2024/05/24 by Nolan Dey, Dey, Nolan, Shane Bergsma +3 · 1 voice · 2 citations
Computer Science · Physics and Astronomy · #Advanced Vision and Imaging #FOS: Computer and information sciences #Machine Learning (cs.LG) #Model Reduction and Neural Networks #Speech and Audio Processing #cs.LG
paper · pdf · doi:10.48550/arxiv.2405.15743
openalex publication_date 2024/05/24 · arxiv published 2024/05/24 · openalex created_date 2024/05/28 · arxiv updated 2026/02/03 · openalex updated_date 2026/07/28
Several challenges make it difficult for sparse neural networks to compete with dense models. First, setting a large fraction of weights to zero impairs forward and gradient signal propagation. Second, sparse studies often need to test multiple sparsity levels, while also introducing new hyperparameters (HPs), leading to prohibitive tuning costs. Indeed, the standard practice is to re-use the learning HPs originally crafted for dense models. Unfortunately, we show sparse and dense networks do not share the same optimal HPs. Without stable dynamics and effective training recipes, it is costly to test sparsity at scale, which is key to surpassing dense networks and making the business case for sparsity acceleration in hardware. A holistic approach is needed to tackle these challenges and we propose SμPar as one such approach. For random unstructured static sparsity, SμPar ensures activations, gradients, and weight updates all scale independently of sparsity level. Further, by reparameterizing the HPs, SμPar enables the same HP values to be optimal as we vary both sparsity level and model width. HPs can be tuned on small dense networks and transferred to large sparse models, greatly reducing tuning costs. On large-scale language modeling, SμPar shows increasing improvements over standard parameterization as sparsity increases, leading up to 11.9% relative loss improvement at 99.2% sparsity. A minimal implementation of SμPar is available at https://github.com/EleutherAI/nanoGPT-mup/tree/supar.