Simulation-Based Hardware Exploration for Shapley Value Calculations

Shapley values are a popular method used to explain the output of machine learning models. They proved to be an important tool for providing trustworthy machine learning systems in healthcare. However, their exact computation is computationally expensive. Various approximation methods have been proposed, often implemented in high level programming languages such as Python. In this work, efficient computations of Shapley values on RISC-V based systems and specialized hardware are explored. Following recent trends of quantized neural networks, an implementation compatible with fixed point implementations is proposed. Using a virtual prototype, the different implementations of Shapley value calculations are benchmarked for different RISC-V ISA extensions and compared to a simulated hardware implementation for quantized Shapley values.