Learning to learn

摘要

Empirical evidence shows that infants 10 months of age can learn about 10 times faster than infants 2 months of age that a novel entity is socially contingent. This suggests that during the period from 2 to 10 months of age infants became better learners. One possible explanation for this change is that new brain structures grow, in a genetically predetermined manner, that support more efficient learning. An analogy for this point of view would be the increase in mastication efficiency due to the growth of teeth. An alternative hypothesis is that the increase in learning efficiency is itself the result of a learning process that operates on the time scale of months. Under this view, better learning is the consequence of learning itself. Here we explore the plausibility of the "learning to learn" hypothesis from a computational point of view. We show that with standard reinforcement learning algorithms using an internally generated reinforcement signal it is possible to develop agents that progressively learn to learn within a period of months. The results fit well at a qualitative level empirical evidence regarding the development of social contingency detection in infants. The learning techniques that we explored have potential application for robots that learn to learn on their own.