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..
Evolving and Regularizing Meta-Environment Learner for Fine-Grained Few-Shot Class-Incremental Learning
Authors: Li-Jun Zhao, Zhen-Duo Chen, Yongxin Wang, Xin Luo, Xin-Shun Xu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our method consistently and significantly outperforms existing approaches. |
| Researcher Affiliation | Academia | 1School of Software, Shandong University, China 2School of Computer and Artificial Intelligence, Shandong Jianzhu University, China |
| Pseudocode | No | The paper describes the methodology using mathematical equations and textual explanations in Section 3.3, but does not include any clearly labeled pseudocode or algorithm blocks formatted like code. |
| Open Source Code | Yes | The source code is provided in the supplemental materials. It includes instructions on data access and preparation, and scripts. |
| Open Datasets | Yes | Following the benchmark setting in [18], we evaluate our method on four fine-grained datasets: CUB200 [25], Stanford Dogs [8], Stanford Cars [30], and FGVCAircraft [15]. |
| Dataset Splits | Yes | For CUB200, 100 classes are used for base training, and the remaining 100 classes are split into 10 incremental sessions. For Stanford Dogs, 80 classes are used in the base session, and the remaining 40 classes are evenly divided into 8 incremental sessions. Each incremental class contains 5 samples, forming a 5-way 5-shot FSCIL setting. For Stanford Cars, 106 classes are used for base training, and the remaining 90 classes are split across 9 sessions, following a 10-way 5-shot setting. For FGVCAircraft, 60 classes are used for base training and 40 for incremental learning. A 5-way 5-shot setting is applied, resulting in 9 sessions in total. |
| Hardware Specification | Yes | Our method is conducted with Py Torch library on a single NVIDIA 3090, and SGD with momentum is used for optimization. |
| Software Dependencies | No | Our implementation is based on the code released by PFR [18] under the MIT license. Our method is conducted with Py Torch library on a single NVIDIA 3090, and SGD with momentum is used for optimization. |
| Experiment Setup | Yes | Following [18], we employ ResNet-12 as the backbone. In our experiments, Fhigh is the output of the final layer of the backbone, while Flow is the output of the penultimate layer. Following previous works [18], the temperature hyperparameter τ is set to 16. N is set as an integer 4 to ensure diversity in fine-grained features and kept small for efficiency. β and γ are set to 0.5 to balance the optimization strength among different components. For comparative methods whose results are not reported in [18], we reproduce their performance under the same experimental settings using the publicly available source code. Please see Appendix A.3 for details. |