Information Theoretic Meta Learning with Gaussian Processes
This provides a unified framework for meta-learning that could benefit researchers in few-shot learning, though it appears incremental by building on existing gradient-based methods.
The authors tackled meta-learning by formulating it through information theory concepts like mutual information and the information bottleneck, learning task representations that predict validation sets, and demonstrated competitive accuracy on few-shot regression and classification problems.
We formulate meta learning using information theoretic concepts; namely, mutual information and the information bottleneck. The idea is to learn a stochastic representation or encoding of the task description, given by a training set, that is highly informative about predicting the validation set. By making use of variational approximations to the mutual information, we derive a general and tractable framework for meta learning. This framework unifies existing gradient-based algorithms and also allows us to derive new algorithms. In particular, we develop a memory-based algorithm that uses Gaussian processes to obtain non-parametric encoding representations. We demonstrate our method on a few-shot regression problem and on four few-shot classification problems, obtaining competitive accuracy when compared to existing baselines.