Neural Collapse Inspired Feature Alignment for Out-of-Distribution Generalization
Authors: Zhikang Chen, Min Zhang, Sen Cui, Haoxuan Li, Gang Niu, Mingming Gong, Changshui Zhang, Kun Zhang
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on four datasets, and the results demonstrate that our method significantly improves out-of-distribution performance. |
| Researcher Affiliation | Academia | Zhikang Chen1 Min Zhang2 Sen Cui5 Haoxuan Li 3 Gang Niu4 Mingming Gong6,8 Changshui Zhang5 Kun Zhang7,8 1 Tsinghua University 2 East China Normal University 3 Peking University 4 RIKEN 5 Institute for Artificial Intelligence, Tsinghua University (THUAI) Beijing National Research Center for Information Science and Technology (BNRist) Department of Automation, Tsinghua University 6 The University of Melbourne 7 Carnegie Mellon University 8 Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) |
| Pseudocode | Yes | Summarily, the overall training process is presented in Algorithm 1. |
| Open Source Code | Yes | Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: We have strictly adhered to the anonymity guidelines when uploading the code. |
| Open Datasets | Yes | Following the work [Gulrajani and Lopez-Paz, 2020], we evaluate our method with baselines on benchmark datasets, using four datasets, namely Colored MNIST, Colored COCO, COCOPlaces and NICO. |
| Dataset Splits | Yes | For each dataset, we partition it into two subsets, d1 and d2, based on the environment, with the ratio of sample quantities between d1 and d2 being 9:1. Subset d1 from the training environment is used for model training, while d2 is employed for testing the models within the training environment. In the testing environment, d1 is utilized to assess the out-of-distribution (OOD) performance of the models, while d2 is utilized according to Domain Bed standards for selecting the best model. |
| Hardware Specification | Yes | In our experiments, we conduct all methods on a local Linux server that has two physical CPU chips (Intel(R) Xeon(R) CPU E5-2640 v4 @ 2.40GHz) and 32 logical kernels. All methods are implemented using Pytorch framework and all models are trained on Ge Force RTX 2080 Ti GPUs. |
| Software Dependencies | No | The paper states 'All methods are implemented using Pytorch framework' but does not specify a version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | Architecture. On Color MNIST, training is conducted using a 4-layer convolutional neural network. For the Colored COCO and COCOPlaces datasets, we adhere to the setup outlined in Ahmed et al. [2020], Gulrajani and Lopez-Paz [2020], employing Res Net8 for training. On the NICO dataset, training is performed using Res Net18. |