CMT Group Meeting
This event is part of the Condensed Matter Theory Seminar Series.
Deep Learning as Renormalization and Holography
Deep Learning is one of the most promising Machine Learning and Artificial Intelligence (AI) techniques for extracting important features from large data sets ( http://www.nytimes.com/2012/11/24/science/scientists-see-advances-in-deep-learning-a-part-of-artificial-intelligence.html?smid=pl-share&_r=2& ; http://www.technologyreview.com/news/524026/is-google-cornering-the-market-on-deep-learning/). Deep Learning employs a neural-inspired architecture with multiple layers and are often trained using an iterative, layer-by-layer algorithm. Despite their tremendous success, the basic logic behind Deep Architectures is not well understood. Here, we show that there is an exact mapping between Deep Learning and real space Renormalization Group techniques from physics. Furthermore, we show that the resulting Deep Architectures have a natural interpretation as holographic duals. We use this correspondence to give a general proof of the AdS-CFT correspondence. We illustrate these ideas using the 1D and 2D Ising Models.