The accuracy-fairness-efficiency Trilemma in mobile image classification: a Pareto benchmark

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deploying deep learning on mobile devices faces challenges balancing accuracy, fairness, and efficiency.
• Existing methods lack a unified framework for these competing objectives.
• No prior work defined a Deployment-Feasible Zone (DFZ) for constrained optimization.

Purpose of the Study:

• To formalize the deployment of deep learning on mobile devices as a constrained multi-objective optimization problem.
• To benchmark various optimization strategies under identical, resource-constrained conditions.
• To define and identify the Deployment-Feasible Zone (DFZ) representing optimal trade-offs.

Main Methods:

• Benchmarking eleven deep learning optimization configurations on a 2,821-image dataset.
• Evaluating performance across 24 demographic subgroups under strict constraints (accuracy, fairness, memory, latency).
• Utilizing 3D-aware augmentation and Protected Fairness Pruning (PFP), alongside an Adaptive Trilemma Weight Scheduler (ATWS).

Main Results:

• The combination of 3D-aware augmentation and PFP (C2) emerged as the Pareto knee point, achieving high accuracy (0.906-0.962), fairness, 6.3 MB size, and 187 ms inference time.
• Fairness-constrained pruning consistently outperformed standard magnitude pruning at similar compression ratios.
• The ATWS strategy improved fairness metrics by 1.3 pp and 0.7 pp over fixed-weight training.

Conclusions:

• A novel framework and the concept of a Deployment-Feasible Zone (DFZ) are established for mobile deep learning deployment.
• The optimal strategy (C2) demonstrates a viable approach to satisfy accuracy, fairness, and efficiency constraints simultaneously.
• Adaptive scheduling methods offer significant improvements in fairness over static approaches.