Researchers presented the FairCompressAgent framework for compressing models under FPGA

Researchers presented FairCompressAgent — an agent-based framework designed for automated compression of neural networks considering fairness requirements and hardware constraints of FPGA, the system combines pruning, incremental quantization, and sparse low-rank factorization into a single workflow. The solution is based on a language model-planner that analyzes neural network profiles and performance metrics to select the optimal compression configuration. The executive layer of the framework takes on the tasks of fine-tuning, quality evaluation, and model selection based on given constraints, in case of inability to fulfill user requirements, the system provides a report on violations. Testing was conducted on the VGG-11 architecture using the Fitzpatrick-17k dataset, with given accuracy constraints, FairCompressAgent reduced the storage volume of inference tensors by 59.54%, at the same time, the average precision grew from 0.5141 to 0.5233, and the equality of opportunity metric (EOpp) decreased from 0.2251 to 0.2168. The framework demonstrated high efficiency of search: to achieve the final configuration, the system required an average of 7.33 candidate evaluations versus 12 when using standard planning methods, the tool supports interactive updating of requirements and re-fine-tuning, which ensures the stability of the optimization process. The development is aimed at simplifying the deployment of AI solutions on specialized equipment, where the balance between computational efficiency and ethical model metrics is critically important, the source code and methodology description are available under the Non-exclusive distribution license.