Power-efficient Approximate Multipliers for Classification Tasks in Neural Networks

Multiplication is a key operation in neural networks. To overcome the power efficiency challenges of designing dedicated hardware for neural networks, designers can explore approximate multipliers to reduce area and power while maintaining tolerable accuracy. In this work, we evaluate the power and accuracy trade-offs of adopting two approximate multiplier structures, AxMultV1 and AxMultV2, for image classification in neural networks. In these multipliers, we explore seven approximate 4:2 compressors from the literature and compare with our proposed MAX4:2CV1 compressor. The adoption of our proposed compressor in multipliers provides power savings up to 56%, a delay reduction of 45.5%, and reduction in transistor count up to 48% compared to an exact multiplier. The multipliers based on the MAX4:2CV1 compressor can be considered suitable for classification tasks in neural networks, achieving 95.54% accuracy on the MNIST using a Multilayer Perceptron and up to 81.27% accuracy on the SVHN dataset with the LeNet-5 architecture, comparable to the accuracy of an exact multiplier.