Context We use Random Matrix Theory and regularization to prune deep neural networks, achieving state-of-the-art results on ImageNet with Vision Transformers. We also provide theoretical justification for our approach. Performance of our RMT-based pruning method on ViT-Base-16 model trained on …