Learning
Typical definition of Learning: the acquisition of knowledge or skills through experience, study, or by being taught.
Wikipedia Definition of Learning: the act of acquiring new, or modifying and reinforcing existing, knowledge, behaviors, skills, values, or preferences
Proposed Definition: Learning is a probabilistic search of solutions to a problem, with qualitative feedback
The typical definitions are difficult when we want to separate learning out from other similar or more general activities. For example, if I buy a book, in some sense I've acquired knowledge. Or a more interesting example might be copying a digital book to a hard drive, did the computer "learn" anything? Another problematic example is when an animal is born with instincts. Those instincts were acquired at some point during it's life. Or if it doesn't display those instincts immediately, but has instinctual behavior triggered at some point, when was the knowledge of what to do acquired? In a sense that's acquired knowledge or skills, but I wouldn't necessarily call that learning.
I would say that just the act of learning, at its most minimum, is essentially guess and check - we try something new and see if it works. We might be trying to figure out a new skill (golf, computer programming, painting, playing piano, etc.) or trying to learn facts (what's George Washington's middle name, what's 24x756, etc.) or trying to solve more complicated problems. But the basic act of learning is looking for a solution to a problem, or more generally a response to a stimulus, and checking for feedback. We either see if there's a good result, or check the result against a desired target (a good golf swing, or the correct product of two numbers). If the result is good, we've found a good solution. I don't believe that learning would be possible without feedback, so that's a required part of the definition.
I think that there's undoubtedly at least part of our brains (and other animal brains) that work this way. They connect stimulus to response and are able to be adapted over time, which would qualify as a form or learning. If we accept that learning is an unconscious way of storing stimulus and response pairs, then we can think of the different ways that could be accomplished.
We can imagine a very simple animal that just has one set of sensory receptors and each individual receptor is simply hardwired to stimulate a particular response - when triggered it flees light for example. And we can imagine more and more of these simple 1-to-1 sensory-to-response connections being made, each working independently of each other. Eventually though for animals with a lot of detailed senses and a lot of possible responses, they would end up with lots of redundant neurological wiring. A better system would be to find areas where there was overlap and create a network that shared data transmission resources and used logic to make sure that the appropriate stimulus-response pairs we're still connected correctly. This would in essence be a form of data compression.
As a hypothetical example we can think of a neural network in a digital animal that connects the stimulus of 4 binary sensors with a 7-segment number display, and the purpose of this neural network is to convert the numeric value from binary to the display. A simple and straightforward way of doing it would be to have 10 neurons, each with 4 dendrites to sense the binary stimulus and 7 dendrites to active the appropriate segments of the display. If we measure the complexity of each neuron by how many boolean logic gates it would take to 'program' it, the average would be around 4 each, for a total of at least 40. This is simple, but not very efficient way of encoding these stimulus-response behavior. For comparison, a typical binary to 7-segment decoder can be made with just over 20 logic gates.
That simple hypothetical seems to have a lot in common with the kind of neurological structures we see in actual animal brains. There's a huge amount of overlap, while still efficiently connecting stimulus and response pairs. If those pairs are fixed, that's usually what we'd call instinct, but if there's a process in place so that the network can change connections to maintain most existing pairs while adding or updating new ones, that sounds like learning.
This kind of 'compression' would also lead to a de facto set of predictions for the kinds of responses that would be triggered by similar stimulus that haven't been observed before. It may be that the response to a new stimulus is close enough to previously learned ones to be useful, or it could even be that a new stimulus is essentially the combination of two previous stimuli and that the way the responses have been learned/compressed creates an adaptive prediction of the correct response for the new situation.
Of course a predicted response could just be an artifact of the 'compression' and might not be adaptive at all, in which case feedback would lead to rewiring that attempts to preserve the previous 'correct' responses while integrating a new response in to the existing 'compression' scheme. Which is essentially a description of learning given this perspective. What kinds of new responses are able to be learned would be dependent on how similar they are to existing responses and the cost/benefit of learning the new responses as opposed to relying on a 'good enough' response. Cases in which it's not evolutionarily advantageous to override a large body of compressed/learned responses to integrate a new and unusual response could be thought of as behavioral biases - consistent non-optimal responses.
Thinking of the brains of animals this way would lead to some testable predictions. We'd expect animals that have more complex environments and/or more complex learned behaviors to have to have larger brains to encode the larger set of responses. It also would make sense that the entire brain wouldn't be wired this way, there would be areas that are hardwired as instincts. We'd also expect to see predictable patterns in the way stimulus-response pairs are encoded/compressed.
We can imagine a very simple animal that just has one set of sensory receptors and each individual receptor is simply hardwired to stimulate a particular response - when triggered it flees light for example. And we can imagine more and more of these simple 1-to-1 sensory-to-response connections being made, each working independently of each other. Eventually though for animals with a lot of detailed senses and a lot of possible responses, they would end up with lots of redundant neurological wiring. A better system would be to find areas where there was overlap and create a network that shared data transmission resources and used logic to make sure that the appropriate stimulus-response pairs we're still connected correctly. This would in essence be a form of data compression.
As a hypothetical example we can think of a neural network in a digital animal that connects the stimulus of 4 binary sensors with a 7-segment number display, and the purpose of this neural network is to convert the numeric value from binary to the display. A simple and straightforward way of doing it would be to have 10 neurons, each with 4 dendrites to sense the binary stimulus and 7 dendrites to active the appropriate segments of the display. If we measure the complexity of each neuron by how many boolean logic gates it would take to 'program' it, the average would be around 4 each, for a total of at least 40. This is simple, but not very efficient way of encoding these stimulus-response behavior. For comparison, a typical binary to 7-segment decoder can be made with just over 20 logic gates.
That simple hypothetical seems to have a lot in common with the kind of neurological structures we see in actual animal brains. There's a huge amount of overlap, while still efficiently connecting stimulus and response pairs. If those pairs are fixed, that's usually what we'd call instinct, but if there's a process in place so that the network can change connections to maintain most existing pairs while adding or updating new ones, that sounds like learning.
This kind of 'compression' would also lead to a de facto set of predictions for the kinds of responses that would be triggered by similar stimulus that haven't been observed before. It may be that the response to a new stimulus is close enough to previously learned ones to be useful, or it could even be that a new stimulus is essentially the combination of two previous stimuli and that the way the responses have been learned/compressed creates an adaptive prediction of the correct response for the new situation.
Of course a predicted response could just be an artifact of the 'compression' and might not be adaptive at all, in which case feedback would lead to rewiring that attempts to preserve the previous 'correct' responses while integrating a new response in to the existing 'compression' scheme. Which is essentially a description of learning given this perspective. What kinds of new responses are able to be learned would be dependent on how similar they are to existing responses and the cost/benefit of learning the new responses as opposed to relying on a 'good enough' response. Cases in which it's not evolutionarily advantageous to override a large body of compressed/learned responses to integrate a new and unusual response could be thought of as behavioral biases - consistent non-optimal responses.
Thinking of the brains of animals this way would lead to some testable predictions. We'd expect animals that have more complex environments and/or more complex learned behaviors to have to have larger brains to encode the larger set of responses. It also would make sense that the entire brain wouldn't be wired this way, there would be areas that are hardwired as instincts. We'd also expect to see predictable patterns in the way stimulus-response pairs are encoded/compressed.