Multiplication Circuit Architecture for Error- Tolerant CNN-Based Keywords Speech Recognition

Abstract

Approximate multiplication is an emerging circuit design technique for AI-based IoT devices which can reduce the energy consumption. A precision reconfigurable tensor multiplication unit (TMU) is proposed to explore the advantages for energy efficient approximate computing for error-tolerant convolutional neural network (CNN) based keywords speech recognition, which is a widely used human-machine interaction system. A positive-negative encoding method and a reorganization structure of partial products are proposed for the tensor multiplication. The partial products are further optimized by a hierarchical addition tree structure with fine-grained precision reconfigurable approximate computing. The imprecision parts of the addition tree are optimized with a low supply voltage. The effects of circuit configuration variations on accuracy and power consumption have been evaluated under 22nm process technology. The proposed approximate designs are applied to a CNN-based keywords speech recognition system. With the proposed approach, the energy efficiency of the CNN accelerator can be greatly improved with a power reduction of up to 44.1% at a loss in accuracy of less than 2%.