Farkas layers: don't shift the data, fix the geometry
Machine Learning
2019-10-08 v1 Computer Vision and Pattern Recognition
Abstract
Successfully training deep neural networks often requires either batch normalization, appropriate weight initialization, both of which come with their own challenges. We propose an alternative, geometrically motivated method for training. Using elementary results from linear programming, we introduce Farkas layers: a method that ensures at least one neuron is active at a given layer. Focusing on residual networks with ReLU activation, we empirically demonstrate a significant improvement in training capacity in the absence of batch normalization or methods of initialization across a broad range of network sizes on benchmark datasets.
Cite
@article{arxiv.1910.02840,
title = {Farkas layers: don't shift the data, fix the geometry},
author = {Aram-Alexandre Pooladian and Chris Finlay and Adam M Oberman},
journal= {arXiv preprint arXiv:1910.02840},
year = {2019}
}