The design section of this book provides a logical description of how Adaptron is built so that it can emulate most of the requirements described in the Requirements Analysis. High level design is called Architecture. This section describes the framework and structural components of Adaptron and how these parts relate.
The detailed design describes how each component functions in this first version of Adaptron. The Architecture for Adaptron is fairly stable and fundamental to its ability to learn and think. The detailed design is where substantial improvement can be achieved through the application of more efficient and effective algorithms. This is where heuristics could be added that would improve performance substantially.
Design must be done using some form of medium / material. In this case it is not a connectionist neural net.
Map all stimulus channels (senses) into one sense input channel - is this feasible.
But attention can be paid to different senses and it is the sense that helps determine how to use the info.
Voice noises are communication and require language recognition, while squeaks and bangs do not.
Associations are made between different sense stimuli.
To use a constrained, well defined and simplified environment. The environment and the interactions with it have been intentionally simplified so as not to have to deal with the complexity that these can introduce but focus on and investigate the primitive principles / functions involved in learning and thinking.
25th Dec 1979 - The environment being used at present [ in my simulation] is a static one. Only new input occurs if an output has occurred to change the environment. A more realistic environment is a dynamic one. Without any output being done the input changes - my reflexive motor on duplicate inputs will no longer function in a dynamic environment which switches between two inputs.
Only use one dimensional environment - should be able to exhibit learning and thinking in this environment.
Representing Complex Stimuli - 117 - In this way the first frame of that concept will come to be the concept and the frame might not be enough to represent the concept.
4th April 1996 - Subconscious action!
While executing - conscious is thinking interrupts on
- inconsistent input with expected
- finished task
- unable to obtain required stimuli
Learned sequence relegated to parallel processing at particular concentration level.
How much of subconscious sequence saved in LTM when interrupt occurs, so it can learn from new situation - just inconsistent stimuli plus reaction? Or a little bit more to give context? How is the new learned fragment integrated into a sequence which can be executed? Does it need to be executed once more to lay it down in LTM? The most recent version replaces (due to recency factor in recall) the older one. This implies it must store all subconscious action sequences every time performed - or at least some moving windows worth of it (STM length?).
+ The human brain seems to record concepts / distinguishes concepts by a change in sensory input. It seems to record each picture that it sees. so if eyes stay absolutely still it records that picture. only when eyes move to new spot on scene being viewed does it record another picture. Thus it is a change in the sensory input that causes a picture to be recorded. Either change due to environment of due to agent. Also a new picture recorded when staring at a constant scene if one changes one's attention and back. + A test for this is to look at one spot and have no change in attention and see how many times you can record that concept i.e. a picture. + Concept formation: How many concepts are made of a constant tone or a still picture? I feel the duration before a second concept is formed can vary considerably when the input is constant. An exercise to understand this is to stare at one spot for a long time and try introspection of your attention and concept formation of the spot. One loses attention toward it and then pays attention to it again thus recording another concept. Similarly with a tone.
+ Consider recognition: of a tune say if it is slowed down or speed up. The concept of a note. How long does the sound have to last before you make a second concept of it?
+ Just as "an input is only stored when attention is attracted and attention is only attracted by a change in the environment" a sequence of inputs should be ignored until something different - a change - occurs. This means attention is attracted to a change in the expected. - Likewise only the change is recorded. Does this also apply to emotions, only recorded when a change in emotional state takes place.
+ Internal ideas should be one character long. I.e. each character represents a concept. Somehow input and outputs can be many characters long. Output handled by processor to do automatic output once initiated. Recognition must be handled by memory recall. Really though these many characters sensory data need not be handled since the character data is really recorded only at change or attraction of attention time.
Attention is reflexive but also controlled. Pay attention to something.
+ the attention is a device which does something just like a hand is. It receives commands on what to do and when to do them. It starts off doing the things because of a reflex action built into it at making time which is that attention is attracted to any change. This starts out as a reflex and all the changes of attention are recorded as though they were actually done by the centre BM even though they were basically reflexes. After a while the BM will start repeating the switches in attention as though it had been doing them all along (it will start due to some cue and because reward was associated with it in the past.) + The attention has a reflex action, anything different coming in on a sense the attention is switched automatically and the command is recorded in central control. This can be over-ridden by command from central control to switch to a sense. When central control is reading off memory into memory reflex in attention is operating but when central control commands attention to switch to something it will switch and reflex will not be able to switch attention while it is being recorded.
+ Then in reverse recall and attention appears to be able to work together to be given an identifier / concept of a thing and then provide specific properties of it. Or given a property to look for, identify things/concepts that satisfy the property.
+ To pay attention to something - form a mental image of what is wanted and it matches the real world seen object that matches the mental object (experienced).
+ Form an image of the goal and satisfaction obtained when solution matches the image of the goal.
+ Input only stored on attraction of attention but it is also stored on purposeful behaviour when we purposely look at something - execute a sequence at a high level of concentration. Often we are looking for an expected thing to happen.
- How many inputs/outputs do you process before your purpose (goal) is not realized?
+ Only becomes more discriminatory with input when it is more meaningful i.e. cues to punishment or reward.
+ How will the brain mechanism take-over and control actions. In the case of the attention if the command to switch attention comes from the BM this has priority over and above the reflex action which might attract attention away from what it is turned to.
15th July 1992 - Let the real world represent itself (model itself) which it does in the form of the experiences recorded in memory.
- 3 - To move or speak it has a recording of all it can do and it just combines these recordings.
Intelligent agent needs motivation to do anything - what are these - not survival.
How about to learn and think?
Interest as a drive. - might be possible by automatically associating pleasure with input or thought which is new - not recognized. The intention being to stimulate the search for new things.
- A possible curiosity drive might be constructed by having output whenever two sequential attentions and data are identical. i.e. a drive to always maintain a changing sensory input.
Some possible boredom detection techniques.
1 New Input - interest
2 Same input - familiar expecting previous associated input - store idea of input
3 Next input - not familiar - not same as expected before(idea of input) -> interest --> any associated output, input etc. go to 1 or 2
- same as expected - reflexive motor
4 Same input - familiar - associated output - store idea of output - do it.
- expecting some input so go to this process #3 - not familiar
- In this scenario - no punish or reward - always trying new things on environment so never have a repetitive input stream.
Cause and effect
- Goals are inputs. We have situation + action -> outcome. The outcome is the goal. The situation is an input and an action is an output. The outcome is a situation also. ==>> Input + Output -> Input. Memory has many such sequences.
Subconscious Action Processing - A goal is the recall of a final state recognized by a stimulus / feature pattern that would be reached by executing a sequence of responses - a habit. So before the subconscious action process is started you have an image of the final stimulus state and the process recognizes completion and stops when that final state is recognized. Attention is attracted to the completion and you get the feeling of satisfied the goal.
- There is the need for the retention of a goal state. I have [in my model] an increase in the level of attention-interruption-concentration as a result of a good thought but no remembering of the goal state trying to be achieved - unless that is implicitly stored in the position of execution.
- The mind sets up a goal and associated pleasure to be activated when goal found. Then goes about purposeful search to acquire (reach) the goal.
How learning could take place. -
Let us explore this motivation to learn or this drive to avoid boredom in a little more detail. Boredom occurs at a conscious level when your attention finds the stimuli boring. How many times do you have to experience the same stimuli to find it boring? What do you do when you are bored? Do you try new behaviour is boring situations to see what will happen?
We require a more accurate and useful definition of boredom because any artificially intelligent agent is not going to have any significant bodily needs to satisfy. When your expectations are the same as your previous expectations and you are not doing something with a purpose. When your inputs are the same as your past experiences and your thoughts are the same as previous thoughts and you are not doing something with a purpose, at a level of concentration.
- Interest is the drive. Interest in what to do (if anything) when situation recognized. Interest in what to do (if anything) when new situation. Therefore a new situation provokes attention - observation - by default second time observed - its familiar probably has an associated input. but that is all (implies it is expecting something) Can do a number of possible things.
1/ reflexive motor NO!
YES! 2/ execute associated input -> listen for it. expect it.
If input different - new situation
If input same as expected - then reflexive motor.
- The ultimate desire of the mind is to generate sequences of responses which can be executed on the cue of an input. The sequences would be the final product of the learning process and would work correctly. The learning mechanism must still be ready, though, to learn new things if the situation should change.
Recall - Associations
10th Feb 1985 - The following is a list of the expected improvements to the simulation from incorporating recognition of a series of "ideas" rather than just the one "frame".
- - Recognize context and act differently on same cue depending on the context.
- - Recognize a series of inputs - act upon recognition of whole series not just single input.
- - Keeps all series that match 'active' so easier to identify which is same - easier to identify if its a new situation and therefore needs attention.
- - emulates hierarchy of inputs as mentioned 21st March 1981.
- - Helps produce generalization and over discrimination as in 21st May 1981.
21st May 1981 - Need to add intermodal associations - i.e. between different input sources - touch - sight -sound etc.
- Generalization and over discrimination require that objects have a number of properties that are common or different. This implies that a single concept is insufficient for generalization and over discrimination. This means my current model can not generalize.
- I need an association which is the next sight, next sound, next idea, next output etc. which will attract attention just like the associations already generated. This gives locality of associations - remain in the context of the original cue and help in the sequentialization of the order of doing things.
- Associations should be made based on the assumption of cause and effect relationship between events.
23rd Feb 1995 - A modification to the pattern matching of a sequence (series) of stimulus / responses over time to obtain the best match and provide the best context to the recognition and recall process and where to execute is to allow stimuli to match over a time sequence independent of the responses matching in a sequence. see 10th Feb 1985.
29th May 1993 - With reference to 29th Nov 1972 CAT has many associated concepts whereas XCQ does not. 29th Nov 1972 - 177 Each concept can appear to have a varying amount of information stored in it. The word CAT can be thought of in one concept where as XCQ is thought of in 3 concepts. Maybe a reference is stored instead of CAT.
We need a smile / cry output state indicator so the external world can react upon it. This is just like a care-giver who detects her child's mood by the way they act (cry, smile etc.) and then s/he reacts to this. On a robot maybe a primitive equivalent is a red (unhappy) and green (happy) light. Both lights off would be a possible state.
Feelings control the flow of thinking + behaviour.
Could put these instinctive reactions in memory so they can be overwritten based on learnt behaviour rather than hard-wire the instinctive reactions and not allow any override, or divide reactions into two classes: those that must be in program and those that can be in memory. e.g. reaction to pain in hard codes, reaction to "have no match in memory" in memory.
Feeling of Boredom and the ability to detect it and instinct (currently hard coded) to do random output when recognized could be put in memory.
How it operates / cycles
23rd Aug 1989 - 1/ Should stop doing past experience script as soon as data input not (same) consistent with script being executed.
2/ However to introduce variety into output I have it so we stop doing as soon as data input is same as script being executed, and keep on executing script when data input is not consistent (same) as script being executed.
1/ and 2/ is a contradiction in objectives. 2/ happens when concentration level=0 i.e. non-purposeful execution. 1/ happens when purposeful doing of script, concentration level>0.
7.2 Detailed Design
(The following is a process like description of what the sub-mind and memory must do to perform a habit.
- Receive a command from the mind to perform a habit.
- Find the command and habit sequence in memory. Recall the first response, response pattern or sub-command and its expected feedback stimulus.
- Perform the response, response pattern or sub-command in the sequence. If it is a sub-command perform it by starting at step 1 with the sub-command.
- Pay subconscious attention to the sense and feature(s) required to detect the expected feedback stimulus.
- Obtain whatever feedback stimulus is actually available on the given sense and feature(s) and compare it to the expected one.
- If the feedback stimulus does not match the expected one then interrupt the attention (mind) and let it know about the failure.
- If the feedback stimulus is the same as the expected then identify the next response, response pattern or sub-command and the next expected feedback stimulus.
- If there is no next response and stimulus then stop processing with successful execution.
- If there is a next response and stimulus continue at step 3.