FairCompressAgent Uses an AI Planner for Fair Model Compression on FPGAs

Summary

The paper introduces FairCompressAgent (FCA), an agentic workflow for choosing model-compression methods and configurations while balancing accuracy, fairness, and deployment cost for FPGA use. FCA provides a common interface for fairness-aware pruning, incremental quantization, and sparse low-rank factorization. A language-model planner uses model profiles and measured results to propose configurations, while an execution layer carries out compression, fine-tuning, evaluation, and constraint-based selection. The framework can incorporate changing requirements and reports any remaining violation when a request cannot be met. In experiments on Fitzpatrick-17k with VGG-11, the authors compared four search methods across 40 measured configurations. For an accuracy-constrained request, FCA selected a model that reduced inference tensor storage by 59.54%, increased validation average precision from 0.5141 to 0.5233, and reduced equalized opportunity from 0.2251 to 0.2168. It reached the same final selection as one-shot planning while evaluating 7.33 candidates on average, compared with 12 under the respective stopping policies. Repeated fine-tuning, held-out testing, and online requirement updates were used to examine stability and interactive operation. The results support measured feedback and explicit constraints as a way to refine fairness-aware compression choices.