Towards practical neural network meta-modeling

Abstract

This thesis primarily focuses on the topic of efficient automated procedures for convolutional neural network (CNN) architecture search. We first introduce a novel approach for CNN architecture architecture using Q-learning, a popular value iteration algorithm from the reinforcement learning community for sequential decision problems. On the task of object classification, the Q-learning agent outperforms all human crafted models that are similar to those in the search space. By analysing the underlying weights of the agent, we are also able to uncover some of the design principles that the agent learned during the search process. Reinforcement learning is generally very sample inefficient; in the case of architecture search most approaches require thousands of unique models to be trained. In the second part of this thesis we introduce simple sequential regression models (SRM) to predict final performance of a candidate CNN from partially observed learning curves. We use these performance predictors and empirical variance estimates for practical early stopping of online optimization procedures. Our SRMs are state-of-the-art in performance prediction and early stopping.