Multirate FIR CNN - Optimizing Deep Learning Implementation

This paper proposes the multirate finite impulse-response (FIR) Convolutional Neural Networks (mCNN) algorithms for an efficient implementation of the 2-Dimensional (2-D) CNN circuits implementation. With the rapid advancement in computational power, Deep Learning (DL) leveraging CNN has expanded the scope of Artificial Intelligence (AI) applications. In traditional CNN architectures, convolution and pooling operations have been enhanced through multirate algorithms. Recent research has introduced polyphase structures to optimize these operations, leading to more efficient multirate filtering implementations compared to conventional non-multirate approaches. We generalize this approach to 2-D CNNs through time-varying coefficients, achieving M-fold line-buffer reduction (where M is the pooling factor) and 1/M-rate MAC operations. A first-stage CNN implementation demonstrates these efficiency gains, validated via MATLAB's digit recognition tool, showing measurable improvements in memory and computational requirements.