Isolation and Impartial Aggregation: A Paradigm of Incremental Learning without Interference
Authors: Yabin Wang, Zhiheng Ma, Zhiwu Huang, Yaowei Wang, Zhou Su, Xiaopeng Hong
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the proposed method on four large datasets. Extensive results demonstrate the superiority of the proposed method in setting up new state-of-the-art overall performance. |
| Researcher Affiliation | Academia | 1 School of Cyber Science and Engineering, Xi an Jiaotong University 2 Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences 3 Singapore Management University 4 University of Southampton 5 Peng Cheng Laboratory 6 Harbin Institute of Technology |
| Pseudocode | Yes | Algorithm 1: Model Training and Algorithm 2: Inference are included in the paper. |
| Open Source Code | Yes | Code is available at https://github.com/iamwangyabin/ESN. |
| Open Datasets | Yes | We evaluate the proposed method on four large datasets. Split-Domain Net: We build the cross-domain class-incremental learning benchmark, Split-Domain Net, based on Domain Net (Peng et al. 2019). Split-CIFAR100 (Wang et al. 2022c) splits the origin CIFAR-100 (Krizhevsky and Hinton 2009). 5-Datasets (Ebrahimi et al. 2020) provides a benchmark for class incremental learning. CORe50 (Lomonaco and Maltoni 2017) is a large benchmark dataset for continual object recognition. |
| Dataset Splits | No | The paper does not explicitly state specific training/validation/test splits with percentages or counts for most datasets. For CORe50, it mentions a test set but no validation split: "Three domains (3, 7, and 10) are selected as test set, and the remaining 8 domains are used for incremental learning." |
| Hardware Specification | Yes | We implement our method in PyTorch with two NVIDIA RTX 3090 GPUs. |
| Software Dependencies | No | We implement our method in PyTorch. The version number for PyTorch or any other software dependency is not specified. |
| Experiment Setup | Yes | We use the SGD optimizer and the cosine annealing learning rate scheduler with a initial learning rate of 0.01 all benchmarks. We use 30 epochs for Split-CIFAR100 and Split-Domain Net, 10 epochs for 5-Datasets and CORe50. We set the batch size of 128 for all experiments. Momentum and weight decay parameters are set to 0.9 and 0.0005, respectively. The candidate temperature set Ψ is from a range of numbers from 0.001 to 1.0 with step of 0.001. We set the energy anchor = 10 and balance hyper-parameter λ = 0.1 for all benchmarks. |