Fisher Task Distance and Its Application in Neural Architecture Search

We formulate a Frechet-type asymmetric distance between tasks based on Fisher Information Matrices. We show how the distance between a target task and each task in a given set of baseline tasks can be used to reduce the neural architecture search space for the target task. The complexity reduction in search space for task-specific architectures is achieved by building on the optimized architectures for similar tasks instead of doing a full search without using this side information. Experimental results demonstrate the efficacy of the proposed approach and its improvements over the state-of-the-art methods.