Probabilistic Group Mask Guided Discrete Optimization for Incremental Learning
Authors: Fengqiang Wan, Yang Yang
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments conducted on standard benchmarks confirm its superior effectiveness compared to existing IL approaches. The experimental setups are carefully aligned with those employed in recent works (Kang et al., 2022). The evaluation encompasses comprehensive performance comparisons, ablation studies examining module effectiveness, group size variations, and adaptability to different training paradigms, as well as an in-depth analysis of computational efficiency, parameter dependencies, and parameter distribution. |
| Researcher Affiliation | Academia | 1Nanjing University of Science and Technology. Correspondence to: Yang Yang <EMAIL>. |
| Pseudocode | No | No explicit pseudocode or algorithm block is found in the paper. The methodology is described through prose and mathematical formulations. |
| Open Source Code | Yes | The source code is available at: https://github. com/njustkmg/ICML25-PGM. |
| Open Datasets | Yes | We use three different popular datasets, including Split CIFAR-100 (Krizhevsky & Hinton, 2009), CIFAR-100 Superclass (Yoon et al., 2018b), Split Tiny Image Net (Krizhevsky et al., 2017). |
| Dataset Splits | No | The paper mentions using Split CIFAR-100, CIFAR-100 Superclass, and Split Tiny Image Net. It also states that 'The experimental setups are carefully aligned with those employed in recent works (Kang et al., 2022)'. While these datasets often come with standard splits for tasks, the paper does not explicitly provide specific percentages, sample counts, or direct details about how the training, validation, and test sets are partitioned within these datasets in its main text. |
| Hardware Specification | Yes | All experiments are implemented using Py Torch on a system equipped with four NVIDIA 4090 GPUs. |
| Software Dependencies | No | The paper states 'All experiments are implemented using Py Torch'. However, it does not specify the version number for Py Torch or any other software libraries or dependencies used. |
| Experiment Setup | Yes | Training employs the Adam optimizer with a momentum of 0.9, with each task trained for a fixed number of epochs to ensure convergence. Each task is trained for 50 epochs on CIFAR-100 and 40 epochs on Split Tiny Image Net. Additional hyperparameter settings are provided in the Appendix B. For instance, we use a modified Alex Net for Split CIFAR-100 and a customized Le Net for CIFAR-100 Superclass. |