Parameter Search Transfer Learning

Abstract

Deep learning approaches have had success in many domains recently, particularly in domains with large amounts of training data. However, there are domains without a sufficient quantity of training data, or where the training data present is of insufficient quality. Transfer learning approaches can help in such low-data problems, but still tends to assume access to sufficient source domain data and a sufficient signal for transfer. In this work, we propose a novel approach for transfer learning called Parameter Search Transfer Learning (PSTL) which directly searches over parameters of a neural network in order to minimize the impact of low training samples in both source and target domains. Across Reinforcement Learning (RL), Regression, and Classification tasks we demonstrate that PSTL meets or exceeds the performance of transfer learning baselines, which we hypothesize is due to its ability to identify a better gradient.