Diagonal structured sparsity for efficient optical and GPU-based neural networks
Abhishek Tyagi, Arjun Iyer, Christopher Kanan, William H. Renninger, Yuhao Zhu · 2025
Diagonally-structured sparse neural networks suitable for optical implementation are shown to outperform other standard sparsity structures across several models and datasets, and are experimentally shown to have faster inference times than unstructured sparsity on GPUs.