Early Termination of the MSDF Computations towards Efficient Inference in Neural Networks

Multi-layer perceptrons (MLPs) are widely used to model complex nonlinear relationships across various applications, but achieving high accuracy while minimizing energy and resource consumption is crucial in resource-limited environments. Computations based on the most significant digit first (MSDF) process digits from left to right, allowing early access to the most significant digits during initial processing. This characteristic enables the early termination of computations once the desired output accuracy is achieved, thereby leading to efficient and cost-effective implementations. This paper explores early termination techniques for applying MSDF computations in the MLP. An ASIC-based MLP architecture is designed to perform serial computations on digits. The efficiency of the proposed MLP design during inference using the MNIST handwritten digit classification dataset is analyzed. Results show that the proposed design achieves a minimum delay and energy consumption reduction of at least 47% while maintaining high accuracy compared to state-of-the-art designs.