Robust Pruning

Robust pruning aims to reduce the size and computational cost of neural networks while preserving accuracy and robustness against various forms of noise or perturbations, such as weight changes or adversarial attacks. Current research focuses on developing pruning algorithms for diverse architectures, including spiking neural networks and large language models, employing techniques like cooperative co-evolution, geometric median matching, and Moreau envelope-based methods to achieve this goal. This work is significant because efficient and robust pruning is crucial for deploying large models on resource-constrained devices and improving the reliability and fairness of machine learning systems in safety-critical applications. Furthermore, research is exploring how to mitigate biases introduced by data pruning techniques.