Batch Kalman Normalization: Towards Training Deep Neural Networks with Micro-Batches

TL;DR

This paper introduces Batch Kalman Normalization (BKN), a new method that improves deep neural network training with micro-batches by considering all layers collectively, leading to more stable training and better accuracy with smaller batch sizes.

Contribution

BKN is a novel normalization technique that models all network layers as a system, enhancing stability and performance especially with micro-batches, outperforming traditional Batch Normalization.

Findings

BKN achieves higher accuracy on ImageNet with smaller batch sizes.

Training with BKN converges faster than with BN.

BKN performs well with batch sizes 64 times smaller on CIFAR and 8 times smaller on ImageNet.

Abstract

As an indispensable component, Batch Normalization (BN) has successfully improved the training of deep neural networks (DNNs) with mini-batches, by normalizing the distribution of the internal representation for each hidden layer. However, the effectiveness of BN would diminish with scenario of micro-batch (e.g., less than 10 samples in a mini-batch), since the estimated statistics in a mini-batch are not reliable with insufficient samples. In this paper, we present a novel normalization method, called Batch Kalman Normalization (BKN), for improving and accelerating the training of DNNs, particularly under the context of micro-batches. Specifically, unlike the existing solutions treating each hidden layer as an isolated system, BKN treats all the layers in a network as a whole system, and estimates the statistics of a certain layer by considering the distributions of all its preceding layers, mimicking the merits of Kalman Filtering. BKN has two appealing properties. First, it enables more stable training and faster convergence compared to previous works. Second, training DNNs using BKN performs substantially better than those using BN and its variants, especially when very small mini-batches are presented. On the image classification benchmark of ImageNet, using BKN powered networks we improve upon the best-published model-zoo results: reaching 74.0% top-1 val accuracy for InceptionV2. More importantly, using BKN achieves the comparable accuracy with extremely smaller batch size, such as 64 times smaller on CIFAR-10/100 and 8 times smaller on ImageNet.