Machine Learning Day:
Towards Automated Deep Learning
Machine Learning Day
AI/Machine Learning/Deep Learning
Big Data Analytics
Performance Analysis and Optimization
TimeWednesday, June 27th9:15am - 10am
DescriptionDeep learning has recently celebrated substantial successes, due to its ability to learn in an end-to-end fashion from raw data, without the need for manually-constructed features. However, since the performance of deep neural networks heavily depends on the chosen network architectures and their hyperparameters, the manual work previously required in feature engineering has merely been shifted over to manual engineering of architectures and good hyperparameters.
In this talk, I will discuss modern methods for effectively searching in this combined space of architectures and hyperparameters, thereby paving the way to fully automated end-to-end deep learning. Specifically, I will discuss various speedup techniques for Bayesian optimization that go beyond the blackbox optimization setting (sometimes resulting in 100-fold speedups in finding good hyperparameter settings) and techniques that exploit network morphisms to search through the space of neural architecture up to 1000 times faster than previous methods. Finally, I will also present competition-winning practical systems for automated machine learning (AutoML) and outline some of the current challenges of the field.
Professor of Computer Science