I have just enjoyed reading a review article in Science entitled Foundations for a New Science of Learning. It does a very nice job of synthesizing a variety of different threads ranging from machine learning to psychology and neuroscience. The central thread is the importance of interactions or, more specifically, the notion of social learning.
A few of their comments are gratifyingly close to viewpoints in my own current work. For instance, they point out the importance of imitation and transfer, but they also note that it doesn’t suffice to do this at a shallow level. Then, in order to have skill transfer at a deeper level, one needs processes that carefully combine imitation and active experience or exploration.
However, one point that strikes me when I read things like this is that these arguments are largely descriptive with little in the way of normative explanation. They are certainly intuitive, but what is so fundamental about these specific mechanisms? So, although I see the practical value of implementing these features in robots to show intelligent behaviour, I would find it a bit more satisfying if these insights were supported by an underlying theory. To give an example from a related area, computational neuroscientists are realizing that most motor control processes can be described well in the language of optimal control and estimation theory. I know from my own training that many of the computational details of these operations, e.g., Bayesian filtering, derive their justification from deeper principles in the calculus of variations and stochastic systems theory. What are the corresponding principles underlying these more general forms of Learning (with a capital L, to indicate learning in open-ended “large world” scenarios)? A few weeks back, I was flipping through Itzhak Gilboa’s book entitled Theory of Decision Under Uncertainty, which provides an elegant overview of the foundational issues in decision theory. It’d be great to figure out how one can lay a similarly principled foundation for the kinds of things Meltzoff et al. are talking about in the above Science paper – showing the necessity and essential structure of these constructs that one observes in behavioural studies.