Continual Learning with Filter Atom Swapping

Authors: Zichen Miao, Ze Wang, Wei Chen, Qiang Qiu

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Being validated on multiple benchmark datasets with different convolutional network structures, the proposed method outperforms the state-of-the-art methods in both accuracy and scalability.
Researcher Affiliation Academia Zichen Miao, Ze Wang, Wei Chen, Qiang Qiu Department of ECE Purdue University {miaoz, wang5026, chen2732, qqiu}@purdue.edu
Pseudocode Yes We provide the algorithm of the proposed method in Alg. 1.
Open Source Code No The paper does not provide explicit statements or links for open-source code availability.
Open Datasets Yes We validate our method with the Class-Incremental (CI) setting with CIFAR100 and Image Net-Subset , which contains 100 classes selected from Image Net (with random seed 1993).
Dataset Splits Yes Details of each dataset are provided in the Appendix C. Table A: Statistics of 10-Split CIFAR100. # Training samples/task 4500 # Validation samples/task 500 # Test samples/task 100
Hardware Specification Yes All methods are tested on a single RTX 2080ti GPU under the class-incremental setting.
Software Dependencies No The paper mentions software like SGD, ResNet, and AlexNet-like architectures but does not specify their version numbers.
Experiment Setup Yes For CIFAR100, we choose SGD with batch-size of 128, learning rate of 0.01, momentum of 0.9 and weight decay 1e-3. The model is trained for 250 epochs, with learning rate drop by 0.1 at the 100-th and 200-th epoch.