Bi-Objective Continual Learning: Learning ‘New’ While Consolidating ‘Known’
Authors: Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Yihong Gong5989-5996
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments on CIFAR10/100, CORe50 and a subset of Image Net validate the BOCL framework. We also reveal the performance accuracy of different sampling strategies when used to finetune a given CNN model. |
| Researcher Affiliation | Academia | 1Faculty of Electronic and Information Engineering, Xi an Jiaotong University 2School of Software Engineering, Xi an Jiaotong University 3Research Center for Artificial Intelligence, Peng Cheng Laboratory |
| Pseudocode | Yes | Algorithm 1 Generating the pillar set. |
| Open Source Code | Yes | 1The code is released at https://github.com/xyutao/bocl. |
| Open Datasets | Yes | We use CIFAR10/100 (Krizhevsky and Hinton 2009), CORe50 (Lomonaco and Maltoni 2017) and a 1000-class subset of Image Net (Deng et al. 2009) as the benchmark datasets. |
| Dataset Splits | Yes | Both datasets contain 60,000 natural RGB images of the size 32 32, including 50,000 training and 10,000 test images. (...) CORe50 dataset. It consists of 164,866 images of 50 domestic objects, which are split into eleven sessions. Three sessions (#3, #7 and #10) are selected for test and the remaining ones for training. (...) Sub Image Net. It contains 250,000 training images that are split into 5 training sessions. We use the original 50,000 validation images as the test set. |
| Hardware Specification | No | The paper does not explicitly describe the hardware used for running experiments (e.g., specific GPU/CPU models, memory, or cloud instances). |
| Software Dependencies | No | The paper mentions using 'Res Net18' as the baseline CNN model but does not specify any software libraries or dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | For CIFAR10/100, the models are trained on each session using SGD with a mini-batch size of 100. When training on S1, we set the initial learning rate to 0.01, and decrease it to 0.001 after 15 epochs. We stop training when the cross-entropy loss becomes stable, which takes about 20 epochs in total. Then, we use a constant learning rate of 0.001 and finetune each subsequent session for 20 epochs. (...) For sub Image Net, we train the models on each session for 60 epochs with a larger initial learning rate 0.1. (...) We use SOMs of sizes 15 15, 32 32, 25 25, and 64 64 for CIFAR10/100, CORe50 and sub Image Net, respectively. We set λ = 10 in Eq. (1), γ = 10 and Kp = 1 (...) We sample Ks = 1, 000 examples for CIFAR10/100 and CORe50, and Ks = 10000 for sub Image Net, from Si to form the training set Si. |