Structured Sparse Ternary Weight Coding of Deep Neural Networks for Efficient Hardware Implementations

TL;DR

This paper introduces a structured sparse ternary weight coding method for deep neural networks that enables multiplication-free inference and significantly reduces weight storage, improving efficiency for hardware implementations.

Contribution

The paper proposes a novel structured sparse ternary coding scheme that allows for efficient, multiplication-free DNN inference with high compression rates.

Findings

Achieves up to 32x weight compression compared to floating-point networks.

Enables multiplication-free DNN inference through structured sparse coding.

Maintains acceptable performance with normalization and pruning techniques.

Abstract

Deep neural networks (DNNs) usually demand a large amount of operations for real-time inference. Especially, fully-connected layers contain a large number of weights, thus they usually need many off-chip memory accesses for inference. We propose a weight compression method for deep neural networks, which allows values of +1 or -1 only at predetermined positions of the weights so that decoding using a table can be conducted easily. For example, the structured sparse (8,2) coding allows at most two non-zero values among eight weights. This method not only enables multiplication-free DNN implementations but also compresses the weight storage by up to x32 compared to floating-point networks. Weight distribution normalization and gradual pruning techniques are applied to mitigate the performance degradation. The experiments are conducted with fully-connected deep neural networks and convolutional neural networks.