AAS 390 Chapter Notes - Chapter 1: Network Architecture, Incorporated Council Of Law Reporting, Reinforcement Learning
Document Summary
Richard shin & charles packer & dawn song. The successes of deep learning in recent years has been fueled by the development of innovative new neural network architectures. However, the design of a neural network architecture remains a dif cult problem, requiring signi cant human ex- pertise as well as computational resources. We evaluate our methods on the udacity steer- ing angle prediction dataset, and show that our method can discover architectures with similar or better predictive accuracy but signi cantly fewer parameters and smaller computational cost. Several recent works have treated neural network architecture design as a reinforcement learning problem (zoph & le, 2016; baker et al. , 2016; zoph et al. , 2017). While these approaches have successfully found interesting architectures for highly-studied benchmark problems, these methods require training (tens of) thousands of models from scratch and testing them on a validation set.