Approximate Ternary Matrix Multiplication for Image Processing and Neural Networks

This paper presents a Carbon Nanotube FET-based ternary matrix multiplication using systolic array architecture for applications towards ternary neural networks and image processing applications. A ternary multiplier is proposed using ternary 4:2 compressors, which reduces the hardware by 18% in terms of the number of CNTFETs over the best existing designs. The compressors are approximated for energy-efficient applications. The approximate 4:2 compressor saves 19% of the energy over the proposed accurate design. A 6 x 6 multiplier designed using proposed approximate compressors saves 20% of the energy over the recent approximate multiplier. The proposed approximate multiplier has a better MRED and results in better image quality when deployed in image multiplication and image smoothing applications. A neural network is implemented using the proposed matrix multiplication for image classification results in an accuracy of 96% to 98% for various errors in multipliers.