Bowen Baker is completing an M.Eng in EECS at MIT and is on route to San Francisco for an internship at OpenAI. Over the last year he has primarily focused on search problems relevant for training deep neural networks.
Designing convnets require both human expertise and labor. In this session we will cover MetaQNN, a Q-learning search algorithm. This approach uses an agent that is presented with the task of sequentially choosing layers in a CNN. Most of this work is described in Designing Neural Network Architectures Using Reinforcement Learning, ICLR 2017 (https://arxiv.org/abs/1611.02167), that said we will cover recent results not included in said paper.
Search is extremely computationally expensive. While MetaQNN is cheaper than vanilla methods, it still can be impractical in large scale situations. A recent paper entitled Practical Neural Network Performance Prediction For Early Stopping (https://arxiv.org/abs/1705.10823) details the use of regression models and empirical variance estimates to achieve speedups on search problems by a factor 2-6x.
6:00pm: Doors open for checkin
6:00 - 6:30pm: Food & Drinks
6:30 - 7:30pm: Presentation
7:30 - 8:00pm: Shmooze