execution of an experience with absolute values in LTM is an A-Habit. The selection of a possible action / response could be based on the absolute LTM experience match or it could be based on the object with different absolute property values found somewhere else in LTM. (I started thinking about this on 5th Feb 2008, 6th, 7th and 15th Feb also.)
So I need to implement a change object for sequential intensity change. But also one for position, size, reflection and negative change. And they need to be available for each sense. This is equivalent to a 5-sensor sense with independent sensors and each sensor has graduated readings. Objects can be formed out of different combinations of the sensor readings. These objects are not position, size, negative nor reflection sensitive. They just form a combination / pattern with a certain intensity. So time to experiment with Sense #1 to recognize these combos.
18th April 2009 Level of Interest
I have mentioned before that level of interest is a graduated intensity reading (27th April 2006, 21st Sept 2006, and 23rd Jan 2007). However, on the 14th March 2007 I said novelty is always greater than the Level of Interest. This is not true – it is possible to miss a novel stimulus because of concentration. This also means a new line may not attract attention if a move line causes a greater change.
19th April 2009 Experience versus objects
On further thinking about keeping / storing / remembering absolute values for position, intensity etc. as part of one experience I have concluded that we remember none of this. We only remember "combinations of changes" represented as objects. We remember sequences of changes too. So what about that memory you have that is so vivid because it really surprised you or caused some other strong emotional reaction at the time. Surely the absolute values were recorded in memory. I would say, “No, they are not, just the amount of change that was caused and the amount of change amongst the part of the experience were recorded”.
This means that we explore the objects in our environment independent of their position, intensity etc. But since context provides a relationship between an object and its surrounding we may explore the situation, which is the object in this context. It also means that recreating a sequence from episodic memory is virtually impossible if any part of the sequence is occurring for a second or subsequent time. There is no episodic memory except for the first occurrence. This will result is some major changes in Adaptron’s S-Habit recognition / performance.
20th April 2009 Gaps and overlaps
If I make all episodic experienced S-Habits into objects then I would be creating pairs of stimuli, trigger and expected goals with an overlap reading as S-Habit objects. But these would be different from the ones that have become permanent. Their overlap is always level-2 parts. What about gaps in sequential habits? Is this meaningful? What about overlaps in parallel / simultaneous stimuli?
16th May 2009 Changes on changes
For senses with graduated readings it makes sense that changes have a graduated value and changes in these changes are stimuli. However as soon as a sense produces discrete / symbolic readings or a multi-modal stimulus (P-Habit which is symbolic) is formed from two or more senses then changes are meaningful stimuli but changes in these changes does not make much sense. I even question, for symbolic stimuli, whether changes are also stimuli or just information on which to base decisions / attention.
For symbolic stimuli the change is binary, it is either 0 = the same or 1 = different, and there is no graduation reading to this change – it is in one state or the other. Thus the change is discrete, not graduated such as when the stimuli readings are graduated.
6th June 2009 Multi-sensors – Discrete
If we have a sense with multiple discrete reading sensors and the sense is discrete (independent sensors) such as sound sensors recognizing words said by different people, then how would this best be represented as a data structure? We can’t use edges because the sensors are not adjacent to each other. A symbolic reading (X) must be recognizable independent of which sensor detects it since the sense is for the one type of information. But the readings must also be associated with a sensor so that X on sensor 1 can be recognized as occurring again. Similarly X on 1 and Y on 2 is different from X on 2 and Y on 3. This would imply that the most primitive objects are the X or Y symbols. But the position of a stimulus object from a graduated (dependent sensors) sense is to be stored in LTM as an experience with a position and the object is a binon that is position independent. However the X and Y combinations are sensor (position) dependent and thus sensor number (position) must be combined with each X or Y before combining them to form the pattern.
Equivalent position dependency is found in the P-Habits but its source objects are all sense dependent – unique to the sense. The position to be stored in LTM with a graduated sense object is only one value and cannot be used to keep the multiple positions for the different sensor readings. But if position is kept with the X and Y pattern as an object how does the “pay attention” mechanism work for a particular sensor? Do I need to recognize the X on sensor 1 is the same X as on sensor 2? On 14th Nov 2007 I appear to disregard the fact that an X on one sensor is the same as an X on another sensor. This means that each sensor has its own set of symbolic values, or if they produce graduated readings each sensor is measuring its own object.
But I do want to recognize a series of different symbolic stimuli from the same (unchanging) sensor as well as a pattern of the same (unchanging) symbolic stimuli from different sensors. What about recognizing a series of different patterns of symbolic stimuli from the same set of sensors or a series of the same pattern of symbolic stimuli from different combination of sensors? Yes, this has to be recognized as well. The solution may be to combine the symbolic stimuli into objects i.e. XY and produce a sensor object from the combination of sensors involved. Then these are kept in LTM together. They can then be recognized independently of each other and changes / difference in stimuli combination or sensor combinations can be determined. But what about XXY, XYX and YXX, aren’t they the same?
7th June 2009 Multi-sensor implementation
I had a solution on the 6th July 2007 (also mentioned on 14th Nov 2007) to implement this using large binary numbers representing which sensors were part of the object. However using the binary tree structure will work better and reuse existing code. To represent two Xs and a “Y” pattern independent of sensors I must first sort the string of sensor raw stimuli that have attracted attention (changed). I call this the stimulus pattern. This would then be recognized in a tree of combinations. The sensor numbers that are rearranged by this sorting must also be recognized. I call this the stimulus sensor pattern. Thus X1, X3, Y5 would produce the sorted tree <(X[X)Y]> and the sensors would be <(1[3)5]>. X1, Y3, X5 would produce the same sorted tree of stimuli but with sensors <(1[5)3]> and Y1, X4, X7 would produce the same sorted tree with sensors <(4[7)1]>. This would allow for the recognition of the same pattern of stimuli independent of sensor position. To represent a sensor pattern such as 1, 3 and 5 independent of their stimuli I must produce the tree of sensor combinations. I call this the sensor pattern. So X1, X3, Y5 produces a sensor pattern as above of <(1[3)5]> and Y1, M3, A5 also produces the same sensor pattern.
It is the stimulus pattern plus the stimulus sensor pattern that uniquely identifies the experience. The object experienced consists of the sensor independent piece (the stimulus pattern) plus the stimulus independent piece (the sensor pattern). Then the experience occurred at the stimulus sensor pattern location. The stimulus sensor pattern indicates where to attend if you want the same stimulus pattern and the same sensor pattern. However since this is going to take some time to implement and test I plan to use the P-Habit recognition algorithm for multi-sensor discrete sense and discrete or graduated readings. This means that the rearranged patterns of stimuli are not recognized.
8th June 2009 Thinking and A-Habits
As soon as I start forming A-Sequences where a R, S, R collapses into a single A-Sequence response then thinking one expectation ahead will allow for thinking two or more ahead because it will be thinking across A-Sequences. For example S is any stimulus (compound, raw, P-Habit or S-Habit) and R is a raw response. S1, R1, S2, R2, S3 is done and the A-Sequence A1 = R1, S2, R2 gets formed and the sequence is stored in LTM as S1, A1, S3. Assume we have the sequence S3, R3, S4 in LTM. Then when S1 occurs again and it is recalled the expectations include S3. When S3 is thought about it produces the expectation S4 and if the A1 is done and S3 reached, R3 will be done. Thus we will have thought forward from S1 to S4.
22nd June, 2009 Multi-sensor expectations
During the running of Testcase #26, the P-Habit ‘_B’ is recognized and the LTM has a recording of experiencing the B part at locations 6, 4, and 2. Location 4 is to be ignored. Should 2 backgrounder S-Habits be started to recognize the location 6 S-Habit B -> a and the location 2 S-habit B -> A? Is this the same concept as stimuli becoming permanent for S-habits? I think that two backgrounder S-Habits should not be started because B is not permanent. And the concept of permanent (not responding to) does not apply to parts of a parallel simultaneously recognized set of stimuli. The pattern ‘_B’ was what attracted attention, and this is what is conscious.
Later in this test case it is conscious of an ‘a’ on sense 1 and it gets this stimulus repeated but an unexpected stimulus ‘A’ on sense 2. Naturally the ‘A’ attracts attention. But it is also executing two S-Habits that are expecting a ‘_’ on sense 1 combined with an ‘A’ and a ‘B’. Does the ‘_A’ reduce the interest in the ‘A’ input? No, I believe the expectation of the S-Habit should only reduce the interest in the combination ‘_A’ or ‘_B’ not any part that matches. This corresponds to my thinking on 14th and 25th Nov 2007.
And yet on 4th Dec 2005 I say that just one of a combination of stimuli may trigger a response due to generalization presuming there is interest in doing the A-Habit.
26th June 2009 Multi-sensor symbolic expectations
I have been thinking about this subject and have come up with the following analysis. Given a 3-binon tree such as:
where B and C are sensory binons and A is a combo – P-Habit binon.
Sensory binons such as B and C get stimuli from their corresponding measurement / detection devices. They / it is always expecting the same input stimulus as before and only lets the P-Habit binon A know about a stimulus when it is different from the previous one. In other words it is executing its own S-Habit. A possible implementation in real neurons is the first stimulus triggers, saturates and fires it, setting its level of activity above threshold and when the next stimulus occurs it is already saturated above threshold and does not pass on the stimulus. Thus the same second stimulus is neutralized; the binon is expecting it. But a different stimulus is unexpected, (a change occurs) and it passes that on; it does not neutralise it.
If the binon ‘A’ is a learnt permanent S-Habit then it is triggered by B and is expecting a C. If C fires then the A S-Habit is saturated and fires saying it has recognized the sequence B, C. If C does not fire then the S-Habit has failed. The phrase “expecting C” means that it is triggered but does not have enough input to fire. The firing of the C binon gives it that input, enough to exceed its firing threshold. Now if it were to behave like a sensory binon it would be expecting itself to occur again, stay the same. This is why there is a repeater S-habit. But I only have it on the last conscious S-habit. Should they also exist for all unconscious background S-Habits?
If however, the binon ‘A’ is a P-Habit both B and C must fire in parallel (simultaneously, at the same time) for A to exceed its threshold and fire. If however A is expected as a goal of an S-habit then the S-Habit is waiting for A to fire. If either B or C does not fire then A will not fire. Thus expectation of a combo P-Habit is an expectation of the AND of the parts not an OR of the parts. Interest level in a part due to a change taking place is not neutralized by the combo A binon being expected unless both the parts occur. So expectation does not propagate down the tree to the leaves. Expectation of the combination should neutralise the interest in the combo of the parts not neutralise the interest in the individual parts.
What if the ‘A’ P-habit has never been obtained before? If both B and C stimuli are different and therefore have a change on the sensory binons but each symbolic value received is familiar to LTM and expected by different S-habits then both B and C’s interest level will be neutralized. Now it depends on what S-Habit is conscious as to what attention is paid to. If attention is paid to the sensory binon level the novel combination will not be perceived. If attention is paid to the A-Level P-Habit binon then the novel ‘A’ combo will be perceived.
The test run from 21st June also presents an interesting question. Based on the ‘A, B, C’ tree above it is equivalent to the B, C combo being novel and A has attracted attention and been stored in LTM. But B is familiar and in LTM could be the trigger for two background S-Habits. Currently backgrounders are only started if the trigger is a permanent stimulus because no action will be done. We can’t have actions started subconsciously. If two background S-Habits were started from the B then what purpose would it serve. It would reduce the interest in their goal stimuli should they occur.
29th June 2009 Conscious / unconscious
Yesterday I had a test run in which I wrote, “If you have just experienced something novel and you are surprised, you are not performing any conscious habits expecting anything. What do you pay attention to next? If you are performing some subconscious habits expecting things, do any of these that get their goal stimulus and complete attract your attention? Or is it any stimulus that is not expected by your subconscious habits that attracts your attention? It is the later, because subconscious habits are there to neutralise expected stimuli and thus highlight unexpected stimuli. Conscious habits however focus on / highlight expected stimuli.” However in the given run both stimuli were neutralized, one by a sensor that was expecting no change and the other by a subconscious habit. So the question becomes; do we then pay attention to the whole scene or may be the same sense as the surprising stimulus so that the novel stimulus becomes familiar?
If we are in the living room with a metronome ticking and we have subconscious habits neutralising these stimuli and a bear appears in the middle we get a big surprise. The running / continuing subconscious habits should not be stopped, and new ones should be started as necessary. Thus, neither the metronome nor the whole scene should attract our attention. I suggest we will pay attention to the bear again.
30th June 2009
Details about the Canadian Revenue Agency’s SR&ED tax incentive program are here.
3rd July 2009 Attention continues
I’ve tried the above idea and it works well except for the situation when we are not performing a past experience – a conscious habit. There may be no conscious expected next stimulus. So I made any unconscious habit that has not been learnt – has a redo-interest greater than neutral – into the conscious habit. This then had problems if there was no such habit. So I made the completed background habits goal stimulus into the conscious one. This also had problems when the goal was an S-Habit but didn’t match any of the next experiences. Maybe the better approach will be to pay attention to the last sense and location that was conscious. If this last conscious stimulus is an S-Habit then the next stimulus to pay attention to is the final sensory stimulus that the S-Habit got.
7th July 2009 Subconscious and Attention Algorithm
I’ve been trying various approaches to attention and subconscious habits. When a subconscious habit completes it may or may not produce a new S-Habit stimulus. If it was performing a sequence (a pair – trigger and goal) which could not be combined into a new S-Habit it would have served its purpose and reduced / neutralised the interest in the goal stimulus. However, if the completed sequence can be combined to form an S-Habit I was putting this on the stimulus list. I have decided this should not take place. Such combinations are only formed by completion of conscious habits because it represents a new learnt stimulus.
Then when it comes to what to pay attention to next the following items are used to determine the priority in attracting attention.
- The existing conscious habits may have all failed to get their next goal stimulus.
- One or more conscious habits may be continuing having got their next expected stimulus.
- One or more conscious habits may have completed. They have either (A) finished and produced a new combination stimulus or (B) just finished.
- There may be some stimuli whose sensory interest has not been reduced by subconscious habits.
- There may be some stimuli that have non-neutral expected redo interest due to recall of past experiences.
- There may be no conscious habits, just subconscious ones. This happens when a novel stimulus attracts attention. The concentration level is neutral.
The algorithm to decide on what next to pay attention to is: (see 17th April 2007 for previous)
If there is a 4/ with an interest level greater than concentration then it interrupts any conscious habits. Subconscious ones continue. Subconscious habits failing, continuing or completing are never stopped or interrupted and never interrupt consciousness. They are there purely to reduce or neutralise stimuli sensory interest. They all should operate at a neutral interest level.
If there is a 5/ with an interest level greater than concentration then it interrupts any conscious habits. I have experienced concentrating on the performance of a practiced habit and in the middle deciding to try an alternate sequence. I had thought that 5/ should be greater than or equal to the concentration level for this to happen. Now I suspect that the practiced habit is being done subconsciously and the concentration on its performance is purely an observation of the resulting stimuli being done at a concentration level lower than the interest in 5/. The practiced habit is being done subconsciously and thus has no interest level; it’s neutral. There must have been an interest to begin it but it executes subconsciously at a neutral interest.
If there is a 1/ then attention is paid to what stimulus happened instead of the one expected by the first conscious habit. There may be many conscious habits. Each is expecting a goal stimulus based on what has been experienced in the past for the given trigger stimulus. The first conscious habit represents the most recent experience and the most likely to happen again if the laws of the environment have not changed. Then the concentration level is neutralized.
If there is a 2/ but no 3/ then attention is paid to the first conscious continuing habit’s expected and obtained stimulus. The concentration level remains the same. The first conscious continuing habit may not be the first conscious habit. The first conscious habit may have failed because the environment has changed and an earlier experience is being repeated.
If there is a 3/ but no 2/ and it is a 3/(A) then attention is paid to the first conscious completed habit’s combination stimulus. If this stimulus is new then it may already have been selected in 4/. Then the concentration level is neutralized so it can be established based on the new stimulus recall.
If there is a 3/ but no 2/ and it is a 3/(B) then attention is paid to the goal stimulus that was expected by the conscious completed habit. Then the concentration level is neutralized so it can be established based on the goal stimulus recall.
If there is a 3/ and a 2/ then attention is paid as above when there is a 2/ and the concentration level is not neutralized.
If there is a 6/ then attention is paid to the same sense, sensors and expected object that was attended to previously. This may not be done if 1/ or 2/ exist with an interest level above neutral.
Test case #9 has made me change my mind on the combination of 3/ and 2/. The reason being that shorter length sequences should be learnt before longer ones. Thus if there is a 3/ and a 2/ then attention is paid as above for the appropriate 3/ and concentration is neutralized. All continuing habits will be stopped.
If I don’t place S-Habits on the stack that results from the completion and collapse of subconscious S-Habits I am faced with a major problem forming sequences with overlaps. Most of them come from the completion of subconscious S-Habits. On further thought these collapsed sequences should be formed by the subconscious habits and available on the S-List should attention be paid to them such as in situation 6/. An example is when listening to one person in a crowd and you stop listening to them you can drag out of your short term memory (STM) the last few words spoken by someone behind you that you had been disregarding previously.
9th July 2009 Overlap calculations
These are the calculations for overlap of parts of an S-Habit. The measurements are all counts of the number of sense or P-Habit S-level 1 stimuli that comprise the S-Habit.
T = the length of the trigger stimulus = its S-Level. .
G = the length of the goal stimulus = its S-Level. .
X = the overlap of stimuli in T and G. .
F = the final T & G collapsed S-Habit (if it is collapsible) length = its S-Level = T+G–X. .
O = when the habit sequence is placed on the Habit list initially the trigger has this overlap length with the last stored LTM stimulus. .
N = the non-overlap length of the habit on the habit list for it to complete = the number of stimuli needed, shorter than or equal to the goal length, cycles to go through, to complete this habit = F – O.
GO = Goal Overlap length, the number of stimuli in the goal that are already in the last stored LTM stimulus = G – N = the unneeded part of the goal when completed. If it is negative then it is a zero overlap. Any final stimuli that are created in the execution of the habit that are a partial match of the goal stimulus will have this same goal overlap.
But each cycle of matching the most recent stimulus with the next expected part of the goal produce a cascade of partial matches, some of which may or may not contain part of the goal overlap with LTM. If N are needed and the 1st stimulus (S-Level 1) occurs then its overlap is zero. Collapse it with the previous stimulus to get an S-Level 2 and the overlap is the 1st one in GO. Collapse this to get an S-Level 3 and 2 of the GO are covered. Etc. If N-1 stimuli are now needed the S-Level 1 and the S-Level 2 collapse have zero overlap of GO and then S-Level 3 overlaps 1 and 4 overlaps 2 etc. If N-H stimuli are now needed the S-Level 1 to H+1 have zero overlap and the overlap increase from then on with each higher S-Level recognized. So if I used the S-Level of the recognized collapsed stimulus and subtracted H+1 then if the result is less than or equal to zero the stimulus overlap is zero. Else it is the value derived.
H = the length of the N that have already occurred. .
S = the S-Level of the collapsed recognized piece of the goal stimulus.
SO = Stimulus overlap the collapsed recognized piece of the goal stimulus = S – (H + 1) = 0 if negative.
There is also an upper limit on SO and that is it cannot be greater than GO.
If SO is negative it would seem there is a gap between the end of the last stimulus stored and this collapsed stimulus. Should this stimulus be used to start any backgrounders? May be to recognize any subconscious sequences and reduce interest but any final collapsed combo should not become conscious?
13th July 2009 Subconscious sequential recognition
I’ve been struggling with whether subconscious sequences made up of permanent stimuli should be made available to conscious attention. Specifically when two sequential stimuli collapse and combine to form a new stimulus. I believe these must be formed and made available to conscious attention. We have the ability to change our attention to the last few things said by somebody even though they were not attended to at the time they were said. We have a sequential buffer accumulating this sequence of subconscious stimuli. Background S-Habits are executing for this purpose. They reduce the interest in known sequences and create or leave new sequences interesting to attract attention.
14th July 2009 Parallel subconscious habits
I have been starting up subconscious / background habits from all stimuli on all senses and sensors and partial S-Habit pieces created in the recognition of larger S-Habit stimuli. However these can result in the creation of novel parallel sequences which contain none of the conscious attended to stimuli. These can then be stored and be inconsistent with previous LTM stimuli. Attention is paid to them because they are novel and no conscious habit are being done and none of any interest. Should they have been executing? Should whatever it gets as a new combo, be thrown away? If it is to be recorded what form should it take? Does it need to be learnt?
It probably should have been done (as it was) as a background to reduce interest in the stimuli from that sense. The sequence needs to be made available as a stimulus as discussed this morning. It’s obviously going to attract attention. May be the question is how to represent it and interpret it as an experience.
Maybe the current LTM representation is adequate. It says the two stimuli overlap. It does not have to be interpreted that the two contain the same part stimuli. This overlap implies that there are two different traces taking place in parallel. So when it comes time to do a recall there will be a match with the actual stimulus but there must also be a match with a parallel overlapping stimulus later in LTM but that started at the same time.
I have decided to only start background / subconscious habits that take as their trigger any part of the attended to stimulus. This means that they all started “consciously”. Thus we only ignore / neutralise those stimuli in the context / situation we are in after we have learnt to do so as part of the conscious stream of stimuli for that context / situation.
17th July 2009 Subconscious habits
I was allowing any subconscious habit that collapsed completely to put its recognized S-Habit (if there was one) on the S-List in case it was interesting / novel enough to attract attention. I have decided that this can only take place if there is no currently active habit. This rule applies even if the active habit is being done at a neutral interest level. There is usually no currently active habit after a novel stimulus has been attended to. It is as though there is a level of concentration associated with performing a habit consciously at the neutral level of interest that is different from not performing a conscious habit at the neutral level (possibly free recall).
If the sequence A, B, C is experienced then A, B has a redo interest as does B, C. If A, B occurs again (assume followed by X), A, B is learnt and its redo interest becomes neutral. If thinking is turned on, looking one sequence in advance and A occurs again then the B is thought about. The B, C sequence is interesting to redo so we repeat whatever action we have learnt for the A, B habit, hoping to get the B. If we get a D instead then the A, B sequence is currently remarked neutral – not worth redoing. And the occurrence of an A will continue to trigger the A, B habit trying to get to the B. What is needed is to distinguish between learnt and not worth redoing and not worth rethinking about. Or this is accomplished by only thinking about the most recently experienced thing that follows B, that is the X rather than the C.
A test run seems to indicate that collapsed completed subconscious habits need to be recognizable on the S-List when the concentration level is neutral, not if there are no currently active habits.
28th July 2009 Repetition ignored
I have decided that just like recognition habits of an array of sensors combine a series of adjacent equal valued sensors into one stimulus I should have S-Habits stop recognizing a series of repeated stimuli. Thus when the pattern forms an S-Habit and this S-Habit repeats it is to be ignored. This radically changes my test cases.
1st August 2009 Change & Graduated Stimuli
From the 18th Nov 2006 last paragraph and 18th Dec 2006 notes I think the correct approach is to record the sequence of P-Habits that combine the graduated sensor readings with the change readings. Then if the same graduated reading OR change reading reoccurs it will partially match the trigger of the S-Habit consisting of the sequence of P-Habits. Then the expectation will be the graduated reading, the change reading or the combination that makes up the goal P-Habit. This will allow the “Happy Birthday” tune to be recognized at any pitch or at any volume because the same change sequence is involved. I should try this strategy with symbolic stimuli.
Based on 12th Dec 2007 I should be implementing S-Habits slightly differently and also do the same for P-Habits. Their objects should consist of the two object parts and a relative reading difference. Then the experiences of these should contain the trigger reading for the S-Habit and the first object part reading for the P-Habit. Redo interest would remain an experience attribute.
3rd August 2009 S-Habit Reflections
I have just changed symbolic stimuli readings to a value of 1 as mentioned 23rd Jan 2009. Now the relative / change in intensity between the symbolic trigger and goal stimulus of an S-Habit is 0 = 1 – 1. However now the sequence A, B is the same object as B, A but just a reflection. And I don’t store reflection experiences in LTM along with the object / binon ID. Which re-introduces the question about action habits being triggered by the trigger object (type of thing) or the experienced trigger that includes properties of reflection, reading etc. We must be able to discriminate between the experiences and have different behaviour based on experience. We may start out with behaviour based on type – generalization but must be able to discriminate / specialize if it makes a difference in reward or interest. (See 16th, 19th April 2009 and earlier at 5th through 15th Feb 2008.)
4th August 2009 Change in Objects
The above difficulty with reflections now makes me realize that each binon / object must also contain the relative difference in reflection, negative and position. It already contains the relative size and reading. S-Habits will contain these changes over time while P-Habits will contain these simultaneously.
14th August 2009 Experience versus object
I have been separating the experience, which is in long term memory, into to three parts. The location the stimulus was perceived (Sense or senses, sensor position/pattern)[where], the object perceived [what] and the absolute values of the stimulus (size, reading). The object must be represented such that it is recognized again at different sensor position / patterns, sizes, readings, reflections (order of parts) and negatives. Thus the object must be independent of these features. Yet it must also record this information when it is relevant in identifying the object due to the need to discriminate between objects of the same object type with different values for these features. So far I am representing an object as a pair of parts with their order, relative position, size and reading. But I need to be able to recognize the general object type made up of the two parts independent of their order, relative position, size and reading if necessary. Also I need to be able to recognize a specialized object made up of two parts in a particular order or relative position, size and / or reading.
An example of what is necessary is the recognition of the middle line of three adjacent lines with readings of 090. From left to right there are two edges E1 and E2 involved (edge is the type of object). E1 has a reading gradient (difference) of 9 and E2 of –9. The middle line L1 (line is the type of object) is made up of the two edge parts. However, given 395 the 1st edge E3 has a gradient of 6 and the 2nd edge E4 has a gradient of –4. The middle line L2 is made up of the two edges E3 and E4. But most people would generalize the lines and recognize L1 and L2 as the same type of object. However should there be the need to discriminate between them L1 and L2 are different. The same is true between the four edges. They are all of type edge but each is a special case.
When we break down the many difference properties / features between two parts that need to be used in discrimination between objects made up of the same two parts we end up with the list:
- Position difference value – width of object – distance between parts
- Position difference direction – object order – image reflection information
- Size difference value – proportional size of parts – size ratio
- Size difference direction – we don’t use this often and don’t have a name for it
- Reading difference value, 9, 6 and 4 for edges
- Reading difference direction – negative image information, -9 and +9
These difference features are object level independent. That is, two lines (level-1 objects) might have a reading difference of +9 and two complicated parts might have a reading difference of +9. This also applies to all possible combinations of these features. This would imply that each difference feature is like a stimulus with a value and P-Habit like combinations must be recognized independent of the two object parts making up an object.
Take this idea and apply it to the four edges and two lines example. This example disregards position and size so no left or right orientation is involved. We need the following primitive objects.
- – -1, -1 – an edge value 0
- – -2, -2 – a reading difference value 9
- – -3, -3 – a reading difference direction value +1
- – -3, -3 – a reading difference direction value –1
- – -2, -2 – a reading difference value 6
- – -2, -2 – a reading difference value 4
We then need to describe objects made up of two parts that are these primitive objects.
- – 1, 2 – a reading difference +9
- – 1, 3 – a reading difference –9
- – 0, 6 – an edge with reading +9 E1
- – 0, 7 – an edge with reading –9 E2
- – 2, 4 – a reading difference +6
- – 0, 10 – an edge with reading +6 E3
- – 3, 5 – a reading difference –4
- – 0, 12 – an edge with reading –4 E4
- – 0, 0 – a line 0
- – 8, 9 – a pair of edges – no position, left right orientation involved
- – 14, 15 – a line with two edges L1
- – 11, 13 – a pair of edges
- – 14, 17 – a line with two edges L2
To be more accurate we must have a common middle part to combine objects. Thus all the parts of objects 1 through 6 would have to be –1’s.
- – 1, 2 – a reading difference +9
- – 1, 3 – a reading difference –9
- – 0, 1 – an edge with reading difference 9
- – 6, 8 – an edge with reading +9 E1
- – 7, 8 – an edge with reading –9 E2
- – 2, 4 – a reading difference of +6
- – 0, 4 – an edge with reading difference 6
- – 11, 12 – an edge with reading +6 E3
- – 3, 5 – a reading difference of –4
- – 0, 5 – an edge with reading difference 4
- – 14, 15 – an edge with reading –4 E4
- – 0, 0 – a line
But this produces a problem. Although combining objects 10 and 11 because of the common object 9 could form L1, this would not make sense. And combining objects 14 and 17 can’t be done to make L2. So what is the common part? It is the side of an edge. A line is made up of two edge sides. Thus the lowest level primitive needs to be an edge side.
The example becomes:
- – -1, -1 – an edge side value 0
- – -2, -2 – on the left value 0
- – -2, -2 – on the right value 0
- – -3, -3 – a reading value 0
- – -3, -3 – a reading value 9
- – -4, -4 – a reading difference direction +1
- – -4, -4 – a reading difference direction –1
- – -3, -3 – a reading value 6
- – -3, -3 – a reading value 4
- – 0, 1 – a left side
- – 0, 2 – a right side
- – 0, 3 – a side value 0
- – 0, 4 – a side value 9
- – 9, 11 – a left edge side with reading value 0
- – 10, 12 – a right edge side with reading value 9
- – 0, 0 – an edge
No, this is not working either. A more accurate representation is that a line is made up of two sides that are of equal value while an edge is made up of two sides that have different values. Thus edges and lines are at the same level. Then alternating sequences of lines and edges can be formed with common sides. Now the example becomes:
- – -1, -1 – a generic Side object type 0
- – -2, -2 – a reading difference 0
- – -2, -2 – a reading difference 9
- – -3, -3 – a reading difference direction +1
- – -3, -3 – a reading difference direction –1
- – -2, -2 – a reading difference 6
- – -2, -2 – a reading difference 4
- – 0, 0 – an generic edge or line object type 0
- – 2, 3 – a reading difference & direction +9
- – 2, 7 – an edge with reading difference 9
- – 8, 9 – an edge with reading +9 E1
- – 2, 4 – a reading difference & direction –9
- – 9, 11 – an edge with reading –9 E2
- – 1, 7 – a line with reading difference 0
Direction values could be +1 for positive, -1 for negative and 0 if there is no difference. But since whenever there is a reading it has to have a direction it would seem natural to always combine a reading difference with a direction. The same should apply to size and position differences. Since left to right reflection is captured in the position difference direction the part object numbers could always be stored in the lower to higher numeric order.
Each sense will use different combinations of the three difference features depending on sense graduated or not and readings graduated or not.
|Sense Type||Reading Type||Use Position||Use Size||Use Reading|
17th August 2009 Concepts
I realized that the combinations of difference in position, size and intensity are all concepts. And as stated above they can be applied to all objects at all levels (level independence). Likewise combinations of the direction of the differences form concepts. Examples are larger on the left, brighter on the right.
I have also realized that the difference in width / size is the second derivative of position. The position of an object A is the “centre of gravity” of the two parts that comprise the object. It’s calculated using the formula (position 2 + position 1) / 2. But the width of the object A is position 2 – position 1. And that is the difference in the positions of the two parts, the 1st derivative. Then when the difference of A’s width to another object B’s width is determined we have the 2nd derivative of positions. But this is the combination of parts A and B. So just by combining A and B we get a representation of the difference in width / size of the A, B combo. This means there is no need to represent difference in width explicitly. But it does mean that lines of different widths must be stored as objects such that they have a position difference.
Thus I need to be able to recognize:
- the type of object independent of difference in the position and intensity of its parts and the concepts of:
- the difference in position of two parts
- the difference in intensity of two parts
- the left to right order of the two parts (position difference direction)
- the intensity difference direction of two parts and combinations of these five.
When generalizing only the type of object is recognized. When specialising / discriminating combinations of types-of-object and concepts are recognized.
But now what is a type of object? At the lowest level edges and lines are primitive types of objects. But as soon as two lines are combined we produce multiple types of objects. Assuming no gaps between the lines then we could, for example, have: two adjacent lines with the right one larger than the left one or vice versa, a pair with the right one brighter than the left one and vice versa, a pair a certain distance apart or a certain brightness difference. Lots more are possible for non-adjacent lines. This seems to imply that a type of object is any object composed of two objects at the level below. Then an object is generic if it is made up of two objects below it and specialised if it is made up of an object at this level plus a combo of concepts.
19th August 2009 Multi-sensor implementation
I have got a pretty good algorithm working in Sense3 for the idea from 7th June 2009. However for Y1, M3 changing to M1, Y3 and then back to Y1, M3 there is no change in the stimulus pattern M, Y and no change in the sensor pattern 1, 3. There has to be an additional piece of information about the stimulus pattern saying it is a reflection or maybe a negative. This is currently determined from the positions if the sense is graduated. However for independent sensors I need to provide information about the sensor numbers so a reflection (position difference direction) can be determined. Thus the independent sensors have an order but no position for distance calculations. Take this to level 3 and the pattern A, B, B could be interpreted as B, A, B or B, B, A depending on which sensors the 3 objects occur on.
I tried representing the sorted list of sensor locations as a binary tree and found it too difficult to maintain the sort order. I have now successfully represented the sensors involved in the form of a string. Each sensor is either a 0 or 1. A 1 if the sensor’s object / reading and is part of the experience or 0 if it is not involved. Would it now make sense to represent the object generically such as A, B, B and then contain the order of them in the sensor location string? For example if B2, A5, B6, occurs in a discrete set of 8 sensors, the stimulus pattern could be A, B, B (sorted order) and the sensor pattern (“where” information) could be 0-2-0-0-1-3-0-0. If the stimulus pattern changes to C2, A5, and B6 the new sensor pattern would be 0-3-0-0-1-2-0-0 and the stimulus pattern ‘A, B, C’. This produces a sensor independent object. However this does not feel right. I would be complicating the where to pay attention information with what it is information. Maybe the answer lies in the generalization / specialisation process. The sorted stimulus pattern forms the general object and then it is combined with order information based on some sensor order to form specialized objects.
20th August 2009 Concept representation
I’m trying to figure out a way of storing the parts of an object, concepts and combos in binons for independent sensors and discrete / symbolic stimuli. For the 3 readings A, B, and C I can store the sorted order patterns of A, B, C, AB, AC, BC and (AB)[BC] where B is the overlap. These are the generic stimuli objects. Then I need to represent a positive and negative PDD (Position Distance Direction) which serves the purpose of sensor order concept. The tree starts to look like:
Ob# Level O1 O2 Represents.
0 0 -2 –2 Order +1.
1 0 -3 -3 Order -1.
2 0 -1 -1 A.
3 0 -1 -1 B.
4 0 -1 -1 C.
5 1 2 2 A.
6 1 3 3 B.
7 1 4 4 C.
8 2 5 6 AB generic.
9 2 6 7 BC generic.
10 2 5 7 AC generic.
11 2 10 0 AC +.
12 2 10 1 AC –.
13 2 8 0 AB +.
14 2 8 1 AB –.
15 2 9 0 BC +.
16 2 9 1 BC –.
17 3 8 9 (AB)[BC] = ABC generic.
18 0 0 1 +1 –1.
19 0 1 0 -1 +1.
20 0 0 0 +1 +1.
21 0 1 1 -1 –1.
22 3 17 20 ABC generic and +1 +1 or.
22 3 13 15 AB BC the combo of specialized objects
I can’t use the pairs of the order objects along with the generic object as in 22 because there are only 4 combos and I need 6 ways of representing the pattern containing ABC. The six are ABC, ACB, BAC, BCA, CAB, and CBA. I can use the pairs of level 2 specialized parts but then I lose the level 3 generic information that is in 22.
Using the definition stated on 17th August that “An object is generic if it is made up of two objects at the level below it and specialised if it is made up of an object at the same level plus a combo of concepts.” Level 3 would be:
Ob# Level O1 O2 Represents.
17 3 8 9 (AB)[BC] = ABC super generic.
18 3 13 15 AB+ BC+ generic = the combo of specialized objects AB BC.
19 3 18 0 AB, BC + = AB, BC specialized order = ABC.
20 3 18 1 AB, BC - = BC, AB specialized order = ? .
21 3 11 16 AC+ BC- generic = AC CB.
22 3 21 0 AC, CB + = AC, CB specialized order = ACB.
22 3 21 1 AC, CB - = CB, AC specialized order = ?
This representation is still challenged. If each of N sensors reads a different object there are N factorial order permutations of those objects. Thus at level L there are L factorial order permutations if each sensor reads a different object.
What I really need is a set of order patterns for each level. At level 2 it would be 2 combinations of 12 and 21. At level 3 it would be the 6 combinations of 123, 132, 213, 231, 312, and 321. Then at each level an object would be represented as the super generic combo and one of the order patterns. The super generic combos would be based on the sorted set of sensory symbols using a binon tree to combine them at different levels. Can the order patterns be represented in a binon tree? Is it necessary to recognize such patterns and sub patterns within them? If we were dealing with relative brightness I would say yes over a linear graduated scale because we recognize such patterns visually e.g. striped, graded etc. So it is probably necessary to recognize the 23 in the 123 and the 231 order patterns.
Using the B2, A5, B6 example I would end up with the following level 2 information:
- Where objects Wh1 = 0-1-0-0-1-0-0-0 and Wh2 = 0-0-0-0-1-1-0-0
- What objects Ob1 = AB and Ob2 = BB
- Order objects Or1= 25 and 56
At level 3 the information would be:
- Where object Wh3 = 0-1-0-0-1-1-0-0 = Wh1, Wh2
- What object Ob3 = AB and BB = ABB = Ob1, Ob2
- Order object Or3 = 25 and 56 = 256 = Or1, Or2
For C2, A5, and B6 reusing the objects already obtained above the results would be:
- No new where objects
- What object Ob4 = BC at level 2. Ob5 = AB and BC = ABC = Ob1, Ob4 at level 3.
- Order object Or4 = 62 at level 2. Or5 = 56 and 62 = 562 = Or1, Or4 at level 3.
This should allow for recognizing the object pattern independent of their order or to recognize a change in order but not in the object pattern on a given set of sensors. The order pattern would be specific to the sensors involved because the sensor numbers are being used. An order pattern on one set of sensors has no comparability with the same order on a different set of sensors because they are independent. This would mean that a LTM experience is made up of:
- a where object for attention focus,
- a generic what object for recognition and
- an order object that provides specialization for discrimination when necessary.
23rd August 2009 Directed Attention - Entire Tree
I’ve integrated the newest line recognition software with Adaptron and have made an interesting observation. I had previously thought that for a multi-sensor sense I only had to create the binon tree of objects that had changed from the previous frame. This is what attracted attention. Then when an S or A-Habit was being done directed attention would create the stimulus tree for the desired sensor pattern and whatever stimulus was there would be attended to. This is what the FindInstead routine was to do. However attention is not only attracted by a changed sensory stimulus but also a stimulus that has an expected interest based on its LTM S-Habit do-interest. This can only be determined by producing the complete binon tree of stimuli and checking their do-interest. A second reason that justifies creating the entire tree is that attention is attracted to hearing ones name mentioned in an unattended to conversation. If the current do-interest is attached to the binon in the tree and kept up to date then this could be compared with the concentration level at recognition time to determine attraction rather than search LTM.
26th August, 2009 Order information
I have the Order-objects being formed and successfully associated with the generic What-objects. But I am sorting the two part objects in object number order for What-objects so I don’t end up with two representations of the same two parts just in a different order. This works fine for the symbolic independent sensors because the readings get sorted first even before making them into and using them as parts. And when they are made into parts the symbolic objects are in order. So for them the two parts in an object do not need to be sorted based on object number. For the graduated independent sensors there is only one primitive object and the sign of the Intensity Difference Direction helps distinguish between two objects using the same two parts when the two parts are in object number order. But they also do not need to be kept in sorted order by object number. Then the IDD’s are always positive and the InD is the reading increase from the first part to second part because their readings have been sorted ahead of time. Note that the What-object only contains the difference in intensities making up the pattern of graduated readings of the independent sensors. Now in both cases it does not matter what order the symbolic objects or intensity readings are created into objects in the object array.
When dealing with graduated readings on independent sensors a change is a sequential (time based) object in which the two parts are the graduated reading objects. The intensity difference between these two would then form part of the Change-object. Or would it? May be the intensity-difference is the experience value and the What-object is a change of a certain type, i.e. a reading-change as opposed to an object type, order or position-change. Multiple independent sensors would produce change objects that are combinations of reading and order-change change-objects. The object type-change change-object would be the one that occurs between symbolic readings on a single sensor. As soon as you have symbolic readings on several independent sensors you then get combinations of object type-change and order-change change-objects. S-habits are change objects in which the two parts may be from different senses or the same sense but different Where objects.
Is there just one kind of reading-change change-object or are there different ones for each sense? Since we appear to be able to correlate the direction and intensity of a reading change across senses it would seem to imply the former.
Do we remember the magnitude of the changes or do we just remember the change type? Or do we just use the change?
27th August 2009 Attention to Change
When we pay attention to change there must be an order in which the different changes are processed. First we are concentrating at a certain change / interest level. We are expecting a certain type and amount of change at a certain Where. Then attention is attracted to a different where based on a change at this location greater than the concentration. The order might be:
- What / object type-change
- How much / reading-change
- Expected change from LTM recall.
I suspect that along with an object it is the differences between parts making up objects and the amount of change in change-objects that gets remembered, not the absolute values / readings.
When I have an S-Habit in which the senses of the trigger and goal stimuli are different the redo interest is first set to the amount of change that occurred to the goal stimulus compared to the previous stimulus on the goal stimulus sense. It is not the amount of change between the trigger and goal. Only if the trigger stimulus is the same sense and where of the goal stimulus can the change between the trigger and goal stimulus be the same as the goal stimulus change.
30th August 2009 Discrimination / Specialization
Now that the order object is stored in long term memory with the where and object type I need to sort out when just object type is used (generalization) and when the combination is used (discrimination). For exploration, actions are initially tried reflexively based on object type matches as triggers. When an A-habit gets repeated an exact match of the trigger stimulus (where, what and order) will start the habit (first case) but so will a match just based on the object type (what) if no exact match is found (second case). In the first case we are expecting an exact goal stimulus. In the second case we are expecting just the same goal object type.
If in the first case the goal obtained
- Is exactly as expected then the habit has been learnt and no change was experienced
- Has the same object type then unexpected where and order changes attract attention.
- Is completely different then the world has changed and a new goal has been learnt.
If in the second case the goal obtained
- Is an exact match to the expected we have a second case in which the same goal can be obtained
- Has the same object type then the general rule has been learnt
- Is completely different then a new rule for the given trigger has been learnt.
But discrimination occurs in the second case when just the object type was used as a trigger match and an unexpected goal is obtained. Getting the completely same expected goal reinforces the general rule between the trigger object type and goal object type. Getting the same expected object type but unexpected order and / or where object would seem to also reinforce the general rule. Getting a different goal object type than expected would seem to provide a new rule that is unique for the actual trigger stimulus combination, i.e. discrimination.
But the stimuli involved here as triggers may be change-objects and not object type, where and order objects. If we just work with readings not graduated i.e. symbolic and independent then the object type-change change-object (Cb, in the code) is either 0 or 1 because an object is either the same type or not. The order-change change-object is also either 0 or 1 because the same order of sensors is involved or not. And the where-change is either 0 or 1 because it is either the same set of sensors or not. A single sense was easy to deal with because only object-type change could occur and this was the only criterion for surprise and boredom.
I currently use sensor numbers in the order object and I feel this is incorporating where information as well. For example the ABC combo found in the 5 sensor reading of XAYBC provides the object type ABC, where of 01011 and the order 245. The order should be 123 saying the A is first, B is second and C is third. This would mean BXYCA and XBYCA would both produce object type ABC and order 312 with different where objects of 10011 and 01011 respectively. Another example of AXBYC would recognize ABC at 10101 and order 123.
1st September 2009 Generalization / Specialization
In generalization we need only one of the two parts of a binon to match to cause us to try the response associated with the binon. This presumes there is no exact match with the two parts. If there is an exact match then that is a more specialized situation that we have already experienced. When we get the results of trying the response based on a partial match we now have a specialized experience. If we then accumulate several specialized experiences in which only one part of the binon is relevant and the other part is irrelevant the response must somehow get associated with the part, lower on the binon tree and the specialized situations are not consulted for the actions to perform. See 11th Feb 2008 as an answer to how. Only the attended to part of the trigger gets stored as the trigger for the habit. We have effectively learnt the generalization.
Specialization however continues to get different goal stimuli based on the exact match of the two parts making up the trigger binon.
Using an example will help clarify the mechanism. If it is a red sky (S1) with the wind blowing (S2) and you clap your hands (R1) and a streetlight comes on (Sr) you learn this interesting sequence. If you later experience either a blue sky (S3) with the wind blowing (S2) or a red sky (S1) with no wind (S4) you will clap your hands (R1) expecting the streetlight to come on (Sr). If the light does come on then you have laid down a memory of the generalization in which the trigger was just that part on which the sequence was based. In the blue sky with wind blowing example the remembered trigger would be only the wind blowing (S2) and the sequence would be S2 – R1 – Sr. This becomes the generalized learnt behaviour. If you then experience a grey sky (S5) with the wind blowing (S2) you will clap your hands (R1) expecting the light to come on. However, if nothing rewarding or interesting happens (Sn) then you are disappointed and remember the exact nature of the situation as S5 & S2 – R1 – Sn. If you then experience the wind blowing with a red sky or any other colour other than grey you would perform the clap hand response. The most recent experience does not outweigh the previous ones. If you experienced a grey sky with or without the wind blowing you probably will be bored and try some other response, but not clap your hands.
If instead of nothing happening (Sn) in the grey sky with wind blowing followed by a hand-clap situation, lightning strikes (Sp) you will remember all the details including the trigger stimulus in full detail, S5 & S2. Then if a grey sky or a wind occurs you would expect lightening and be fearful and not clap your hands. You may perform some other reflexive response. Even if it is a red sky and the wind is blowing you will not clap your hands. This is because the most recent experience overrides the previous ones.
3rd September, 2009 Generalization in Recognition
The lowest binons on the recognition tree are the most general and the higher you go the more specific it becomes. New binons are created when new pairs of parts are recognized. But maybe already existing binons are activated if just one of their two parts is recognized, thus allowing one to recognize a hammer even though the handle part is obscured. It could also account for many optical illusions.
Interest in P-Habit and Combo parts
It is important that the parts of a combo or P-Habit only gain their own interest and get explored when they start occurring and changing independently. If apples always come with pips and pips always come with apples and apples and pips never occur by themselves then for all intents and purposes they form the one object. There is no way to separate out the parts and react differently to them. Thus when an expected P-Habit occurs the parts are also expected and lose their interest along with the whole.
4th September, 2009 Order information
I am struggling with how to deal with order information. The Where information is important because it provides the location to pay attention to. It is also at the where location that the change occurs. We do not generalize our behaviour based on where an object is using the Where combination of sensors that are used in Adaptron. Where the object is in relation to other objects we do use and generalize on this. However we do generalize our behaviour based on the type of object e.g. all dogs or we specialize our behaviour on the type of object and order. If we talk about position order then an example is a dog facing left or right. But the order that I am analysing right now is the order of the sensor readings. This is somewhat of an artificial thing used to detect when the readings exchange positions to different sensors but still form the same combination of readings i.e. the same type of object. There has been no change in the where. The readings have moved amongst the same sensors. The specialization is the type of object in this order. The generalization is this type of object. We don’t generalize the order and base behaviour on it just like the where information. Maybe order is the next level of attention detail after where. Sense is the highest level of attention detail, very general. The Where level is the next level of detail. It is a little more specialized and order is the most detailed level and specialized.
When an object gets recalled it can match the same specific sense, where and order. Or it could match more generally the sense and where but a different order. Or it could match the sense and order but not the same where. Or it could be found on that sense at a different where and order than ever experienced. Then it can match just one part of an object. If we learn a behaviour that is generalized on an object independent of its where and order how does this get represented. For parts of a binon the general rule is represented because the part is recorded as the trigger and not the more specialized object it was part of.
More Generalization / Specialization
On 1st Sept. 2009 I wrote about the red sky (S1) with the wind blowing (S2) and you clap your hands (R1) and the streetlight comes on (Sr). If you later experience a blue-sky (S3) with the wind blowing (S2) you will clap your hands (R1) expecting the light to come on (Sr). If the streetlight does come on then you have laid down a memory of the generalization, in which the trigger was just the wind blowing. This is now independent of the sky colour. If however, nothing rewarding or interesting happens (Sn) then you are disappointed and remember the exact nature of the situation as S3 & S2 –> R1 –> Sn. Even more so if you are punished. How does the trigger change from being sky independent to sky dependent after the Sr or Sn? It has already been recorded. What must be done is that the complete trigger combo is stored and when it comes time to use the learnt behaviour generalization is only done on the most recent partial match and only if it was rewarding. Now this is more consistent with where and order information being stored with every experience but only used in generalization if redo interest is rewarding.
This would mean that when an object is attended to and there is no exact specialized match worth using the process of generalization would take the two parts of this object and do generalization on both of them and pick a rewarding response if one exists. Or is it possible the stored trigger gets modified after the result to indicate whether the generalization worked or it is a special case. This would mean removing or making irrelevant those parts of the trigger that did not play a part in selecting the response if it is rewarding. But as soon as this is done there is no way to get generalization transfer to the removed parts. This would seem to reinforce the fact that all the details of the specialized trigger are stored and generalization takes place in the process of using the stored information.
5th September, 2009 Exact / Partial match
Thus an exact match of an attended to stimulus to a LTM experience becomes a match of the same sense, where, order and object type. Partial matches must match the same sense but could be at a different where; at a different order provided it’s the same where and / or match just one part of a LTM object. The same is true for any part of the attended to stimulus.
Now that I have implemented this, a question arises about retrying partial match behaviours. Take the red sky; wind blowing, hand clapping and streetlight scenario. When the blue sky and wind blowing occurs and you clap your hands are you expecting the streetlight to come on or are you just trying the same response because it was interesting before. If the former then when the light does come on the sequence has been learnt and has a neutral redo interest. However the original red sky wind blowing triggered habit should remain interesting.
6th September 2009 Past Experiences
I believe that I can always change the past experience redo interest to neutral as soon as I use it as a trigger for a conscious habit. This is because if the expected goal stimulus is obtained then the original habit is learnt and if a different goal stimulus occurs then the world has changed and the more recent experience has a redo interest of neutral or more.
The learning cycle when including partial matches is as follows:
- The attended to situation occurs and is recalled.
- A remembered goal stimulus is obtained either due to (A) an exact match with the situation or (B) a partial match.
- The goal stimulus is interesting and wanted so start a conscious habit.
- The appropriate response is produced and goal expected based on the habit performed.
- The redo interest in the trigger is neutralized for an exact match (A).
- The goal situation occurs.
- If the goal is expected then the redo interest is neutral.
- If the goal is unexpected the redo interest is wanted.
7th September, 2009 Partial match rule
A LTM partial match of the attended to stimulus can only be used as a trigger for a possible response to try if the match is more recent than any exact LTM matches. This is necessary because we don’t want to keep trying the partial match and keep getting the same expected goal stimulus. We have to get on with trying some other possibilities. Thinking is required to use any other previous partial matches when a more recent exact match has occurred. We need to think about which partial matches have been tried and which have interesting goals and have yet to be tried.
8th September, 2009 Change and Object Identification
When a stimulus is symbolic it can immediately be uniquely identified. However when it has a graduated reading it continues to identify the same object but the readings can change. The reading does not further identify the object. It is always the difference between the readings on two parts that identify a composite object or the change in readings on one part that identify a change object. Although I store in LTM the readings on an object it should not be used to identify it. But the change object is an object with a reading, which is the amount of change. Should I create a different object for each change amount if I don’t store readings in LTM? It comes back to the same question as before, can we remember absolute pitches or do we only remember their relationship to the sounds before or after them?
14th September, 2009 Readings and change
The answer is that we do not remember absolute readings. Even when looking at something the brightness of an object is determined based on its relative brightness to the surrounding objects. When recognizing lines I use edges to recognize where they are. There is only one type of edge object and its intensity change is part of its experience. For lines, there is one type of line object for each size (edge position difference) but I use the line’s reading as part of its experience. Then when two lines are combined I produce unique objects with size difference and reading difference. This works for a situation in which the sun goes behind a cloud. The entire scene changes in brightness but nothing else changes. So the objects must be identical before and after the brightness change.
Now I need to apply the same principle to sequential readings such as from a range finder or the volume of musical notes. The change in reading is equivalent to the edge. Only one type of change object is required and the change amount is part of the experience. There should be one type of reading object for each size. Size now becomes duration. Since we don’t have a sense of time there will be just one reading object right now. Each reading will be part of its experience. Then two reading objects in sequence will form a combination object with both parts from the same sense. The change in reading will be the defining property that helps to identify the combination object. I also need to remove the P-Habit that is the combination of the reading object and change object.
So each sense at each interval will produce a reading object, a change object and a combination object using the previous reading object. Upon start up or upon attention switched to a sense there will only be the reading object and change object available. To recognize the combination object you have to be conscious of both its parts. See 18th Dec. 2006. Does this mean that I need a second change to recognize the second reading object? This second change may be an external change or the polling timeout of our attention. A series of readings with polling timeouts might be interpreted as follows. At start up the 1st reading of 4 occurs. There is no previous reading so no change object exists. Similar to the processing of the left edge of my multi-sensor graduated sense with graduated readings. Attention polls the sense again and gets the reading of 4 and the change object with value of zero is experienced. This is an edge and the reading is a reading object. But the reading object does not get produced until the next change occurs because that delimits its length in time. So given Cp is a polling change and Ir is an input reading we have the sequence Ir1, Cp1, Ir2, Cp2, Ir3. Cp1 produces a change object (edge in time) with an experienced reading of zero since all Ir’s have a reading of 4. Cp2 produces a change object with an experienced reading of zero but also a reading object Ir2 with an experienced reading of 4. It needs another Cp4 to produce the Ir3 object with a reading of 4. At this point we will have two reading objects the same, will be bored and perform a reflexive response. This is assuming we get bored of reading objects and not of change objects. But an Ir4 with a reading of 4 was necessary to produce the Cp4. Is this Ir4 thrown out because the response was done after the Cp4 in response to the Ir3 and in parallel with the Ir4 because the readings have been obtained by polling? If the changes are non-zero then we could presume the environment causes the changes and the response is done before the next change but after the reading.
Do we ever pay attention to the change object? If the reading stimulus is long and continuous and then it changes to another reading value and stays there our attention is attracted by the change. But do we record it or the new reading? If we remember sequences of two reading objects with a reading difference then this is equivalent to the change between the two readings. This pair is also independent of the absolute values of the readings of the two parts.
15th September, 2009 Readings and Change
I asked if we needed a 2nd change to recognize the 2nd reading object. I would say since there is no concept of time the 2nd reading obtained due to polling or an external change is assumed to have lasted for the shortest discernible time interval. That time is up even though there has not been a 3rd change to end the 2nd reading object. Thus the occurrence of the 2nd reading object is good enough to signal the end of the 2nd reading object. At this point the sequence of the two reading objects can be formed.
16th September, 2009 Change and Reading Objects
Once again I am faced with the question, should the readings be explored even though the same object is being observed? According to previous thinking the readings are not meant to be kept in LTM or used to identify the object. But I have been using them in the recall of triggers for S-Habits. Maybe the things that need to be explored are the changes. The actions are done to cause change and no change (boring) causes a reflexive response. Once a change and response results in the same goal change the sequence has been learnt. But I have previously concluded that changes are represented by a sequence of two reading objects with a change in reading between them. Thus I should start exploring the S-Habits formed from two stimuli / objects which are from the same sense. This then begs the question, what about responding / exploring the reading object without considering its reading?
I think I should try making the reading changes into objects at level 0 like edges. Then combine two edges to form a reading object at level 1. This would not have a reading value but be unique because of the two changes involved. Then I should explore both change and reading objects. Have I tried this strategy with lines and edges? No I am using one generic edge object and one generic line object per size. The edge difference is being stored as the experience in the reading, not as the intensity difference. Similarly the reading of a line is part of its experience not the object. Why have I been using this strategy? See 19th, 23rd, 24th June 2008.
If I change the line recognition strategy then level 0, edge objects will exist for all possible differences in intensity but have no position difference. Level 1, line objects will consist of 2 edges with a position difference and no intensity difference. Level 2 line combinations will form with common edges if the lines are adjacent and with no common edges if there is a gap between the lines. The problem now arises in identifying the same level 1 line between different edge pairs. It would mean the line with intensity 4 in 1114111 would not be the same object as the line with intensity 5 in 1115111 because they do not consist of the same edge objects. But it would be the same as the line with intensity 5 in 2225222 because the pair of edges is the same. Maybe what we have is an exact match to 2nd example with the background intensity of 2 and a partial / generalization match to the 1st example with the background intensity of 1. In the second example the intensity of the entire scene changed, the line object did not. In the first example the central line changed. The two lines share the same width and gradient direction of edge intensity difference but not the magnitude of the edge intensity difference. And it is this change in the magnitude of the edge intensity difference that gets interpreted as a change in the lines reading.
For discrete independent sensors with graduated readings there are no adjacent readings to form edges. The solution would be to assume each sensor has two edges with readings of zero or –1 on both sides.
I have said that we generalize before we specialize. Maybe I should loosen the criteria for the selection of reflexive actions when coming across new experiences so that partial matches to trigger are chosen first.
What I need to do is create for each sense a time based change object which is used before and after a sense reading object to form the reading object. These time based change objects would be level-0 edges in time and the reading objects would be level-1 events. However I have the same issue about the reading object cannot be produced until the change occurs after it. My thinking of 15th Sept. works if a reading object does not consist of a pair of change objects. This would mean that the reading object is level 0 and a change is level 1 comprised of two reading objects. And a range finder always detects the same object so there is only one reading object at level 0. Its experience contains the reading. The level-1 change objects would then have an intensity difference. But this presents a representation problem because currently I have a position edge at level 0 that is paired to produce the reading object at level 1 which has a position difference (width) and no intensity difference. Two level-1 reading objects will need to be paired to form a level-2 change object that will contain the reading change in the intensity difference entry. I can shift all these levels down 1 if necessary. Or can I store the reading change as the intensity difference in a level 1 reading object? This might work for level 1 reading objects but what about level 2 and higher objects where the intensity difference is position based rather than time based.
Basically what I am trying to do is represent reading change as an object so that it can be explored independently of the object that has the reading. Reading change is sequential and is produced when a reading is obtained. The reading object that has the change may be a simple level 1 object from a single sensor or a composite object from several sensors in which there are intensity differences between the parts. And the reading change must be sense specific. Maybe the solution is to add another property to an object that is the intensity change. This means that when I get to process graduated sensors I will probably then have to add position change. If I do this then generalization will have to recognize the same object with the same parts, position difference and intensity difference but different intensity changes and / or position changes.
28th Sept 2009 Response
Because the world produces a continuously varying sequence of stimuli human responses are not produced after every stimulus. In speech a long series of syllables can take place before we respond. We start to react when it gets boring which is far less frequent than the rate at which stimuli are perceived. It's only when it starts getting boring - repeated inputs that we start to respond / react. If the world is changing rapidly we take in the stimuli at that rate up to the time resolution of our senses. If it is not changing then we pay attention / poll at a rate based on our internal conscious sense of time which is slower than the maximum rate at which we can distinguish sequential versus simultaneous stimuli. This results in a response / action / behaviour which does not responding to every individual stimulus. The patterns of stimuli are complex and stretch out sequentially in time with infrequent repetition. It is only when we get the repetition that we are bored and react.
The challenge is that Adaptron starts to respond to the lowest level syllables first, assuming that atomic stimuli frequency is the response frequency. This is not true when people are talking. However, when performing physical activity the reaction speed must be in synch with the stimulus frequency to act / behave correctly. This also applies for the sense of touch in ones mouth for forming syllables during speech.
14th Oct 2009 Procedural versus Declarative
Learning to recognize -> Declarative.
Learning to do / act -> Procedural
Generalization as a Concept
When facing an unfamiliar situation, which has similarities to a known situation, a partial match occurs and an action may be done. The partial match as an expectation / idea may be the generalized concept that we then use as a trigger for this action. This would mean storing this generalization / concept and doing recalls on them.
16th Oct 2009 Edges and Objects
When considering lines maybe there should be only one edge object type and each occurrence has a particular reading, negative or positive. It has no width, the reading just mentioned and a position, which form its experience. Then a line type object at level 1 exists for each possible width but they have no intensity difference. Each occurrence has a particular position and two particular edges with readings that form the experience. Could there be just one line type object and its width is part of the experience? This would allow a change in width to be the same object just with a change to the edge positions. Then two adjacent line objects can form a level 2 object with a relative intensity and a relative width.
18th Oct 2009 Edges and Lines
The idea of a single line object seems correct based on the idea that any part of an object that can change is part of the experience, not part of the object. I distinguish between objects and experiences in that we explore objects with different responses / actions but not the experiences. The experiences are the values of readings which are always relative to adjacent readings and basing actions on them is fruitless. The same applies to position and width, which are relative. Experiences are used to identify the objects but should not be stored in LTM. Thus a single line object is as generic as possible. Two edge objects are required to recognize it. Line objects have no relative intensity (intensity difference) or size (position difference). Position of the line can change. Intensity of the line can change. Width (position of the two edges) of the line can change. The same rules apply to an edge. An edge is recognized as a change in intensity. The generic edge object has a relative intensity and no size. Its experience also has a width that is zero. However its position and intensity can change.
A two-line object is recognized as a pair of line objects. It has no relative intensity or size. The relative intensity (intensity difference) of its two parts is part of its experience. This would also apply to the position difference of the two parts. Extrapolating this approach would mean that each level only has one generic object. But what do we really recognize as an object. I have this statement that an object consists of parts which all change together, which can be phrased, parts of an object do not change independently. Making this more specific means that an object cannot be picked out of a simultaneous set of stimuli until a change occurs in sequence. However potential objects can be identified using edges in space. Let’s start with an edge experience. It is at a position and has an intensity difference. Both of these can change. Its two parts are a left side and right side. We would say it is the same object if the intensity difference between its left side and right side does not change. However, if the right side or left side intensity changes while the other side does not then it is a different edge object.
Let’s apply this thinking to a line. It is at a position and has an intensity, both of which can change. Its two parts are a left and right edge. We would say it is the same object if the intensity and / or the position changed. The two parts can change independently and still be the same object. This contradicts the statement about the two parts changing together. But when we consider two lines the principle can be applied. The relative intensity and relative size must remain constant for it to be the same object. Also the relative separation to size of the parts must remain constant. An earlier line recognition algorithm used this approach. The newer algorithm has made each line of a different width into a separate object and combinations of two lines have a position difference giving the separation information.
A different approach is to consider the line at level 0, the edge at level 1 composed of two lines and a pair of lines at level 2 composed of two lines. Two edges are never combined to form anything but are used to delimit lines. This would mean there is only one level-0 line object. A line’s position, width and intensity provide its experience and could change. There would be many edge objects each consisting of two lines with an intensity difference but no position difference. The position, width of zero and intensity of the edge would provide its experience and could change. Two lines would be combined to form a pair. The pair object would have an intensity difference and an intensity experience that could change. Position and width of the whole can both change. But the ratio of the two line widths must stay the same and the change in the position difference, if any, must be the same as the line width changes. The actual ratios that must stay constant for the same pair object are the line1 width / line2 width, and gap width / line1 width or gap width / line 2 width.
If this approach is then taken with sequential input of graduated readings from a single sensor each stimulus is a line of a given intensity. Only one stimulus / line object exists and the intensity is the experience. Two sequential stimuli / lines would form an edge object – a change in reading based on the difference of intensity of the two. There would be many edge objects (changes), each with an intensity change. The experience of the change would be at a particular intensity but this would not be kept in LTM. We would be exploring objects and since different changes are different edge objects they would be explored.
Therefore, an important rule is that only a change in object type will attract attention. Thus, if a change in position or intensity is to attract attention there must be change objects / edges for them.
5th Nov, 2009 Reflections and Negatives
I need the generic line, edge and composites of these to be size, position, orientation (reflection) and intensity independent. Thus the absolute size, actual position, actual orientation and absolute intensity need to be part of the experience. The generic objects must therefore contain relative size, relative intensity, relative position and no orientation of their two parts. I don’t believe I want relative negativity. In a negative an object is only recognized due to it having a particular shape which is relative size and position of the parts. But it does mean that a very complex object should be recognizable due to the relative sizes and positions of its parts independent of the relative intensity of its parts. And maybe the reverse is also true. Recognize a complex object based on the relative intensity of its parts independent of their relative size and positions? I have also noted previously that the relative position of two parts gives its size. So if an object can be recognized by shape (relative size and position of its parts) then are shape and intensity two different trees of object types or is it similarity recognition based on shape while disregarding intensity?
In humans orientation is important for recognizing faces but not for recognizing a hammer. We have difficulty recognizing a face upside down but not a hammer. I would prefer to make the generic object orientation independent and thus an object can be recognized at all orientations / reflections.
But how does the experience information get represented? The object gives the 'what'. The experience contains the where, order (orientation), size and intensity. Absolute intensity may not be kept but relative intensity is. Absolute position may not be kept but relative position is. Absolute size may not be kept but relative size is. Actual orientation may not be kept but relative orientation is. For similarity recognition we need one or more of these to match. For exact recognition they all match. Should objects then be a P-Habit combination of the many relative dimensions? Each relative dimension then must have a primitive level-0 object type and a value. Then all objects become P-habit combinations of these. How would a line and edge be represented using this idea?
9th Nov, 2009 Representing differences
I’ve realized that for graduated intensity lines on a graduated sense I need to build independent binon trees for the differences in intensity, width and position. This is the same idea as I currently have for order and where for discrete valued stimuli on a discrete set of sensors. Then at each level in the tree I need to combine the 3 difference binons in a P-Habit. The intensity difference tree is based on the following table.
|Level||Absolute Intensity||Intensity Difference||Intensity Orientation||Consists of Levels|
|two lines||2||yes||yes||yes||0, 0|
|two edges||3||yes||yes||yes||1, 1|
|three lines||4||yes||yes||yes||2, 2|
|three edges||5||yes||yes||yes||3, 3|
Line’s and edge’s intensities are part of their experience and not their object. The same object must be recognizable at any intensity. The edge’s intensity would be the value of the left side line’s intensity. The edge’s intensity difference binon would be stored as the two values of the first instance found with that particular ratio, no need to reduce it to common denominators. The smaller intensity value would be first. The intensity orientation of the edge must be part of the experience so that the same edge object can be detected at either orientation.
Similar logic would apply to two lines and two edges. They may be adjacent or separated by a gap. Their left side object would be used for the intensity experienced. For two lines the intensity difference and orientation would be as determined for an edge. Two edges may be separated by a gap or not and the right side of the left edge may not necessarily be the same intensity as the left side of the right edge. The intensity difference of the two edges will be the difference in their intensities that are on their left sides. For three lines and three edges they will need to share a common line or edge experience consisting of an absolute intensity and intensity difference binon.
The same ideas have to apply to the width. This table captures the possibilities.
|Level||Width||Width Difference||Width Orientation||Consists of Levels|
|two lines||2||yes||yes||yes||0, 0|
|two edges||3||yes||0||0||1, 1|
|three lines||4||yes||yes||yes||2, 2|
|three edges||5||yes||yes||yes||3, 3|
Building a binon tree of width objects introduces the need to deal with zero values, which also have to be handled for intensities. Storing the two intensity values for the difference will avoid the divide by zero error and allow a 0 to 0 ratio to be equal to a 1-to-1 ratio. Also a 0 to X ratio should be different from a 0 to Y ratio where X and Y are different values.
For position the left side of the object will be used so there is no need to do any arithmetic to find a middle location. However since the position difference will always be the right position minus the left position the object orientation will always be positive.
|Level||Absolute Position||Position Difference||Object Orientation||Consists of Levels|
|two lines||2||yes||= width 1st line + gap||yes||0, 0|
|two edges||3||yes||yes||yes||1, 1|
|three lines||4||yes||yes||yes||2, 2|
|three edges||5||yes||yes||yes||3, 3|
10th Nov, 2009 Lines for edges
We often outline an area that has an edge by replacing the edge with a line. We can also do this by outlining it with any combination of lines and still recognize the area. The area inside the outline is recognized by its shape. In the case of a single linear dimension this would be equivalent to recognize a two or more lines that are the areas.
Changes in time must also be represented as binons. I will need binons for position change, width change and intensity change. These will have a positive or negative value. All possible combinations of these will need to be represented.
I’ve said the orientation is not part of the binon so that two objects with the same differences but different orientations can be recognized to be the same object. However for a three-part object I think orientation needs to be part of the binon. Let’s consider either width or intensity, it does not matter. The line binon will be A. The 1st line has a value of 1. The 2nd line has a value of 2. The two lines form a binon G with the value ratio of 1:2. The two lines have an experienced value of 1. A third line has a value of 9. The 2nd and 3rd lines form a binon H with the value ratio of 2:9. The experienced value for H is 2. Now the three line combination 1, 2, 9 is binon M with value ratio of 1:2 and is recognized as a G and H with the common 2nd line. All of these experiences have a positive orientation.
Now consider the three-line combination 9, 2, 1. The 1st two lines are binon H with an experience value of 9 and negative orientation. The 2nd two lines are binon G with an experience value of 2 and negative orientation. The three-line combo is binon M with an experience value of 9 and a negative orientation. The difficulty arises when dealing with the three lines 4, 2, and 9. The 1st two lines are binon G with an experience value of 4 and negative orientation while the second two lines are binon H with experience value 2 and positive orientation. The three-line combo will form a new binon N with the value ratio 4:2 and experience value 4 and a positive orientation. But the two parts are binons G and H.
A table might help to summarise this information and expose the challenge and help with a solution.
|Line Pattern||Binon||Binon Parts||Binon Value Ratio||Experience Value||Experience Orientation||Orientation Ratio|
|1||A||-1, -1||0 : 0||1||+ve||1 : 1|
|2||A||-1, -1||0 : 0||2||+ve||1 : 1|
|9||A||-1, -1||0 : 0||9||+ve||1 : 1|
|1, 2||G||A, A||1 : 2||1||+ve||1 : 1|
|2, 9||H||A, A||2 : 9||2||+ve||1 : 1|
|1, 2, 9||M||G, H||1 : 2||1||+ve||1 : 1|
|9, 2||H||A, A||2 : 9||9||-ve||1 : 1|
|2, 1||G||A, A||1 : 2||2||-ve||1 : 1|
|9, 2, 1||M||G, H||1 : 2||9||-ve||-1 : -1|
|4, 2||G||A, A||1 : 2||4||-ve||1 : 1|
|4, 2, 9||N||G, H||4 : 2||4||+ve||-1 : 1|
The value of the two parts is already incorporated into the binon’s value ratio. The solution would be to also incorporate the orientation ‘ratio’ (relative orientation of the two parts) in the binon as well.
I think I should be changing the levels as specified in the previous tables such that level 0 is for edges and lines and level 1 is for any pair of these while level 2 is for any triplet.
11th Nov, 2009 Discrete / independent sensors
With this new structure I need to represent stimuli patterns from independent sensors. The where information, which identifies the sensor positions, will still be needed and will work fine since it does not involve any orientation (positions are always left to right). All stimuli are a width of 1 so there is no need for width ratios or orientation (both would be 1). However the readings vary so relative intensity is needed. However, since all sensors are independent there is no orientation of the intensity (the 1st value must always be smaller than the second value). Thus 020, 002 and 200 will need to map onto the same binon as will 123, 132, 213, 231, 312, and 321. 123 will produce pairs of 12, 23, and 13. Only 12 and 23 will be combined because of the common 2. 321 will also produce 23, 12, and 13, which will only combine one way based on the previous rule. 020, 002 and 200 will all produce 02 and 00, which will only combine based on the common 0.
I’ve decided that the level should correspond to the number of lowest level stimuli that make up the stimulus. Thus edges and lines are at level 1 and pairs of these are at level 2. This corresponds to the levels used in S-Habits where a single stimulus has S-level of 1 and the first S-Habit combo of two stimuli has S-level of 2.
14th Nov, 2009 Position and Width
For width and intensity I am creating the relative width and relative intensity (ratio) objects with orientations. Both of these are based on the two parts making up the object. For position I need to keep the experienced position and create a position difference object, not a ratio of positions. The position difference is the width of the first part plus the width of the gap between the parts.
15th Nov, 2009 Position and Width
This means the position is the experience and the difference in position is the object. Since it is a position difference the values of the two parts should always be 0 & a position difference. Orientation is not relevant.
|Level||Absolute Position||Position Orient||Position Difference||Orient Ratio||Consists of Levels|
|Line||1||yes||1||0 : Width||1 : 1||0, 0|
|Edge||1||yes||1||0 : 0||1 : 1||0, 0|
|two lines||2||yes||1||Width 1st line + gap||1 : 1||1, 1|
|two edges||2||yes||1||0 : Dist||1 : 1||1, 1|
|three lines||3||yes||1||0 : Dist||1 : 1||2, 2|
|three edges||3||yes||1||0 : Dist||1 : 1||2, 2|
Discrete / independent sensors
The challenge now is to devise an algorithm to keep the reading binons in sorted order based on their parts. I will need a comparison routine that will tell me if one binon comes before or after another so that I can combine them in the right order.
20th Nov, 2009 Sorted Binons
So given the pairs of binons AB and AC I need to identify the common part (A) and then order them such that they are sorted by intensity reading. This also applies if the two pairs are AB and BC or AC and BC. Using the first example it is obvious that A<B and A<C. But also the common part in the AB pair is by definition sorted. Thus let us say AB is XSY where A is XS and B is SY. “S”, the overlapping common part is sorted. X and Y will be level-1 objects. For the AC the common part must also be S. A is XS and let C consist of SZ. Now based on the comparison of Z and Y it is possible to decide whether B (SY) comes before or after C (SZ). If B comes first we need to find the object that consists of AB and BC. If C comes first we need to find the object that consists of AC and CB. If the two original pairs are AB and BC then there is no need to do the comparison. If the two original pairs are AC and BC then AC = XTY and BC = ZTY. Comparing X and Z will determine whether A<B or A>B and from this it can be determined to find AB and BC or BA and AC.
This works if X <> Y <> Z. But if B = C then Y = Z. This occurs at level-3 when 090 is processed. AB =09 and BC (sorted) is also 09. We need to be given an indication of the common part and then the process can make the correct choices.
21st Nov, 2009 Common part of sorted binons
I have decided that I need to use the orientation information that is part of the experience of discrete sensor readings to identify the order of the readings of left and right side parts. The binon recognition routine (ExistBinon) can then decide which side is in common for the given pair of binons. – I tried it and it has problems too.
22nd Nov 2009 Independent sensors
I started working on independent sensors on the 6th June 2009. On 19th and 20th August 2009 I discussed the sorted order of readings. Now that I find it so difficult and complicated to represent a sorted order of readings for independent sensors I question whether we really need or are capable of recognizing the similarity of 9540, 4590, 0594, 5049 etc. when the sensors are independent. If they are truly independent then maybe no pattern similarity is needed. Each combination is a unique one. If the readings are graduated then, given the same sensors, ‘123’ is the same as ‘246’ and ‘241’ is the same as ‘482’. If the readings are symbolic then every set of values is unique.
23rd Nov, 2009 Shape recognition
If a shape with an edge, caused by a change in intensity, is to be seen as the same shape in which a line outlines the edge then position difference patterns for lines and edges must be common to both edges and lines. Also since the outline could be any combination of lines (at any level) or widths the position pattern must be level independent. The relative width patterns would still be different because the edges have no widths.
Graphically all of these linear shapes must be recognizes as the same.
-----xxxxxxx-----xxx----- -----|------|----|--|----- -----H------H----H--H-----
The H is meant to be two vertical lines.
For this to work the difference in position of two lines must be the difference in position of two edges. This also means that the position difference for a single line must use the same generic position difference for a single edge. Thus the position difference for a line cannot be its width; it must be zero just like an edge.
25th Nov, 2009 Change determination
I need to figure out how to determine change when it comes to the generic concept binons of position, width and intensity. Each generic concept binon captures a pattern made up of the relative values. They can be found in frame 1 and 2 and the change can be determined. The changes that need to be determined are change in position, width, and width orientation, intensity and intensity orientation. For the position experience there is only the one level 1 generic binon and each experience captures its position. To determine a change in position at level 1 I could separate the line and edge experiences to help determine what moved where. Other than that I can’t use intensity or width to help in the change determination. Any edge could move to any other edge or any line could move to any other line. There may be new edge and line positions and some may disappear. At level 2 the same relative positions can be found between frames and experiences of pairs of objects that have moved can be determined. To determine that two lines have moved independently you need to be able to identify them not by position but by width and intensity information. At level 3 the position generic binon is starting to exhibit a specific pattern of relative distance between the experienced lines. Now expansion and contraction should be possible in the experience but still use the same position generic binon. Also one would expect orientations to be possible and map to the same position generic binon.
Distance – separation
I have realized I needed a distance apart / separation generic binon which kicks in at level 2 when 2 lines or edges have been recognized. Then patterns of distances apart can be recognized independent of patterns of relative width and relative intensity. But I still need the position for each experience of these three patterns at all levels so I can determine their change in position. This position is no longer the experienced value for the distance apart. The experienced value is the distance apart for the left most parts of the generic binon.
27th Nov 2009 Position, Separation and Width
To address the problem documented on 23rd Nov, 2009 I have the following solution. The problem is how to represent the edges of a shape when replaced by lines or double lines. The idea is not only keep the position of each experience for detection of position change but also to keep the separation of each experience at all levels to compare with separations at other levels. The separation or distance between objects is the sum of the separation for the two parts. The gap is used as the second part at level 2. I have created the graphic below to try to capture the issue.
This resulted in matches of the separation at the next level up. In the case of the double line it matched at two levels. But maybe what I should really be using is the position and separation of the edges of the lines that outline the shape and not the lines themselves. This works better than the separation of the lines. If the separation / distance between the edges of an object retain the same ratios then the edges all belong to the same object. This will handle objects coming toward or away from you and objects moving left and right. It will also allow other objects to exist in the gap between two lines (consisting of 4 edges) and allow it to move independently but still combine the two lines into the same object. Maybe I can reuse the relative width binons for edges as the relative separations because relative widths don't make much sense for edges.
Level 1, 3 Lines with 6 edges
Input: -----xxxxxxx-----xxx-----xxxxx Positions: . . . Separations: ---A--> B-> --C-> Level 2, 1st 2 pairs Position: . Separation: -----D-----> A & B Relative Separation: ------> --> Position: . Separation: ---E---> B & C Relative Separation: --> ----> Level 3, A Combo Position: . Separation: -------------------> D & E Relative Separation: ------> --> ---->
Level 1, 6 Lines at edge positions
Input: -----|------|----|--|----|----| Positions: . . . . . . Separations: > > > > > > Level 2, pairs Positions: . . . Separations: ---A--> B-> --C-> Relative Separation: 1 1 1 1 1 1 Positions: . . Separations: --F-> --G-> Relative Separation: 1 1 1 1 Level 3, Combos Positions: . . Separations: -----D-----> ---------> A & F, G & C Relative Separations: ------>---->1 ---->---->1 Position: . Separation: -------> F & B Relative Separation: ---->-->1 Position: . Separation: ---E---> B & G Relative Separation: -->---->1 Level 4, Combos Position: . Separation: --------------> A&F & F&B Relative Separations: ------>---->-->1 Position: . Separation: ------------> B&G & G&C Relative Separations: -->---->---->1
Level 1, 6 pairs of lines at edge positions
Input: -----H------H----H--H----H----H Positions: .. .. .. .. .. .. Separations: >> >> >> >> >> >> 1,2,3,4,5,6,7,8 etc. Level 2, some pairs Positions: . . . . . . Separations: > > > > > > 1&2 3&4 5&6 7&8 etc. Relative Separation: 11 11 11 11 11 11 Positions: . . Separations: -----> ---> 2&3 4&5 Relative Separation: 1 11 1 Positions: . . Separations: ---A--> B-> 1&3 5&7 Relative Separation: ------>1 -->1 Level 3, Combos Positions: . . Separations: ---A--> ---> Relative Separations: 1----->1 1-->1 1&2 & 2&3 5&6 & 6&7 Position: . Separation: ------> Relative Separation: ----->11 2&3 & 3&4 Position: . Separation: ----> Relative Separation: 1--->1 3&4 & 4&5 Level 4, Combo Position: . Separation: -------> Relative Separations: 1----->11 12&23 & 23&34
Width and relative width for the patterns described above look like:
Level 1, 3 Lines with 6 edges
Input: -----xxxxxxx-----xxx-----xxxxx Widths: ------> --> ----> Level 2, 1st 2 pairs Width: --------------> Relative Width: ------> --> Width: ------------> Relative Width: --> ----> Level 3, A Combo Width: ------------------------> Relative Width: ------> --> ---->
Level 1, 6 Lines at edge positions
Input: -----|------|----|--|----|----| Widths: 1 1 1 1 1 1 Level 2, pairs Width: ------> Relative Width: 1 1 Width: ----> Relative Width: 1 1 Width: --> Relative Width: 1 1 Level 3, Combos Width: -----------> Relative Width: 1 1 1 Width: -------> Relative Width: 1 1 1 Width: -------> Relative Width: 1 1 1
As a result of this new approach what are the generic difference binons? I still need the absolute intensity in the experience and a relative intensity generic binon. This is position and width independent and applies to each edge at level 1 and pairs of lines at level 2. For position I now would have the absolute position for all experiences. I will need a relative position difference / distance between / separation between edges at level 2 and higher. The separation experience will have a separation value and an orientation. This could be kept as the relative width binon for edges. For lines at level 1 and combinations of them I will need the relative width binon. The width experience will have a width value and an orientation.
Now the question is will patterns of edges be combined with patterns of lines to form objects or will edges combine with edges and lines with lines and never mix? Let's consider the later idea. A line could be a combination of a line with width, width ratio (0), intensity, intensity ratio (0) and a pair of edges with positions, a separation but no relative intensities. Then two lines without a gap would be two lines with width, width ratio, intensity, intensity ratio plus three edges. The three edges would have relative distance to each other and it would be combined with a single edge intensity difference for the centre edge. Two lines with a gap would be two lines with a width, width ratio, intensity, intensity ratio and four edges. The four edges would be a pair of three edges with separation ratios but no intensity ratios. This will work well at the borders because these edges don't have relative intensities but do have positions and could participate in the separation ratios. Maybe we don't need the edge intensity ratio for the centre edge between two adjacent lines. Then edges are only used for separation ratios and lines are only used for intensity ratios. Representing three lines will then require 4, 5 or 6 edge combinations depending on the number of gaps.
Is it possible that the separation pattern for the edges can be represented just using the left edge of all the lines and thus have the same number of edges and lines forming an experience at any level? This would work except expansion and contraction would have to be detected by making sure the width and separation changes were the same. This would mean that edges are not kept as experiences with intensity ratios. Edges are used to determine the positions for the left side of a line and its width. Left side positions are used to form separation ratios and width for width ratios. Thus a line combination experience consists of a position, width and intensity experience and their orientations plus a combination of relative separation, width and intensity objects.
28th Nov, 2009 Change Determination
For a graduated sense I want to find the object in the second frame that has the same separation, relative width and relative reading objects but could be at a different position, at a different intensity or have the same change in separation and width. The three ratio objects must be the same else it is a different object. And only the sensors that have changed should be involved.
30th Nov, 2009 Change Objects - Sense of Time
Change over time is like edges in space. The change in reading over time forms a pattern of relative reading objects similar to the difference in readings between lines. The duration of the reading is equivalent to its width and the pattern of these relative durations forms a pattern of relative duration objects. The timing separation between the readings is equivalent to the spatial separation and these form a pattern of relative separation objects. These last two patterns can't be created yet because the sense of time has not been introduced. Although, duration of a reading (width in time) might be possible if we were to consider repetition (boring stimuli) as occurring at a regular time interval. Except this time interval is the one associated with paying attention to the next stimulus because nothing has happened to attract attention. The duration that is really needed is the one created by a timing signal of some sort.
It is possible that S-Habits are these change patterns for readings in which the sensor values are discrete. And if the readings are graduated then S-Habits become the generic relative reading change pattern objects.
10th Dec 2009 Object identification
The same object is found in two frames if the relative separation, width and reading patterns are the same. The absolute position, separation, width and / or reading may have changed. This means that I should produce a unique object representing the combination of the three relative patterns. It then has a position that can change.
After the same object has been found in the two frames then change objects can be created. For discrete / symbolic readings the change object is the S-Habit. This object has two object parts and the two values are both 1. For graduated readings the change object is like an S-Habit in which the two object parts are the same but the values are different. This difference represents the change.
12th Dec, 2009 Linear versus Logarithmic Input
The raw data values can range from 0 to a maximum value and must be integers (no decimal places allowed, valid values are 0, 1, 2, 3, 4, etc.). If the raw data values provided are over a wider range of values than 0 to 255 then the data is treated logarithmically. Examples are volume readings being mapped to decibels or frequency of sound being mapped to piano notes. The raw values go through a 2-stage mapping. First the value is mapped onto the logarithm of its value. For raw values 1 and higher they will be mapped to the range 0 to 254. Since the raw value 0 cannot be represented as a logarithmic value it is mapped to -1. The second stage of the mapping then adds 1 so that the -1 (raw 0) becomes 0 and 0 (raw 1) becomes 1. The comparison process will use subtraction between the values to produce the ratios. However if either or both of the values are zero because the raw data was 0 the algorithm below for treating zeros is used.
If the raw data is linear within the range 0 to 255 then no mapping will take place and the ratios between the values will use division. In this case 1:1 is the same as 2:2 and 3:3. 1:2 is the same as 2:4 and 3:6. However the value of zero is interpreted as no intensity and thus 0:1 is different from 0:2 which is different from 0:3. 0:0 will be interpreted as the same as 1:1.
13th Dec, 2009 Graduated Change
For graduated readings from one sensor the following approach should work. Given the series of values 4, 3, 7, 5 the same object is being measured / detected. Assuming that this object A has been habituated / permanent Adaptron should start paying attention to the changes. The following series should result.
Input reading: 4 3 7 5.
Object: A A A A.
Change Value: ? 4 3 7.
Change Object: ? 1 2 3.
Change Orientation: ? -1 1 -1
Where the change objects are:
Change Object # Part1 Part2 Value1 Value2.
1 A A 3 4.
2 A A 3 7.
3 A A 5 7.
4 1 2 3 4
Once the change objects have become permanent then the same sequence should collapse the sequence of change objects # 1 and 2 into #4. Brackets indicate what was in LTM before the second change occurred.
Input reading: 4 3 7 .
Object: A A A .
Change Value: ? (4) 4 .
Change Object: ? (1) 4.
Change Orientation: ? (-1)-1
15th Dec, 2009 Position & Separation
I believe position is used as the value for level-1 separation objects for graduated senses. It provides the distance / separation that the object is from the border. Then a two line separation contains these two values which when subtracted from each other give the separation / distance between them. For a discrete sense each sensor's position is meant to symbolically represent a source of readings. Thus the position itself is a separation object and the values are all irrelevant.
31st Dec, 2009 Sequence versus Change
Change is sequential but not all sequences are changes as detected by the sensors. This is illustrated in the test run from yesterday for test case #25. When attention changes from one sense to another or from one combination of independent sensors to another combination then we end up with a sequence in LTM that is not the same as a change. For example, let's consider 2 senses with stimuli A, and B from sense #1 and X, and Y from sense #2. The sequence A,B from sense #1 with no change in sense #2 produces the sequence the same as the change A,B. The same would be true for X,Y. But if the sequence is AX, BX the change is only A,B. In the case of test case #25 the memory sequences are X/,A and X,A. X/, X, and A have all become permanent and S-Habits have recognized the X/,A sequence and the X,A change as separate objects. Which attracts attention first? Or should the sequence X/, A be formed? Or should the sequence be processed only for the change part that it contains which is the X, A change? I think this later approach is appropriate because there was no change in the / sensor and that is why only the 'A' was attended to.
Generalizing this concept means that an S-Habit formed from any two stimuli in sequence in LTM is expecting the changes on the same senses and independent sensors between the two stimuli. In the case of AX, BX the S-Habit is expecting the A,B change and no change on sense #2. AX,B is also expecting the A,B change on sense#1 and no change on sense #2. AX,BY is expecting the A,B change on sense #1, the X,Y change on sense #2 and the AX,BY change for the combo P-Habit objects. The A,X sequence would result from a switch in attention from one sense to another and since there is no sense or sensor in common no change can be expected and no S-Habit can be formed. There needs to be a sense #2 stimulus with the 'A' or a sense #1 stimulus with the X to detect a change on either sense. Only then can one determine the impact on one of the senses from a response performed between the two stimuli. This is because action is meant to cause change and with the A,X sequence one cannot determine if attention switched between the senses due to attraction of a change on sense #2 or performing a past sequence.
The other possibility for the A,X sequence is to store in LTM the A stimulus and then the ?,X change binon that caused attention to be attracted to sense #2. This would mean the sequence could form an S-Habit expecting the change binon.