Memory Replay with Data Compression for Continual Learning
Authors: Liyuan Wang, Xingxing Zhang, Kuo Yang, Longhui Yu, Chongxuan Li, Lanqing HONG, Shifeng Zhang, Zhenguo Li, Yi Zhong, Jun Zhu
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We extensively validate this across several benchmarks of class-incremental learning and in a realistic scenario of object detection for autonomous driving. |
| Researcher Affiliation | Collaboration | Liyuan Wang1,2,3 Xingxing Zhang1 Kuo Yang6 Longhui Yu6 Chongxuan Li4,5 Lanqing Hong6 Shifeng Zhang6 Zhenguo Li6 Yi Zhong2,3 Jun Zhu1 1Dept. of Comp. Sci. & Tech., Institute for AI, BNRist Center, THBI Lab, Tsinghua University 2School of Life Sciences, IDG/Mc Govern Institute for Brain Research, Tsinghua University 3Tsinghua-Peking Center for Life Sciences 4Gaoling School of AI, Renmin University of China 5Beijing Key Laboratory of Big Data Management and Analysis Methods 6Huawei Noah s Ark Lab |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is included in supplementary materials. |
| Open Datasets | Yes | CUB-200-2011 (Wah et al., 2011) is a large-scale dataset... Image Net-full (Russakovsky et al., 2015) includes 1000-class large-scale natural images... Image Net-sub (Hou et al., 2019) is a subset derived from Image Net-full... SODA10M (Han et al., 2021), a large-scale object detection benchmark for autonomous driving. |
| Dataset Splits | Yes | The labeled images are split into 5K, 5K and 10K for training, validation and testing, respectively, annotating 6 classes of road objects for detection. |
| Hardware Specification | Yes | We run each baseline with one Tesla V100... We run each baseline with 8 Tesla V100. |
| Software Dependencies | No | The paper mentions software components like 'Res Net-18 architecture', 'SGD optimizer', 'Faster-RCNN', 'FPN', 'Res Net-50 backbone', 'Pseudo Labeling', and 'Unbiased Teacher', but does not provide specific version numbers for any of these components or their underlying libraries. |
| Experiment Setup | Yes | We train a Res Net-18 architecture for 90 epochs, with minibatch size of 128 and weight decay of 1 10 4. We use a SGD optimizer with initial learning rate of 0.1, momentum of 0.9 and cosine annealing scheduling. ... For each incremental phase, we train the network for 10K iterations using the SGD optimizer with initial learning rate of 0.01, momentum of 0.9, and constant learning rate scheduler. The batch size of supervised and unsupervised data are both 16 images. |