neural networks, are the networks that  Inspired from brain of human. and fuzzy logic

Inspired from Inference of human. another ways, such as Q-learning is inspired from behavior of human.

for example, learning in brain of a human (children or older man) is a network that can learn anything with perception, train, test, evaluate many and many times to learn.

for learning these methods, u can study machine learning & soft computing courses.

in Artificial Intelligence, have other ways that inspired from nature, for example genetic algorithm, bee colony, ant colony and other.

for learning these methods, u can study evaluation algorithms course.

NLP (Natural Language Processing) is similar to the human way of learning. Please consult my book. The web link is given below:

https://www.researchgate.net/publication/272888259_Intelligent_Tutoring_System_Changing_Role_of_ICT_in_Education

Book Intelligent Tutoring System: Changing Role of ICT in Education

Regardless the availability of many machine learning algorithms inspired by some hypothesis/theory of how our brains learn, I believe we are still far from having a fully understanding of how we (or children) learn things.

Human learning is so vast and diverse that we only are scratching the surface of understanding how it works. Neural networks and deep belief networks roughly model how a neuron based brain works. This model is not exact, since it does not model the 3D proximity of the neurons. This may have significant impact, for example if you aim at modelling creativity.

We are probably on our way to understanding at a low level (neuron level) how the brain learns, but we still do not understand how to model higher level cognitive functions. For example, how our short-term and long-term memory interacts to perform learning. We are starting to learn how to model some functions, for example motor control, as Vijay is doing research on. There has been recent impressive papers on how a deep belief network can learn to play archade games, but this is still far from modelling learning in a human-like way.

The question is broad and general, which makes it interesting to consider.

I agree with Nils.

Moreover, it depends on what you are learning and how best you learn. For example, some people learn to memorize music by memorizing how the printed page looks visually. Then they "read" the music off of the mental image of the printed page. Others (such as myself) memorize music by listening to how it sounds and forming an auditory "image" of the music. With a "tape-recorder" memory, I  can "play" back this image any way I want mentally, and edit the tempo, key, instrumentation, or any other aspect of the music. This method is superior because it works even if there is no printed page to see, which is the case when inventing new compositions. Does anyone know of a single AI technique that could capture these very different learning methods?

Learning spatial reasoning is different from learning music, logic, language, or chemistry. Some learning is difficult to define because the people who do it best cannot describe to others exactly how they do it. Lacking a description may pose an obstacle to developing some AI techniques, but it may not be a problem for other techniques.

Conversely, some AI techniques may be able to show experts WHAT they are doing even if the experts cannot define how they do it. Presenting data sets analyzed in different ways can give experts and researchers who study them insights into various approaches to knowledge management, of which learning can be considered a subset.

Knowing the subject of interest will help to select the AI technique that is most applicable. Some AI techniques will work better in some domains and others will work best in other domains. Some domains also do not have well defined AI techniques either because no one is interested in those domains, or because the domains may be outside of the scope of most AI techniques.

My suggestion is to search for review articles on AI techniques and note the domains, if specified, in which the AI technique(s) work best. The result of a study like this could provide input for a review article reporting which domains have been studied for AI techniques, and which domains lack published AI techniques.  Anyone interested in co-authoring a paper??

Sounds like an interesting paper, Marion. I would be happy to contribute with what little knowledge I have.

You proposed music as one learning domain. Reading music off a printed page is a job that is well suited for a computer, but we do not really know how this processed in the brain. Learning and understanding music is somewhat related to understanding language, since similar areas of the brain trigger for music and language. However it is language analysis (voice to text analysis) that traditionally has got most attention by the AI community. Hidden Markov Models (HMM) is one technique that has been successful for this task. Voice-to-text analysis is however only a decoding operation. It does not solve how we learn or understand music or language.

An interesting book in this respect seems to be "Guide to Brain-Computer Music Interfacing":

http://www.springer.com/us/book/9781447165835

Hello Nils, you wrote:

"Reading music off a printed page is a job that is well suited for a computer, but we do not really know how this processed in the brain."

I agree with the first part and most of the second part. A pattern-recognition program could scan a printed page into a computer, but this would not lend us insight into how humans would process the results. For example, some people who play instruments read music by associating notation on the printed page with hand, finger, and in some cases, foot positions on an instrument. Tablature notation is particularly well suited to capture exactly where to put your fingers on fretted stringed instruments.

This works fine until you study classical singing. The voice is the most difficult instrument, partly because you can't take it apart, analyze it, and understand exactly how it works on a mechanical basis. Tablature won't work here. You need staff notation. Singers need to learn music by associating notes on the printed page with pitches and then memorize the pitch sequence.

language learning and music are related as you have observed. Some languages are tonal languages. Also, studying one language helps to learn another, particularly in the same language group. What would be interesting is to find an AI technique that could capture the structures and contexts of language and music learning, as mastery of both relies heavily on contexts.

The advantage of voice-to-text analysis is that it enables the researcher to use computational linguistics approaches. 

What can be challenging about AI is that is, an AI technique may perform well on a task but still not provide an insight as to how the learning was achieved, thus impeding comparisons to human learning. A technique may be able to capture enough context and structure in a domain to predict successfully an outcome of a question or problem, such as "Complete the sentence in language XYZ" or "Complete the following musical phrase" but the mechanisms that yields success in AI may differ considerably from the mechanisms that humans use.

Then there is the challenge of capturing how humans learn. It is possible to memorize with no understanding of theory. This is "wrote" learning. A better way, at least for those people who use it, is to figure out how to solve a problem based on theory and principles each time it needs to be done. Both methods may yield the same result for different reasons. Can we find an AI technique that can help to distinguish between the various learning methods?

Scientists throughout the ages have endeavored to understand how the human brain works. Frequently, it has been compared to the most complicated device known at the time. Historically, it was compared first to a clock, then a computer, and now we try to understand it with AI techniques because this represents the state of the art so far. Could it be that we are approaching what limits the scope of what we can achieve with AI? I would like to think that we can make further advances, but I fear that Yasser is right in his observation that we are far from a complete understanding of how human learning takes place.

I agree with Marion and Yasser. We have made impressive progress in use of AI, and will continue to do so in the future, but we should not oversell these techniques as being more brain-like than they are. We are far from understanding how human learning takes place, and are only scratching the surface on understanding how quite simple functions work (e.g. the grid cells which acts as the brain's internal GPS, how information from the eye is being decoded, the coarse function of parts of the brain etc.).

Generally speaking, AI techniques try to mimic human perception, learning and reasoning to solve complex problems. 

You said "How do children learn how to call things e.g. “this is a tree, and this is a flower” . This maybe regarded as a kind of pattern recognition so you can use available techniques such as artificial neural network for it.

Confabulation theory is an interesting theory. The theory contains many hypotheses and assumptions that are yet to be proven, however it is a plausible model of how simple thought processes like concept dictionary lookup may work. It is however not clear how confabulation theory explains more complex cognitive functions. For example how learning actually occurs, how time and space awareness is implemented, motor skills etc.

Thank you, Osama, for posting the interesting article by Hecht-Nielsen. The main problem with this and also with other cognitive theories that I have seen, including the one I published regarding mental pattern management, is that it is difficult to devise an experiment that will distinguish between the different proposed thought mechanisms. This may be because each theory has some merit, that is, it focuses on a particular aspect of thinking. It also could be that experimental technology has not kept pace with theories. Whatever experiments we can design do not seem to be fine-grained enough to yield specific enough observables. This has been a frustration for researchers involved in the study of expertise, decision making, leadership, as well as cognition.

From what I can tell, we may need multiple theories to explain different types of cognition. From Hecht-Nielsen's abstract, "billions of distinct, potentially viable, conclusion sets are considered. At the end, only one remains."

Here, Hecht-Nielsen has implied a significant reduction in the search domain for a single thought. This appears to be very similar to the reduction of search domains, which is a factor in expertise. Experts automatically exclude from consideration anything that cannot be a solution to a particular problem. Novices cannot exclude as many possibilities and thus have a larger search space. Finding a solution is harder if your search space is large. This is one reason why experts are better at making decisions than novices are.

Even experts cannot explain how they make decisions because their thought processes are so automatic that they don't know how they do it, especially when the available data do not linearly lead to a particular "logical" conclusion. It is hard to explain the role of insight, and just what insight is.

This brings up the next question: "What constitutes a "thought?" Do we have a generally accepted scientific definition of "thought?" Where does one thought stop and another start? Can complex thoughts be divided into smaller component thoughts or is each though "atomic?" We all think we know what a thought is as we have them all the time, but does one definition support or conflict with another definition?

If I can find my paper on expertise and pattern management, I'll post it on my profile. 

I totally agree with the first two answers. But I will try to give an specific contribution to the discussion.

Regarding reading and writing learning, there are at least two different theories (one from Sociology and Politics - Socio-interactionist theory - also called "Direct Way": from the World, from the learners' social context  to the critical reading and writing, in opposition to the Oral way, which stars by Phonic way and by B+A = "BA" and so on... - See www.lecture.org for more of this; and the other from Neuroscience, associated to Stanislas Dehaene, which states to have built "The new science of Reading" - and of how to learn to read and write.) I am used to practice Action-Research (consistent with the first one), but I also recommend Dehaene's book as an interesting reading.

The relatively new field of "Developmental Learning", inspired by the work of the psychologist Jean Piaget on cognitive development in children, is a very promising approach for learning in robotics (particularly humanoid robots). According to this approach, robots learn by interacting with their environment. For instance they can learn the consequences of their actions and even the limitations of their own body. Another cognitively plausible AI technique is Case-Based Reasoning/Learning that allow machines to learn by experience. The basic hypothesis is that similar problems might have similar solutions. Then, when facing a "new" problem, a CBR system finds out the similarities with some previously solved problem and adapt the past solution to fit the new problem. Each newly solved problem is a candidate to be retained for future reuse, therefore the experience grows. Reinforcement learning is another example of technique that has similarities with how we learn. Learning by imitation is another one. In summary, there are quite a few learning techniques that are somehow "bio-inspired".

I would like to commend Alexander for this most interesting question and thank all who have contributed. Finally, I found my publications list on another disk. Attached are two cognition-related papers. They do not solve the problem entirely, but like the Hecht-Nielsen paper, I hope they will provide food for metacognition and perhaps stimulate other researchers to share their ideas so we can all look at this problem in a different light.

I agree with Ramon that CBR and the bio-inspired learning techniques are important and very different ways of learning.

One thing we have not brought into the discussion (yet) is the chemical nature of the brain and how it differs considerably from the "dry" environment of computers. Various chemicals in the brain, such as dopamine, and even Na+ are involved in brain function, a situation that we have not even tried to duplicate in AI primarily because we don't understand it well enough to do so. The chemical nature of learning is demonstrated well by the bio-inspired learning. Ants deposit pheromones on their way to find food. The pheromones decay with time, a key feature of the learning process. Ants can find the shortest pathway to the food among many complex 3D pathways by following the strongest scent of the pheromones. The scents of the longer pathways diminish over time.

I would like to add to the discussion the aspect of "firmware" in the brain, that is, a set of abilities (and knowledge?) which are already available to a newly born living being (widening learning from humans to living beings in general) and which are probably essential for survival especially during the initial stage.

However, from this also arises the question of when the formation of a cognitive capacity begins (at what point during gestation in the womb?).

Thus, we could say that nature "bootstraps" the learning process, which then enters into the stage of knowledge adquisition per se which may be by example/counter example, contextual situations, etc., and is influenced by the quality of the learning environment, examples presented and mentors (especially the parents).

There are different interpretations, see by typing "bootstrap learning" into "google scholar".

So, employing ad-hoc AI techniques would not be enough to learn human-like. An integrated, multistage "framework" which mimicks some of the above stuff would get closer, in my opinion....

For an AI techniqe similar to the human one see: www.intelligent-systems.com.ar/intsyst/robotBrain.htm

Dear Mr. Fritz, thanks for your helpful answer and the very interesting paper. I had a look on your paper and found out that “A concept is the basic element of thought”.

Now I’m interested in how the concepts are learned and I identified three key questions for me:

1)      Does every new observed situation/pattern lead to a new concept or when is the observed situation/pattern worth to be put into a concept?

2)      When is a new stored concept worth getting a label?

3)      How are the concepts labeled?

Would be great if you could help me with the answers!

Hello Alexander,

You have posed some very interesting questions, all of which relate to metacognition and expertise.

1. Every newly observed situation does not necessarily lead to a new concept if it can be categorized into an ontological framework that contains a concept that appropriately characterizes the situation. If enough new situations are observed that cannot be thus characterized and categorized, it may lead to the development of a new concept. One can say that a pattern of newly observed situations that have enough similarities could instantiate a new concept. This holds true no matter whether a person is learning a new concept that others understood already or whether the person is a researcher breaking new ground.

2. If a new concept cannot be categorized into any existing category in one's mental ontology, it may be worth creating a new label for it. It may be worth noting that a single situation or concept may fit into multiple locations in the ontology. For example, "green dress" may fit into the collection of objects that instantiates the color, "green" and it also fits into the ontology of concepts that characterize clothing.

3. Concepts are labeled in terms that the thinker understands. Experts tend to have short labels because they use specialized terminology whereas a novice will use the definition of the term, not knowing that there is a one-word term for the concept. For example an expert will use the label, "pleiotropy" where a novice will use the label "the observation that a single gene or gene group determines multiple seemingly unrelated characteristics."

Novices who aspire to become experts try to expand their vocabulary. This really amounts to expanding the number of labels from which to choose when observing a new situation Whether they know it or not, they are expanding their ontology.. Experts can categorize a situation much more accurately and succinctly than novices because their mental ontology is much more developed and elaborate.

Hope this helps,

Marion

Hi Alexander,  Answer to your first question: In seeing, if the situation has the same 12x12 pixles as the old one, the old number is used. If not, a new (the next) number is asigned.

Answer to second and third question: The number asigned is the label.

Best regards, Walter Fritz