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..
Unveiling m-Sharpness Through the Structure of Stochastic Gradient Noise
Authors: Haocheng Luo, Mehrtash Harandi, Dinh Phung, Trung Le
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments We evaluate three optimization methods: SGD, SAM2, and RW-SAM on CIFAR-10 and CIFAR-100 (Krizhevsky et al., 2009), training three models from scratch: Res Net-18, Res Net-50 (He et al., 2016), and Wide Res Net-28-10 (Zagoruyko and Komodakis, 2016). We use a batch size of 128 and a cosine learning rate schedule with an initial learning rate of 0.1. SAM and RW-SAM are trained for 200 epochs, while SGD is trained for 400 epochs. We apply a momentum of 0.9 and a weight decay of 5e 4, along with standard data augmentation techniques, including horizontal flipping, padding by four pixels, and random cropping. ... We repeated three independent experiments and reported the mean and standard deviation of the test accuracy in Table 1 and Table 2a. |
| Researcher Affiliation | Academia | Haocheng Luo Mehrtash Harandi Dinh Phung Trung Le Monash University, Australia EMAIL |
| Pseudocode | Yes | Algorithm 1 Reweighted SAM 1: while not converged do 2: Forward pass to obtain fγk(xk) 3: for q = 1, . . . , Q do 4: Estimate per-sample gradient norm using Eq. (21) 5: end for 6: Normalize estimated per-sample gradient norm 7: Compute weight p using Eq. (20) 8: Compute perturbation ϵk using Eq. (18) 9: Compute perturbed gradient gk = fγk(xk + ρϵk) 10: Update model parameters: xk+1 = xk ηgk 11: end while |
| Open Source Code | No | All experiments use publicly available datasets, but our code is not yet available. We plan to release it upon acceptance. |
| Open Datasets | Yes | We evaluate three optimization methods: SGD, SAM2, and RW-SAM on CIFAR-10 and CIFAR-100 (Krizhevsky et al., 2009), training three models from scratch: Res Net-18, Res Net-50 (He et al., 2016), and Wide Res Net-28-10 (Zagoruyko and Komodakis, 2016). ... For large-scale experiments, we train a Res Net-50 on Image Net-1K (Deng et al., 2009) for 90 epochs with an initial learning rate of 0.05. ... We also finetune a pretrained Distil BERT model (Sanh et al., 2019) on the GLUE benchmark (Wang et al., 2018). |
| Dataset Splits | Yes | We evaluate three optimization methods: SGD, SAM2, and RW-SAM on CIFAR-10 and CIFAR-100 (Krizhevsky et al., 2009), training three models from scratch: Res Net-18, Res Net-50 (He et al., 2016), and Wide Res Net-28-10 (Zagoruyko and Komodakis, 2016). ... For the additional hyperparameter λ in RW-SAM, we performed a grid search over {0.25, 0.5, 1.0, 2.0} on a validation set and found that 0.5 consistently yielded strong performance across experiments. |
| Hardware Specification | Yes | All experiments were run on NVIDIA RTX 4090 GPUs. |
| Software Dependencies | No | The paper does not explicitly state software names with specific version numbers (e.g., "Python 3.8, PyTorch 1.9, and CUDA 11.1") that would be needed to replicate the experiment. |
| Experiment Setup | Yes | We use a batch size of 128 and a cosine learning rate schedule with an initial learning rate of 0.1. SAM and RW-SAM are trained for 200 epochs, while SGD is trained for 400 epochs. We apply a momentum of 0.9 and a weight decay of 5e 4, along with standard data augmentation techniques, including horizontal flipping, padding by four pixels, and random cropping. For SAM and RW-SAM, we set ρ = 0.05 for CIFAR-10 and ρ = 0.1 for CIFAR-100. |