35 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Sample-complexity of Estimating Convolutional and Recurrent Neural Networks
Carnegie Mellon University
Recording
Abstract
Aarti Singh studies how many examples are needed to estimate linear recurrent and convolutional networks. The analysis compares structured architectures with fully connected networks, finds sample complexity proportional to their intrinsic dimension, and gives lower bounds that are tight up to logarithmic factors.