Online Continual Learning with Maximal Interfered Retrieval
Authors: Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, Lucas Page-Caccia
NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We present our approach based on a replay buffer or a generative model in Sec. 3 and show the effectiveness of our approach compared to random sampling and strong baselines in Sec. 4. |
| Researcher Affiliation | Academia | Rahaf Aljundi KU Leuven rahaf.aljundi@gmail.com Lucas Caccia Mila lucas.page-caccia@mail.mcgill.ca Eugene Belilovsky Mila eugene.belilovsky@umontreal.ca Massimo Caccia Mila massimo.p.caccia@gmail.com Min Lin Mila mavenlin@gmail.com Laurent Charlin Mila lcharlin@gmail.com Tinne Tuytelaars KU Leuven tinne.tuytelaars@esat.kuleuven.be |
| Pseudocode | Yes | Algorithm 1: Experience MIR (ER-MIR) and Algorithm 2: Generative-MIR (GEN-MIR) are provided in the paper. |
| Open Source Code | Yes | We release an implementation of our method at https://github.com/optimass/Maximally_Interfered_Retrieval. |
| Open Datasets | Yes | MNIST Split splits MNIST data to create 5 different tasks with non-overlapping classes... CIFAR-10 Split splits CIFAR-10 dataset... Mini Imagenet Split splits Mini Imagenet [38] dataset... |
| Dataset Splits | Yes | CIFAR-10 Split splits CIFAR-10 dataset into 5 disjoint tasks as in Aljundi et al. [3]. However, we use a more challenging setting, with all 9,750 samples per task and 250 retained for validation. |
| Hardware Specification | No | The paper does not provide any specific details regarding the hardware used for running the experiments, such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper does not specify any software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions) that would be needed to replicate the experiments. |
| Experiment Setup | Yes | We use the same learning rate, 0.05, used in Aljundi et al. [3]... The number of samples from the replay buffer is always fixed to the same amount as the incoming samples, 10, as in [8]... More hyperarameter details are provided in Appendix B.2. |