Events for 01/25/2018 from all calendars

Center for Approximation and Mathematical Data Analytics

Time: 3:00PM - 4:00PM

Location: BLOC 220

Speaker: David Rolnick, MIT

Title: The power of deeper neural networks for expressing natural functions

Abstract: The field of deep learning has gained great popularity in recent years due to the ability of artificial neural networks to fit unknown functions of input data. Neural networks are compositions of linear and nonlinear functions, with the "depth" of the network measuring the number of successive compositions. In practice, deeper networks have proven to be more powerful at expressing natural functions. In this talk, we will show (1) that arbitrarily good approximations to polynomials are possible with a fixed network size and (2) this fixed size decreases exponentially as the depth increases. These results justify behavior observed in practice and suggest rules of thumb for picking the depth of a network.