Approximate Logarithmic Multiplier For Convolutional Neural Network Inference With Computational Reuse

The design of efficient hardware for Convolutional Neural Networks (CNNs) has always been a challenging engineering task. This is due to the myriad of multiplication operations performed during CNN training that leads to computational complexities. This research paper proposes a new approach to address this issue by employing logarithmic number system (LNS) to replace the multiplication operation with addition. It makes use of the Mitchell's Algorithm to perform the task of multiplication coupled with computational reuse techniques to perform MAC operations more efficiently. This shows a power improvement of 11.4% from the synthesis results of the proposed design. The improvement comes with negligibly small loss in inference accuracy, which makes our approach suitable for devices with constrained energy and storage capacities. More importantly, computational reuse techniques are introduced in this work to avoid repetitive data movement and computation of redundant CNN parameters during MAC operations. This significantly reduces the associated computation and communication costs producing an improved power consumption and area efficiency.