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..
Distribution-Aware Tensor Decomposition for Compression of Convolutional Neural Networks
Authors: Alper KALLE, Théo Rudkiewicz, Mohamed Ouerfelli, Mohamed Tamaazousti
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on several CNN architectures (Res Net-18/50, and Goog Le Net) and datasets (Image Net, FGVC-Aircraft, Cifar10, and Cifar100) confirm the advantages of the proposed method. We validate our proposed methods through comprehensive experiments on several prominent CNN architectures, including Res Net-18, Res Net-50, and Goog Le Net, across diverse benchmark datasets such as Image Net, FGVC-Aircraft, Cifar10, and Cifar100. |
| Researcher Affiliation | Academia | Alper Kalle EMAIL Theo Rudkiewicz EMAIL Mohamed-Oumar Ouerfelli EMAIL Mohamed Tamaazousti EMAIL *These authors contributed equally Université Paris-Saclay, CEA, List, F-91120, Palaiseau, France TAU team, LISN, Université Paris-Saclay, CNRS, Inria, 91405, Orsay, France |
| Pseudocode | Yes | Algorithm 1 CP-ALS-Sigma 1: function [U (T ), U (S), U (W ), U (H), m] = CP-ALS-SIGMA ... Algorithm 2 Tucker2-ALS-Sigma 1: function [U (T ), U (S), G, m] = TUCKER2-ALS-SIGMA ... |
| Open Source Code | Yes | We provide in the supplementary material the code for the algorithm 2 Tucker2ALS-Sigma as well as the code for the computation of the sigma matrix. These are the two essential blocks that correspond to our contribution, as a complete compression module would be too complex to include. These two blocks will be made publicly available on Git Hub after paper acceptance: https://github.com/alpkll/Parafac_Sigma |
| Open Datasets | Yes | Experiments on several CNN architectures (Res Net-18/50, and Goog Le Net) and datasets (Image Net, FGVC-Aircraft, Cifar10, and Cifar100) confirm the advantages of the proposed method. |
| Dataset Splits | No | Specifically, we consider the case where only 50,000 images from the Image Net training set are available to estimate the matrix Σ and to fine-tune the model. Fine-tuning was conducted on the CIFAR-10 training dataset, employing the Adam optimizer and testing various learning rates from 10 3 to 10 7, selecting the best-performing learning rate. Fine-tuning was performed on the CIFAR-100 training dataset, employing the Adam optimizer and testing various learning rates from 10 3 to 10 7, selecting the best-performing learning rate. |
| Hardware Specification | Yes | Hardware: All CP-ALS and Tucker2-ALS experiments were executed exclusively on CPU, using a machine with an x86_64 architecture and 1.5TB of RAM. In contrast, all CP-ALS-Sigma and Tucker2-ALS-Sigma experiments were performed on systems equipped with either an NVIDIA A100 GPU (80GB) or an NVIDIA H100 GPU (100GB), alongside the same CPU and memory configuration. |
| Software Dependencies | Yes | Software Environment: We used Py Torch 2.6.0 and CUDA 12.4. Experiments were run on a Linux system with Python 3.10.14. |
| Experiment Setup | Yes | For the network compressed using the standard ALS algorithm under the Frobenius norm, we fine-tune it with the Adam optimizer, selecting the optimal learning rate from the range 10 5 to 10 10. Fine-tuning was conducted on the CIFAR-10 training dataset, employing the Adam optimizer and testing various learning rates from 10 3 to 10 7, selecting the best-performing learning rate. Fine-tuning phases were run for 30 epochs, taking approximately 1 6 hours per model. |