Farkas layers: don't shift the data, fix the geometry

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.

References in corpus (7)

• Instance Normalization: The Missing Ingredient for Fast Stylization
• Improved Regularization of Convolutional Neural Networks with Cutout
• Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour
• Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs)
• All you need is a good init
• Fixup Initialization: Residual Learning Without Normalization
• Batch Normalization is a Cause of Adversarial Vulnerability