Pruning neural networks without any data by iteratively conserving synaptic flow

Authors: Hidenori Tanaka, Daniel Kunin, Daniel L. Yamins, Surya Ganguli

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We first mathematically formulate and experimentally verify a conservation law that explains why existing gradient-based pruning algorithms at initialization suffer from layer-collapse, the premature pruning of an entire layer rendering a network untrainable. ... and demonstrate empirically2 that this algorithm achieves state-of-the-art pruning performance on 12 distinct combinations of models and datasets (Sec. 7).
Researcher Affiliation Collaboration Hidenori Tanaka Physics & Informatics Laboratories NTT Research, Inc. Department of Applied Physics Stanford University Daniel Kunin Institute for Computational and Mathematical Engineering Stanford University Daniel L. K. Yamins Department of Psychology Department of Computer Science Stanford University Surya Ganguli Department of Applied Physics Stanford University
Pseudocode Yes Algorithm 1: Iterative Synaptic Flow Pruning (Syn Flow). Input: network f(x; 0), compression ratio , iteration steps n 0: f(x; 0) ; .Set model to eval modea 1: µ = 1 ; .Initialize binary mask for k in [1, . . . , n] do 2: µ µ 0 ; .Mask parameters 1 ; .Evaluate Syn Flow objective @ µ µ ; .Compute Syn Flow score 5: (1 k/n) percentile of S ; .Find threshold 6: µ ( < S) ; .Update mask end 7: f(x; µ 0) ; .Return masked network
Open Source Code Yes All code is available at github.com/ganguli-lab/Synaptic-Flow.
Open Datasets Yes datasets (CIFAR-10/100 and Tiny Image Net)
Dataset Splits No No explicit dataset split percentages, sample counts, or detailed splitting methodology for training, validation, and testing were provided in the main text.
Hardware Specification No No specific hardware details such as GPU/CPU models or detailed compute environment specifications were provided.
Software Dependencies No No specific software dependencies with version numbers (e.g., library names with versions) were mentioned in the paper.
Experiment Setup No The paper states 'See Appendix 13 for more details and hyperparameters of the experiments' but these details are not present in the provided main text.