Strategic Model Compression for Sustainable and Societal AI: When to Prune, Quantize, and Optimize

The rapid growth of deep neural networks (DNNs) has delivered exceptional progress across numerous AI applications but at a significant computational, environmental, and economic cost. Compression techniques such as pruning and quantization offer practical solutions to reduce model size and energy consumption, but the decision of which technique to apply depends critically on the model architecture, deployment context, and societal importance of the application. In this paper, we propose a unified roadmap for compression, offering guidelines for when to prune, when to quantize, and when multi-objective optimization is necessary. Our approach leverages automatic evolutionary metaheuristics, which outperform traditional heuristic methods by adaptively optimizing compression configurations based on actual model performance and robustness. We further discuss sustainability concerns regarding the optimization process itself and argue that the one-time resource investment is outweighed by the significant energy savings during repeated inference. This framework is supported by several of our recent contributions, including layer and filter pruning using binary particle swarm optimization (BPSO), mixed-precision quantization via bare-bones PSO, and multi-objective pruning for audio-visual speech enhancement (AVSE). Finally, we discuss how sustainable compression supports societal goals such as healthcare, accessibility, and global AI democratization.