Matus Telgarsky - A margin perspective on neural networks


This talk will survey the role played by margins in optimization, generalization, and representation of neural networks. A specific highlight will be that applying gradient descent with reasonable step sizes and initialization to a network whose width is polynomial in the inverse margin but merely polylogarithmic in other problem parameters (e.g., training set size) suffices to achieve arbitrarily small test error.

Seminar Talk
60 Oxford St, Room 330. Cambridge, Massachusetts