Is there any recent work on modeling how we rapidly acquire new knowledge?

Question

I work with neural network models of human cognition a lot, and one thing that bugs me about them is the timescale: they learn over thousands of trials whereas humans seem to learn after a couple examples. So I've been wondering if this discrepancy has studied, and if so, what terminology has been used to study it. Specifically, I'd be very interested to see some work trying to get neural network models to learn at a more human timescale.

A good example is learning to play a new video-game. We can do this in a few minutes if there is a tutorial explaining the basic game mechanics to us. However, modeling the same process would involve slowly strengthening perceptual-motor associations and could take a very long time indeed.

Note that I know about one shot learning, and that is not quite what I'm talking about. That literature mainly revolves around image recognition and learning after a single presentation, whereas I'm curious about rapid learning more generally. This sounds a little vague, but the question "why do our models learn several orders of magnitude slower than us?" seems specific enough to have come up in the literature.

Answer 1

Humans actually exhibit both slow and fast learning and they have somewhat different properties. One distinction is between "declarative" memory (for example, facts like "tigers have stripes" or "Paris is the capital of France") and "procedural" learning (such as perceptuo-motor skills like riding a bike or playing a musical instrument). Declarative memory lends itself to fast learning but procedural memory requires a lot of practice.

Perhaps even more relevant to your question is that neural network models are also capable of fast and slow learning, depending on the sparseness of the representations and tolerance for disruption. One of the foundational papers on this topic (McClelland, McNaughton, & O'Reilly, 1995) described a neural network model with two "complementary learning systems": (1) A fast-learning hippocampal memory system that relies on sparse representations and is not able to maintain previously-learned information when learning new information (sometimes called "catastrophic interference"). (2) A slow-learning neocortical system that relies on distributed representations and is able to integrate many different examples into an overall structure that can support generalization.

References:

McClelland, J. L., McNaughton, B. L., & O’Reilly, R. C. (1995). Why there are complementary learning systems in the hippocampus and neocortex: Insights from the successes and failures of connectionist models of learning and memory. Psychological Review, 102(3), 419–457.

Answer 2

The question of how "rapid" learning could be possible relates to Hume's problem of induction -- how can we learn so much from so little. Historically, in both philosophy and psychology, the solution has fallen into one of two camps: either some form of the knowledge was already there to begin with (a 'nativist' view), or we use statistical inference to slowly update our beliefs (some early connectionist models fit into this camp, though in the last 20 years, there are forms of connectionist models that might suddenly/dramatically 'shift' weights in ways that can lead to the appearance of rapid learning).

A solution to the 'rapid learning' problem comes from Bayesianism - the idea that we're doing 'backwards inference' to get to a model of the world. This inference to the best model is weighed by both the likelihood of the data given a model, as well as (and here's the important part) our prior beliefs about how likely the model was before we observed new data. Bayesianism is not just a middle ground between nativism and empiricism; the reason it provides solutions to 'rapid inference' is because you're integrating the idea of 'probabilities' into the mix. That is, rather than there being one absolute truth, we can think of the learner as weighing multiple possible truths - with some a priori belief in each; new data comes in and allows us to rapidly update those possibilities, based on the likelihood of the new data under each.

The Bayesian framework also allows for hierarchical inferences (each set of prior beliefs is given by a more abstract set of prior beliefs) which can tie in nicely to ideas like 'core knowledge' (which fall under the nativism camp), while still providing a learning mechanism (which fall under the empiricist camp).

There is long standing research in vision and language which cover these Bayesian ideas, but in the last decade an explosion of research has come out of labs such as the Tenenbaum lab at MIT, the Griffiths lab at Berkeley, etc. that show how other cognitive/higher-order reasoning (such as causal inference) might be rapidly acquired under this framework. There are several overview papers on these themes which give background on the Bayesian approach, as well as empirical predictions that fall out of specific Bayesian models (many can be found by visiting the lab websites of the aforementioned individuals).

Bayesianism is complementary to models that include additional constraints, such as pedagogical inferences (see Pat Shafto's work) and other pragmatic inferences (e.g. see Mike Frank or Noah Goodman's work) -- all of which to lead to even more rapid inferences from limited data.

Answer 3

The question is slightly confused I feel, so I am not sure there is a good answer.

Neural network models are evolutionary and so possibly analogous to the evolution of the human brain (where learning takes place extremely slowly), not to simple human learning.

The brain does not learn skills by evolving a neural network to perform that skill (that would just be too slow). It is probably something like: the brain processes information then encodes its conclusions into memory (neurons).