Generalization or specialization

From generalization to specialization
Intelligence uses generalization at first. Intelligence shall use hasty generalization. Hasty generalization is the best strategy whenever we are lacking experience:
Similar stimuli are supposed to be processed in a same way, leading to the same conclusions and to the same actions.  Is there a lack of experience on a given field? Even if stimuli are not so similar they will lead to the same conclusions too.
This kind of generalization shall be the default behavior of every intelligent process. Proximity is taken into account in order to generalize.
Later, inaccurate inductions will be adjusted by experience: adapted responses will be associated to the now-identified special situations.
Fewer and fewer, intelligent systems evolve from global rough responses to more sophisticated ones.
From samples to concepts

From several different samples, an intelligent being builds concepts. New concepts are partially inaccurate. Future experience refines concepts into more extended ones (aggregation) or more specific ones (excluding exceptions).

Recall

Knowledge can be represented as strings and trees of nodes connected together. In order to retrieve information and obtain a logical behavior these strings and these trees are scanned in every direction. A thought is a flow of linked recollections. A recall starts from sensors. Thought may follow its process a long time after sensors triggered an initial recall. But at least from time to time sensors trigger recalls. Thought from previous recalls may interfere with new triggered recalls. A reader receives stimuli from vision sensors and they interfere with previously induced thoughts. However an explanation of a recall process shall be explained from sensors in order to remain as simple as possible.
The following process may probably work:
A bunch of stimuli is memorized in a time line. This bunch activates a node.  This node becomes a template. This template will be dropped or reinforced in the future.  All new similar records will be attached to this reinforced node. The recall process propagates both forward and backward.
- Forward from sensors to higher levels.
- Backward from higher level to sensors and actuators.
The forward propagation is a kind of hypothesis and needs confirmations by a sufficient amount of active entries. The back propagation looks like a conclusion. True or not, anyway it shall be taken into account.
If this conclusion fails, and the upper level fails too, then a new template shall appear.
The final conclusion may contain sequences of elementary actions.
A dynamic graph and a simulation  would help: more in a few days.

Thought

Forecasting future by replaying the past.

A thought is a sequential activation of linked pieces of information. Information may have been stored as crisp unique experiences. Multiple crisp past experiences have often been mixed up together into templates representing a fuzzy mean value.
From present activated stimuli or from already activated ideas, templates are activated. This activation tries and moves forward through the upper levels of memorized knowledge.
Several kind of levels shall be specified:
- Time scale levels, from tenth of seconds to hours.
- Association levels, from local sensor groups to large associations.
- Abstraction level, from crisp automatic response to philosophy.
In order to reach the upper levels, this activation needs support from the lower levels. Several rules shall drive this activation process.
- A node is activated if a sufficient amount of its inputs are activated.
- An activated node may include non-activated inputs.
- Non-activated inputs in an activated node represent predictions or actions to be done.

A thought is triggered either by a stimuli or by another thought.
A thought is driven by past experience.

Thought and percolation threshold

Persistence of thought
Thought may persist within a human brain a long time after the last significant amount of stimuli fired it up. For example, you read this phrase, close your eyes and think of your own experience about thinking... And now I think of you thinking about thinking. Human brain is very strong at this kind of role playing exercise. These strange skills take part in self-consciousness and deserve a whole dedicated chapter.

 Our present topic needs a simpler example of thought:
You are lying back on your bed in the darkness and you are planning your next week-end. During several minutes, you are building a project. Maybe you will postpone this project and you will resume it tomorrow.
Paths of thought
This continuous thought moves from an idea to the next one. Thought may be considered as a flow of small pieces of knowledge linked together. This recollection itself is memorized: we can recall our past thought. An intelligent being shall link together all different events in order to take into account a complex world. A path can be found from every idea to every other one. However, when we plan our next week-end, we avoid a  total recollection of our past  holidays. We avoid an entire recollection of our whole geographical knowledge. Not all information is activated. This is a matter of proximity. If a part of knowledge is close enough to the present thought, then this part is more or less activated. What happens if activated knowledge is not enough? The thought process is broken. What happens if too much knowledge is activated? A huge amount of irrelevant and useless path is to be searched. An optimized thought would be just below the system calculation power limit.
In order to obtain an efficient thought we need:
- Clever activation rules in order to trigger knowledge activation steps.
- Clever real time control rules in order to monitor the total amount of path being explored. This feedback loop is compulsory.
Without this feedback loop, we encounter a percolation phenomenon.
Thought can be compared to a blaze in many ways.

Percolation
A forest fire needs propagation conditions. Let us suppose there is no wind. Consider tree density as a main propagation condition. Below a given threshold of tree density, the fire decreases because of gaps remaining between groups of trees. "Islands" of trees are all surrounded by paths of bare ground. If you outreach a density threshold, a fire grows up more and more, because of paths of trees linking most groups of trees together. The configuration has suddenly moved from islands of trees to islands of bare ground.
This kind of phenomenon is called percolation.
Thought proceeds in the same way. Below a given threshold, thought needs external stimulations to be fired up. Above this threshold, thought runs through the entire mind as a fire does in a dry forest.

Online demo applets of percolation
This simple applet comes from this page. Try it out first.
Another applet on this page.
A nice applet on this page simulates forest fires. This is worth playing a while with its control panel. You can run it without reading comments.
You will see how tree density alters fire propagation. A slight modification on this tree density from 0.58 to 0.62 induces big consequences on fire propagation (tree density=probability cursor).

Learning, induction, deduction

Learning process starts from induction.
For example, I discover a chair. I learn that such a thing is a chair.
But what is "such a thing"? No previous experience.

I suppose I may ask questions about it. Next time I see a chair quite different from the first one I probably don't recognize it as a chair.

But I ask and I get an answer again. Later, I recognize many chairs without any help.

From different chairs, I have built a general idea of what a chair may look like.
Induction
Inductive reasoning is moving from a set of specific facts to a general conclusion. This type of reasoning leads to over-generalization. Not all seats are called "chairs". Learning process tends to remove these errors through more experience.
Deduction
Deductive reasoning reaches a conclusion by following a logical inference from general rules applied to a specific fact.
Back to the first example of a chair. Deductive reasoning may be used in order to validate or invalidate a guess.
"From my point of view, this is a chair." How to be sure? I use general rules.
This must be of a given size and shape allowing human beings to sit on it...

Induction is in process whenever experience is building up new knowledge.
Deduction is working whenever knowledge is being used.
This assertion remains true within science domain. Theories are validated from inductive reasoning. Deductive reasoning applies once theories are validated.

Knowledge

Knowledge is the sum of all added information from the birth of an intelligent being. What exactly is information for an intelligent being? Footprints of past days. In order to be depicted as "information", these footprints must be "computable" by the entity.
Knowledge is different from skills. Knowledge is used to enhance skills. Not all skills come from knowledge. Not the whole knowledge is used to enhance skills.
Knowledge shall be stored. This storage must save available resources. How exactly is knowledge stored within an intelligent being memory? Clues have been given by neurosciences but the whole system complexity still remains globally out of reach. However, knowledge management can be studied from its external appearance.
First, knowledge shall be stored.
Moreover, knowledge shall be re-usable. A past experience has to be compared to a new similar situation.
No new real event is entirely similar to a previous one. The storage of knowledge shall ease approximative comparison and classification. 
Of course, neural networks rather well perform such a task.
The ability of generalization is a powerful feature of evolved intelligent being brains. Generalization operates within time and space.
Storage
Stored knowledge starts always from stimuli.
Are elementary stimuli close together in space and in time?
If they are close enough in space and in time, this proximity is recorded as an elementary knowledge.

On most sensor events, only time proximity is significant. The sequence doesn't matter. However chronology  is recorded on a time line and can be used later if relevant. The elementary sense of cause and consequence comes later from the sense of sequence.
Recall
New stimuli activate stored elementary knowledge. Stimuli must be numerous enough and located in a zone of sensors previously involved in a stored elementary knowledge. The recall can be triggered even if new stimuli are slightly different from stored stimuli.

Competitive recall
If several elementary pieces of knowledge in the same zone are activated by new stimuli, then they will compete. At a given time in a given zone, most activated piece of knowledge will be considered as most suited for the present situation.

Self-learning

Self learning ability is a key feature of an intelligent being.
However, if you test an intelligent being within a limited short time, self learning doesn't immediately appear as an essential component. Anyway, you may conclude by a quick test that you are more intelligent than a monkey; a monkey is more intelligent than a rat; this rat appears as  being more intelligent than a lizard; a lizard is more intelligent than a clam!
Comparing Man and Chimpanzee, please, don't be too condescending against Chimps. Otherwise you would be a little bit surprised if I selected the following test: Look at this video: a monkey performing a short-term memory test.
An intelligence test gives an evaluation of a total amount of skills, including genetically inherited behavior, learned behavior, learned knowledge, problem solving skills, conceptualization...
Back to learning and self learning:
Learning includes memorization. Inherent value of learning is re-usability.
Personal experiences may apply whenever similar circumstances occur.  A mere memorization is not reusable. Taking into account the huge amount of possibilities, there is never two identical situations. Learning process and memorization include approximation and global description. This is a condition for re-usability of experience. The need of a global fuzzy description of a past experience is a first step on the way to conceptualization.
Self learning requires self marking of learned patterns.
Learned patterns are used to forecast the consequences of an action.
If this forecast fails, the pattern will be downgraded from good forecast to bad forecast.
But this pattern was just selected because it was in a way similar to the present situation. In order to make a difference between the new situation and the ancient one, there is a need of complete memorization of each situation beside the need of global and fuzzy memorization. Thanks to memorized details of the old situation, the learning process tries and shapes the differences between the new situation and the ancient one. Both remain fully memorized for a future comparison to another situation, if needed. However, the content of learned patterns lies in fuzzy contours just sufficient enough to underline the differences between two situations. If patterns were crisp and sharp, they were never applicable to any other new situation. These patterns may link to acute recalls, but learned patterns are of a different nature. However, "acute" recalls may be partially rebuilt from patterns linked together with short-term memory. This is a known issue for eyewitness accuracy (more about this topic). This is another story.


If you liked the above video there is a longer one about the same topic: Ayumu the Chimp.