R&D researcher focusing on applied Machine Learning, i.e. biometric applications, affective computing, signal processing. Holds a PhD in theoretical physics and worked as a researcher in Germany and Israel (Max-Planck Society and Weizmann Institute of Science), developing theory and algorithms for various complex-systems problems spanning from infection propagation in social networks to design of novel quantum states and magnetic materials. His current research interests concern statistical physics of neural networks and application of tensor network algorithms in machine learning.
Topic: Deep Learning with physics insights
Short Description: High capacity of Deep Learning models is often linked to the massive number of redundant degrees of freedom which complicates the training. A plethora of hand crafted neural network architectures were proposed to achieve the best performance for the data and problem at hand, but such architectures are not necessarily the most optimal one. Recently the connection of NN with statistical physics,has been revived: training of DNN’s resembles a well known in physics renormalization procedure while deep neural nets were recognized to effectively approximate quantum wavefunctions. Some powerful physics methods could be applied in DL to extract the relevant features of the data and optimize the dimensionality of the networks in a controllable way. I will review recent advances which make deep learning less a “black box” and provide an example of using toolbox from quantum physics.