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..
Correlated Low-Rank Adaptation for ConvNets
Authors: Wu Ran, Weijia Zhang, ShuYang Pang, Zhu, Jinfan Liu, JingSheng Liu, Xin Cao, Qiang Li, Yichao Yan, Chao Ma
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments conducted across various mainstream vision tasks, including image classification, semantic segmentation, and object detection, illustrate that Co Lo RA significantly advances the state-of-the-art PEFT approaches. Notably, our Co Lo RA achieves superior performance with only 5% of trainable parameters, surpassing full fine-tuning in the image classification task on the VTAB-1k dataset using Conv Ne Xt-S. Code is available at https://github.com/VISION-SJTU/Co Lo RA. Section 4 is titled "Experiments" and contains detailed setup and evaluation results on multiple benchmarks. |
| Researcher Affiliation | Collaboration | 1 Mo E Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University 2 CISDI Information Technology CO., LTD. |
| Pseudocode | No | The paper describes methods using mathematical equations and descriptive text, but does not include any explicitly labeled "Pseudocode" or "Algorithm" blocks. For example, Section 3.3 "Correlated Low-Rank Adaptation" and Section 3.4 "Integrating Co Lo RA into Conv Nets" describe the methodology without pseudocode. |
| Open Source Code | Yes | Code is available at https://github.com/VISION-SJTU/Co Lo RA. |
| Open Datasets | Yes | Comprehensive experiments conducted across various mainstream vision tasks, including image classification, semantic segmentation, and object detection... on the VTAB-1k dataset using Conv Ne Xt-S. Our evaluation encompasses a range of vision tasks, including classification on the large-scale visual adaptation benchmark VTAB-1k [50], object detection on the MS-COCO [53] dataset, and segmentation on the ADE20K [20] benchmark. |
| Dataset Splits | Yes | For VTAB-1k classification tasks... Each VTAB-1k sub-task is fine-tuned with a batch size of 32 for 100 epochs. For ADE20K segmentation, we employ both a primary decoder head and an auxiliary head... standard 160k iteration schedule. We conduct object detection experiments using the Faster R-CNN framework... adhering to a standard 1x schedule over 12 epochs. The NeurIPS checklist further states: "Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results?" Answer: [Yes] Justification: "Please see Sections 4, G and H for experimental settings.". |
| Hardware Specification | Yes | All experiments are conducted on NVIDIA A800 GPUs using the Py Torch [54] framework, supported by APEX [55] for mixed precision training. |
| Software Dependencies | No | All experiments are conducted on NVIDIA A800 GPUs using the Py Torch [54] framework, supported by APEX [55] for mixed precision training. While PyTorch and APEX are mentioned, specific version numbers are not provided in the text. |
| Experiment Setup | Yes | We set the compression factor to γ = 16 and ensure that the ranks of the shared and layer-specific matrices are equal, i.e., rs = rl. Tables 10 and 11 present the complete specifications of baseline hyperparameters and training configurations. Table 10 (Res Net-50): Optimizer Adam W, Learning rate 1e-4, Weight decay 0.05, Layer weight decay, Linear warmup 1500 steps with ratio 1e-6, Mixed precision training APEX. Table 11 (Conv Ne Xt): Optimizer Adam W, Learning rate 1e-4, Weight decay 0.05, Stage-wise decay, Linear warmup 1500 steps with ratio 1e-6, Mixed precision training APEX. For VTAB-1k... batch size of 32 for 100 epochs. For ADE20K... standard 160k iteration schedule... batch size of 16. For MS-COCO... learning rate of 2e-4, standard 500-iteration warmup period, batch size of 16 distributed across two NVIDIA A800 GPUs, standard 1x training schedule over 12 epochs. |