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..

Dynamical Low-Rank Compression of Neural Networks with Robustness under Adversarial Attacks

Authors: Steffen Schotthöfer, Lexie Yang, Stefan Schnake

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments across standard architectures, datasets, and adversarial attacks show the regularized networks can achieve over 94% compression while recovering or improving adversarial accuracy relative to uncompressed baselines.
Researcher Affiliation Academia Computer Science and Mathematics Division, Geospatial Science and Human Security Division Oak Ridge National Laboratory Oak Ridge, TN 37831 USA EMAIL
Pseudocode Yes Algorithm 1: Single iteration of Robust DLRT.
Open Source Code Yes A reference implementation is provided at https://github.com/Sc Steffen/Robust DLRT.
Open Datasets Yes Extensive experiments across standard architectures, datasets, and adversarial attacks show the regularized networks can achieve over 94% compression while recovering or improving adversarial accuracy relative to uncompressed baselines. ... UCM dataset ... Cifar10 dataset ... Image Net1k dataset
Dataset Splits Yes We normalize the training and validation data with mean [0.485, 0.456, 0.406] and standard deviation [0.229, 0.224, 0.225] for the rgb image channels. ... For CIFAR10, we augment the training data set by a random horizontal flip of the image, followed by a normalization ... The test data set is only normalized. ... The Image Net dataset consists of 1000 classes and over 1.2 million RGB training images... The test set is only resized and center-cropped to 224 224, followed by normalization.
Hardware Specification Yes All experiments in this paper are computed using workstation GPUs. Each training run used a single GPU. Specifically, we have used 5 NVIDIA RTX A6000, 3 NVIDIA RTX 4090, and 8 NVIDIA A-100 80G.
Software Dependencies No The paper mentions 'pytorch implementation' and 'Optimizer Adam W' but does not specify version numbers for these or other software libraries/dependencies.
Experiment Setup Yes Table 8: Training hyperparameters for the UCM, Cifar10, and Image Net Benchmarks. The first set hyperparameters apply to both DLRT and baseline training, and we train DLRT with the same hyperparameters as the full-rank baseline models. The second set of hyper-parameters is specific to DLRT. The DLRT hyperparameters are selected by an initial parameter sweep. We choose the DLRT truncation tolerance relative to the Frobenius norm of b S, i.e. ϑ = τ b S F , as suggested in [38]. It lists Batch Size, Learning Rate, Number of Epochs, L2 regularization, Optimizer, DLRT rel. truncation tolerance, Coefficient Steps, Initial Rank.