Towards Meta-meta-learning

Abstract

The ability to generalize prior knowledge and experiences to a novel problem is a hallmark of human intelligence. This “learning how to learn”, known as meta-learning in machine learning, could help make machine learning models have similar generalization capabilities as humans while using much less data. Results from cognitive science suggest that humans model the world and conduct inference via Hierarchical Bayesian models (HBM), which enables human meta-learning.

We aim to model how humans do meta-learning in machine learning models by distilling the inductive biases of an HBM into neural networks. We do this by training neural networks using a popular meta-learning algorithm called Model-Agnostic Meta-Learning (MAML) using data sampled from an HBM. We find that some neural networks were able to obtain priors similar to an HBM and generalize on a task an HBM is capable of. However, the neural networks do not behave like an HBM overall. Thus, we think that meta-learning as we have implemented is insufficient for modeling human meta-learning, and we may need meta-meta-learning, for which we provide an algorithm.