Aggregated Learning: A Deep Learning Framework Based on Information-Bottleneck Vector Quantization

TL;DR

This paper introduces AgrLearn, a novel deep learning framework based on vector information-bottleneck quantization, which improves learning efficiency and reduces data requirements in neural networks.

Contribution

The paper proposes AgrLearn, a new framework that extends scalar IB quantization to vector IB quantization, leading to enhanced neural network training performance.

Findings

AgrLearn improves accuracy on image recognition and text classification tasks.

It reduces training data requirements by up to 80%.

AgrLearn outperforms standard neural networks in experiments.

Abstract

Based on the notion of information bottleneck (IB), we formulate a quantization problem called "IB quantization". We show that IB quantization is equivalent to learning based on the IB principle. Under this equivalence, the standard neural network models can be viewed as scalar (single sample) IB quantizers. It is known, from conventional rate-distortion theory, that scalar quantizers are inferior to vector (multi-sample) quantizers. Such a deficiency then inspires us to develop a novel learning framework, AgrLearn, that corresponds to vector IB quantizers for learning with neural networks. Unlike standard networks, AgrLearn simultaneously optimizes against multiple data samples. We experimentally verify that AgrLearn can result in significant improvements when applied to several current deep learning architectures for image recognition and text classification. We also empirically show that AgrLearn can reduce up to 80% of the training samples needed for ResNet training.