Harnessing Neural Unit Dynamics for Effective and Scalable Class-Incremental Learning

Authors: Depeng Li, Tianqi Wang, Junwei Chen, Wei Dai, Zhigang Zeng

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that our method achieves strong CIL performance in rehearsal-free and minimalexpansion settings with different backbones. and 5. Experiment
Researcher Affiliation Academia Depeng Li 1 Tianqi Wang 1 Junwei Chen 1 Wei Dai 2 Zhigang Zeng 1 1School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, China 2School of Information and Control Engineering, China University of Mining and Technology, Xuzhou, China. Correspondence to: Zhigang Zeng <zgzeng@hust.edu.cn>
Pseudocode Yes We summarize its CIL procedure in Algorithm 1 in Appendix B.
Open Source Code No The paper does not provide a direct link to the source code or an explicit statement of code release for their own work. It only refers to 'original codebases' for baselines.
Open Datasets Yes We experiment on multiple datasets commonly used for CIL. Small Scale: Both MNIST (Le Cun et al., 1998) and Fashion MNIST (Xiao et al., 2017) are respectively split into 5 disjoint tasks with 2 classes per task. Medium Scale: CIFAR-100 (Krizhevsky et al., 2009) is divided into 10 (25) tasks with each task containing 10 (4) disjoint classes. Large Scale: Image Net-R (Hendrycks et al., 2021)
Dataset Splits Yes When conducting experiments with different datasets, we keep about 10% of the training data from each task for validation.
Hardware Specification Yes We run experiments on extensive datasets adapted for CIL under different widely used backbones, which are implemented in Py Torch with NVIDIA RTX 3080-Ti GPUs.
Software Dependencies No The paper mentions 'Py Torch' but does not provide a specific version number. No other software dependencies are listed with their version numbers.
Experiment Setup Yes We use the SGD optimizer with an initial learning rate (0.001 for MNIST, Fashion MNIST; 0.01 for the remaining)... In our method... we use Lmax(t) = 200 and R(t) = 99%; for CIFAR-100, we use Lmax(t) = 1000 and R(t) = 90% (CIFAR-100/10), and Lmax(t) = 500 and R(t) = 80% (CIFAR-100/25). Similarly, we empirically set the step size l = 10 for node expansion each time and the maximum times of random generation Tmax = 50... r(t) = 0.9 and µL(t) = 1 r(t) / (L+1).