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..
Error Feedback under $(L_0,L_1)$-Smoothness: Normalization and Momentum
Authors: SARIT KHIRIRAT, Abdurakhmon Sadiev, Artem Riabinin, Eduard Gorbunov, Peter Richtarik
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Numerical evaluation. We implemented ||EF21|| using the stepsize rules derived from our theory, and compared its performance against EF21. Both algorithms were evaluated on three learning tasks: minimizing nonconvex polynomial functions, solving logistic regression with a nonconvex regularizer, and training Res Net-20 on the CIFAR-10 dataset. |
| Researcher Affiliation | Academia | Sarit Khirirat, Abdurakhmon Sadiev, Artem Riabinin, and Peter Richtárik are with the Center of Excellence for Generative AI, King Abdullah University of Science and Technology (KAUST), Thuwal, Saudi Arabia. Eduard Gorbunov is with the Department of Statistics and Data Science, Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), Abu Dhabi, United Arab Emirates. |
| Pseudocode | Yes | Algorithm 1 Normalized Error Feedback (||EF21||) Algorithm 2 Normalized Error Feedback with Stochastic Gradients & Momentum (||EF21-SGDM||) |
| Open Source Code | Yes | Our source codes can be found in the link to error-feedback-generalized-smoothness-paper. |
| Open Datasets | Yes | training Res Net-20 on the CIFAR-10 dataset. LIBSVM [51]: Breast Cancer (n = 683, d = 10, and scaled to [ 1, 1]), and a1a (n = 1605, d = 123) |
| Dataset Splits | Yes | We used a top-k compressor over 50, 000 training images, with evaluation on 10, 000 test images. The dataset was evenly distributed among 5 clients, each using a mini-batch size of 128. |
| Hardware Specification | No | Our experiments on (1) minimizing simple functions and logistic functions, and on (2) Res Net20 training can be run on a machine with a single GPU. |
| Software Dependencies | No | We implemented EF21 and ||EF21|| on training the Res Net-20 model by using Py Torch. |
| Experiment Setup | Yes | For EF21, we selected the stepsize γk = 1/ L + L pβ/θ with L = q Pn i=1 ˆL2 i /n, θ = 1 1 α, and β = (1 α)/(1 1 α), given by Richtárik et al. [8, Theorem 1]. For ||EF21||, we chose γk = γ/ K + 1 with γ > 0 from Theorem 1, by setting γ0 = 1, K = 100 for the generated data and Breast Cancer, and K = 400 for a1a. Both algorithms were run for 100 epochs with a constant stepsize γ = 5. Here, one epoch refers to a full pass through the entire dataset processed by all clients. Each using a mini-batch size of 128. |