Let’s first start to define:

What is learning?

I think one could define learning as the acquisition, consciously or unconsciously, of new information or let’s say knowledge about the world, our environment, our routine, the people around us, things we need to know how to do, things we need to remember, in order to reach our goals and adapt to our environment. Learning usually requires an improvement (i.e. an increase) of performances.

As an example, an unconscious type of learning is the basic common grounds about the world or about our environment, as what is usually found in a kitchen, how to grab a cup of coffee or what the usual distance between two steps (in the stairs) is so that you don’t need to look anymore while going down the stairs. Our brain processes this information unconsciously for us. However, if one of those examples does not match what you’ve learned – unconsciously – you will then have an immediate reaction of less than a second.

For instance, if you see a bed in a kitchen or if the last step is lower/higher than the other steps in the staircase of a building you are going for the first time, your brain would immediately send you a “message” (you will have a conscious reaction) to let you know that something doesn’t match what you have previously learned (the usual distance between two stairs), your expectations, your knowledge about the world or about your environment.

To give you another example. Let’s imagine you come back to your apartment after work and something (e.g., your furnitures’ configuration) changed compared to when you left it on the morning. You nevertheless didn’t change anything in your apartment for a long time so that you brain processed it as your usual “environment/routine” around you. This unexpected change in your flat would therefore create a reaction: you will immediately notice something has been modified, even if you were not aware that your brain – out of your consciousness – processed the previous configuration of your flat as the usual one.

This way, our brain can immediately perceived a mismatch between what it has previously learned (i.e., what it expects; this is called top-down processing) and what it senses (through our senses (vision, audition, smell, touch) sending information to our brain, in this particular example through the eyes, and called bottom-up processing). In resume, if what your senses sense does not match what your brain expects/learns, then you have a reaction, like “Oh what’s going on here?!”.

A conscious learning could be how to play a music instrument, golf or tennis, or how to write a poem in a classical form. In this case, it is easy to understand that learning implies performances improvement. A lot of practice would lead you to get better and better at it. Let’s take the example of a 2-year old child who learns how to eat and drink on her own for the first time. At first, she will probably spill the glass of water and make quite some food fall while trying to grab it from her fork, fork that she also just learned how to grab and hold. However, you will show her how to do so in a proper way, gently, you might also want to guide her while she is trying to do it; that way, she will improve. She will probably still struggle for a while but usually after a week or two she should be able to do so in a much better way, till completely learning it, therefore further doing it as a routine behavior. This part of the learning process requires a lot of practice and feedback, whether our own feedback or the one from something or someone else, till we are able to successfully complete the task at hand.

What does “a computer learns” mean then?

A “computer learns” means that it is able to acquire new knowledge based on previous similar knowledge acquired or based on learning rules and self-play. I will develop those two aspects separately below.

By self-play I mean that a computer plays against itself, using a method of trial-error, i.e. starting/trying from random/from scratch, getting feedback from it thus learning from it, trying again, getting feedback from it, gaining even more knowledge, i.e. learning more what has to be done, trying again, till succeeding.

For example, how to climb the stairs in a staircase or ride a bike in the case of a robot. In the field of AI, this concept of self-play is pretty recent, as the computer is still implemented by learning rules but not with data anymore. The old terms given before the self-play revolutionize the field of AI were expert system and decision tree. I do not like particularly those terms but they were used to mimic the role of an expert and help in the process of decision-making.

These two types of “learning” (as a machine, not as a human) are either data-based or self-play based. At least so far. Ideally, a computer would be able to learn as humans by being able to learn “on its own” new knowledge about the world, to adapt to new situations (without human intervention), i.e. to “learn how to learn”. These concepts would be developed in more details below.

As a positive note about humans: ever since Homo Neanderthalensis, 400.000-300.000 years from now, who was able to survive 20.000 generations before Homo Sapiens came, according to experts, “primitive humans” were already able to adapt to their environment to survive for such a long time the theory of evolution by natural selection stated by Darwin in the 19^th century. Therefore, these humans, even such a long time ago, were already more “intelligent” than the best AI computer nowadays as they knew how to adapt to their environment, hence already knew to “learn how to learn”!

How does a computer “learn”?

A computer learns by several ways, using to do so different methods. Either by supervised learning, which is data-based, unsupervised learning, which consists in figuring patterns, thus also called pattern recognition, or by reinforcement learning, which is based on learning rules and self-play, as already defined above. A computer can also learn by a combination of supervised and reinforcement learning.

Recently, it has been shown that a computer can learn only by reinforcement learning, which is already a more advanced version towards human skills; a step forward but still not any close yet to compete with human race abilities. At least so far and as far as I know.

Here you have reached the end of my first post. Hope you enjoyed it. Do not hesitate to share any comment or ask any question! To read more about “Supervised Learning” you can click here.