设计神经网络架构.pdfVIP

DESIGNING NEURA WORK A RCHITECTURES USING REIN MENT LEARNING Bowen Baker, Otkrist Gupta, Nikhil Naik & Ramesh Raskar Media Laboratory Massachusetts Institute of Technology Cambridge MA 02139, USA {bowen, otkrist, naik, raskar} @ 6 1 0 2 At present, designing convolutional neura work (CNN) architectures requires v both human expertise and labor. New architectures are handcrafted by careful o experimentation or modified from a handful of existin works. We propose a N meta-modeling approach based on rein ment learning to automatically gen- 0 erate high-performing CNN architectures for a given learning task. The learning 3 agent is trained to sequentially choose CNN layers using Q-learning with an - greedy exploration strategy and experience re y. nt explores a large ] but finite space of possible architectures and i tively discovers designs with im - G proved performance on the learning task. On image classification ben arks, the L agent-designe works (consisting of only standard convolution, pooling, and s. fully-connected layers) beat existin works designed with the same layer types c and are competitive the state-of-the-art methods that use more complex [ layer types. We also outperform existing meta-modeling approaches fo work design on image classification tasks. 2 v 7