Simulation-Based Hardware Exploration for Shapley Value Calculations.

Area of Science:

• Computer Science
• Machine Learning
• Hardware Acceleration

Background:

• Shapley values are crucial for explaining machine learning models, enhancing trust in AI systems, particularly in healthcare.
• Exact computation of Shapley values is computationally intensive, necessitating efficient approximation methods.
• Existing approximations are often implemented in high-level languages like Python, limiting hardware-specific optimization.

Purpose of the Study:

• To investigate efficient Shapley value computation methods tailored for RISC-V architectures and specialized hardware.
• To develop and evaluate a fixed-point implementation of Shapley values, aligning with quantized neural network trends.
• To benchmark the performance of various implementations against a simulated hardware accelerator for quantized Shapley values.

Main Methods:

• Exploration of efficient Shapley value calculation algorithms on RISC-V platforms.
• Development of a fixed-point arithmetic implementation suitable for quantized neural networks.
• Benchmarking using a virtual prototype to compare different RISC-V ISA extensions.
• Performance comparison with a simulated hardware implementation for quantized Shapley values.

Main Results:

• Demonstrated feasibility of efficient Shapley value computation on RISC-V systems.
• Quantized Shapley value implementation shows promise for hardware acceleration.
• Performance gains identified through specific RISC-V ISA extensions and hardware simulation.

Conclusions:

• Efficient Shapley value computation is achievable on RISC-V architectures and specialized hardware.
• Fixed-point implementations offer a viable path for accelerating Shapley value calculations in quantized neural networks.
• This work lays the foundation for hardware-accelerated trustworthy AI in resource-constrained environments.