Statistical thinking in machine learning
Padhraic Smyth, UC Irvine
12-1pm 9th Jun 2017
Abstract
Recent work in machine learning (for example in deep learning) has led to significant advances in areas such as image recognition, speech recognition, online advertising, and ranking of Web search results. The first part of this talk will be a general discussion of the role of statistics in modern machine learning. Statistical theories and models have long provided a foundational basis for many of the techniques used in machine learning. But even for machine learning approaches which appear on the surface to have no explicit probabilistic or statistical semantics, such as neural networks or decision trees, there are underlying links to statistical ideas. The second part of the talk will explore these links in a specific context, focusing on modeling of categorical sequences, with a discussion of recent ideas in sequence modeling that combine techniques such as Markov models, self-reinforced random walk models, and recurrent neural networks.