Neural Collapse To Multiple Centers For Imbalanced Data
Authors: Hongren Yan, Yuhua Qian, Furong Peng, Jiachen Luo, zheqing zhu, Feijiang Li
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 Experiments In this section, we (1) propose a cosine loss function for fixed classifier; (2) verify NCMC induced by the cosine loss through experiments; (3) show how f and θ influence the learning performance; (4) Compare long-tail classification performance to SETF method with fixed classifier and other classical methods with learnable classifier. |
| Researcher Affiliation | Academia | Hongren Yan, Yuhua Qian , Furong Peng, Jiachen Luo, Zheqing Zhu, Feijiang Li Shanxi University |
| Pseudocode | No | The paper presents mathematical formulations and descriptions of its methods but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | The code is modified from [26], and can be found in the supplementary material. |
| Open Datasets | Yes | We set long-tailed classification tasks on five datasets, CIFAR-10 [38], CIFAR-100 [38], SVHN [39], STL-10 [40], and large dataset Image Net [41] |
| Dataset Splits | No | The paper describes training details and uses 'test' performance metrics, but does not explicitly mention a 'validation' dataset split or its size. |
| Hardware Specification | Yes | We run experiments with backbone Res Net50 and Dense Net150 on the four datasets (CIFAR-10, CIFAR-100, SVHN and STL-10) by 1 A100 GPU, and run Res Net50 on Image Net by 2 A100 GPU with an extra linear layer that expand the dimension of the backbone feature to be larger than f K. |
| Software Dependencies | No | The paper mentions using specific backbone networks (Res Net50, Dense Net150) and that the code is modified from another source, but it does not specify software versions for libraries like PyTorch, TensorFlow, or Python. |
| Experiment Setup | Yes | We train the model on the four dataset for 200 epochs, with a step learning rate initialized to 0.1 decaying to 0.01 and 0.001 at epoch 160 and epoch 180, batch size of 128, a momentum of 0.9, and a weight decay of 2e 4. |