Vision and Language Synergy for Rehearsal Free Continual Learning
Authors: Muhammad Anwar Masum, Mahardhika Pratama, Savitha Ramasamy, Lin Liu, H Habibullah, Ryszard Kowalczyk
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experimental analysis shows the superiority of our method to existing SOTAs in CIFAR100, Image Net R, and CUB datasets with significant margins i.e. up to 30% final average accuracy, 24% cumulative average accuracy, 8% final forgetting measure, and 7% cumulative forgetting measure. Our historical analysis confirms our method successfully maintains the stability-plasticity trade-off in every task. Our robustness analysis shows the proposed method consistently achieves high performances in various prompt lengths, layer depths, and number of generators per task compared to the SOTAs. We provide a comprehensive theoretical analysis, and complete numerical results in appendix sections. |
| Researcher Affiliation | Academia | 1University of South Australia, Mawson Lakes, SA, 5095, Australia 2Institute for Infocomm Research, A*STAR & IPAL, CNRS@CREATE |
| Pseudocode | Yes | In this section we present the procedures of LEAPGen training and inference that are shown in algorithm 1 and 2 respectively. Algorithm 1 LEAPGen Training Algorithm 2 LEAPGen Inference |
| Open Source Code | Yes | The method code is available in https://github.com/anwarmaxsum/LEAPGEN for further study. |
| Open Datasets | Yes | Datasets: We evaluate our method in three CIL benchmarks i.e. CIFAR100(Hendrycks et al., 2021), Image Net-R(Belouadah & Popescu, 2019), and CUB(Wah et al., 2011). |
| Dataset Splits | Yes | We follow 5, 10, and 20 task splits as in the SOTAs (Roy et al., 2024; Gao et al., 2024; Kurniawan et al., 2024). |
| Hardware Specification | Yes | All the methods are run under the same environment i.e. a single NVIDIA A100 GPU with 40 GB RAM. |
| Software Dependencies | Yes | All the consolidated methods are run under the same machine and computing environment i.e. single NVIDIA A100 GPU with 40 GB memory, python 3.8 and Pytorch 2.2.0. |
| Experiment Setup | Yes | We utilize Adam optimizer, set 128 batch-size, and cosine learning scheduler similar to Conv Prompt and CODA-P. The initial learning rate is set to 0.01, 0.05, and 0.005 for CIFAR100, Image Net-R, and, CUB respectively, with 20 maximum epochs. ... The λ2 and λ3 are set to 1.0, while λ1 is set to 1.0 for CIFAR100, and 0.1 for CUB and Image Net-R. The number of generators per task that is the same as top-k descriptors is set to 3. |