Can someone explain what is the process of learning? What does it mean to learn something?


That is a wonderfully fundamental question.

Learning is the use of a system to change another system so that, instead of doing what it did before, which may have been nothing, it does something else.

In the human brain, the system is the way that genetic expression caused the directed mutability of that brain so that human intentions and responses to external stimuli would alter. The direction of alteration is based on incentives and deterrents. In classes of children, there are things a teacher may do or a culture in the school that motivate learning. Misbehavior is deterred through other mechanisms. In that case, the first system is the educational process and the second system is the brain's ability to use what it was taught.

Even without classes, children learn things that create comfort, so comfort is the internal incentive. It is wired into us as a kind of teacher to dislike groin moisture, so we learn to use bathroom receptacles. It is wired into us to like praise from parents and teachers, so we tend to perform to get it. Things like that are complex losses and rewards and primary systems to incentivize learning.

Learning in artificial networks is not nearly as complex as in human brains. In some cases the artificial results are better. In other cases the abilities of the human brain cannot be approached by artificial networks yet.

The functioning of an artificial network often begins arbitrary and completely useless, but it's a parametric function, meaning the function can be changed by changing numbers called parameters. Each time the function is used to process an example drawn from a data set, the result is evaluated and the evaluation is used to modify the parameters. Care is taken to not over-modify the parameters, otherwise confusion can occur. Mathematicians call this type of system confusion chaos.

Repeating this carefully directed and moderated process ideally leads to something called convergence. The learning system is set up so that the result of convergence is a set of parameters that minimize losses, maximize gains, or both.

Sometimes initial learning is not enough and there are other related things to do later.

  • Adjusting learned behavior to adapt to new conditions or information
  • Unlearning things that no longer produce benefit so that new learning can replace them
  • Relearning things that had faded from memory
  • Developing greater confidence in what was learned so that unlearning requires greater impetus

There are additional technical terms for the above concepts, more details that can be grasped, many categories of learning system types, varieties within those, and the mathematics that was used as a foundation to construct all this and make it work. Because the question was fundamental, those details and technicalities were omitted.


Turing had much difficulty explaining how a computer could learn, and in his 1950 paper asked, How could a machine learn? Its possible behavior is completely defined by its rules of operation (program), whatever the machine's history (past, present, future) might be. His proposed solution was ephemerally valid rules, whatever they might be - and he doesn't say.

So learning can be understood as the machine acquiring new behaviors but not by the behaviors being programmed into the machine by a human.

Perhaps a better way to look at this is causally. A human can define the causation (possible behavior) of a computer by programming it into the machine. But this is a case of the human using their knowledge to define how the machine will react to situations. Learning is the case where the machine itself acquires now behaviors or possible behaviors (not by a human using their knowledge of the world). And any such alleged intrinsic learning can easily be tested. If it helps the machine survive in a complex and hostile world, then it really is learning.

  • $\begingroup$ To be precise, Turing did not ask "How could a machine learn?", but "Can machines think?", at least, in section 1 of the paper. I didn't yet fully read the paper and there's a section "7. Learning Machines". In any case, I don't find the exact question "How could a machine learn?" in the paper. Maybe you can clarify the specific part of the paper that address the "learning". Yes, sure, the imitation game is framed as a way of assessing "human-like behaviour", but that doesn't necessarily imply learning. $\endgroup$ – nbro Jan 24 at 18:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.