Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Sign-In to the Lottery: Reparameterizing Sparse Training
Authors: Advait Gadhikar, Tom Jacobs, chao zhou, Rebekka Burkholz
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform experiments on standard vision benchmarks including CIFAR10, CIFAR100 (Krizhevsky et al., 2009), and Image Net (Deng et al., 2009) training a Res Net20, Res Net18, and Res Net50 model, respectively. All training details are provided in Table 4 in Appendix E. |
| Researcher Affiliation | Academia | Advait Gadhikar Tom Jacobs Chao Zhou Rebekka Burkholz CISPA Helmholtz Center for Information Security, Saarbrücken, Germany EMAIL |
| Pseudocode | Yes | Algorithm 1 Sign-In Input: objective L, scaling β, frequency p, epochs T, mask S, stop rescaling epoch T2 Initialize L(S m0 w0) such that m0 w0 = θ0 and m2 0 w2 0 = βI. for i = 1 to T do if i mod p == 0 and i < T2 then Rescale mi wi = θi such that m2 i w2 i = βI end if mi+1, wi+1 = Optimizer (L (S mi wi)) end for Return Model parameters θT = S m T w T |
| Open Source Code | Yes | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: We have submitted code. |
| Open Datasets | Yes | We perform experiments on standard vision benchmarks including CIFAR10, CIFAR100 (Krizhevsky et al., 2009), and Image Net (Deng et al., 2009) training a Res Net20, Res Net18, and Res Net50 model, respectively. |
| Dataset Splits | Yes | We perform experiments on standard vision benchmarks including CIFAR10, CIFAR100 (Krizhevsky et al., 2009), and Image Net (Deng et al., 2009) training a Res Net20, Res Net18, and Res Net50 model, respectively. |
| Hardware Specification | Yes | We gratefully acknowledge the Gauss Centre for Supercomputing e.V. for funding this project by providing computing time on the GCS Supercomputer JUWELS at Jülich Supercomputing Centre (JSC) Jülich Supercomputing Centre (2021). |
| Software Dependencies | No | Appendix M Code for Sign-In in Py Torch We provide an example of integrating Sign-In in any training setup in Py Torch by simply replacing the Conv (or Linear) layer with 1. import torch import torch.nn as nn |
| Experiment Setup | Yes | All training details are provided in Table 4 in Appendix E. Table 4: Training Details for all experiments presented in the paper. Dataset Model LR WD Epochs Batch Size Optim Schedule CIFAR10 Res Net20 0.2 1e 4 150 512 SGD Triangular Table 5: Training Details for the additional Sign-In parameters. Dataset Model β T2 Frobenius Decay Weight decay CIFAR10 Res Net20 1 75 1e 4 0 |