Designing Neural Network Architectures Using Reinforcement Learning

We model Convolutional Neural Network (CNN) architecture selection as a Markov Decision Process. The system can be distributed across many machines and generates models that out perform all others that use the same layer types. Caffe code for the top models generated can be found on my github, and the paper can be found on arxiv or the openreview site for ICLR 2017. I recently gave a presentation on the paper for the CSAIL Vision Group, and our work was recently featured in the MIT Technology Review!