Centroids Matching: an efficient Continual Learning approach operating in the embedding space
Authors: Jary Pomponi, Simone Scardapane, Aurelio Uncini
TMLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that Centroids Matching achieves accuracy gains on multiple datasets and scenarios. |
| Researcher Affiliation | Academia | Jary Pomponi EMAIL Department of Information Engineering, Sapienza University of Rome, Italy Simone Scardapane EMAIL Department of Information Engineering, Sapienza University of Rome, Italy Aurelio Uncini EMAIL Department of Information Engineering, Sapienza University of Rome, Italy |
| Pseudocode | No | The paper describes the methodology using mathematical equations and descriptive text, but no explicit pseudocode blocks or algorithms are presented. |
| Open Source Code | Yes | The code containing all the files necessary to replicate the experiments is available here. [hyperlink: https://github.com/SapienzaVision/CentroidsMatching] |
| Open Datasets | Yes | Dataset: We conduct extensive experiments on multiple established benchmarks in the continual learning literature, by exploring both TIL as well as the harder CIL. The datasets we use to create the scenarios are: CIFAR10, CIFAR100 (Krizhevsky, 2009), and Tiny Image Net (a subset of Image Net (Deng et al., 2009) that contains 200 classes and smaller images). |
| Dataset Splits | No | The paper describes how classes are grouped into tasks (e.g., CIFAR10 into 5 tasks with 2 classes each) and mentions using 'test splits of the tasks'. However, it does not provide explicit train/validation/test split percentages or sample counts for the original datasets (CIFAR10, CIFAR100, Tiny Image Net) before task creation, nor for the support sets that are extracted from the training set. While it implies standard approaches, specific details are not provided. |
| Hardware Specification | No | The paper describes the model architecture (ResNet20) and training parameters but does not specify any hardware details such as GPU models, CPU types, or memory used for running the experiments. |
| Software Dependencies | No | To perform all the experiments, we used the Avalanche framework, which implements the logic to create tasks and evaluate the CL approaches. While 'Avalanche framework' is mentioned, no specific version number is provided for it or any other key software libraries, making the software environment not fully reproducible. |
| Experiment Setup | Yes | Hyper-parameters: for each method, we searched for the best hyper-parameters, following the results presented in respective papers. For EWC, we used 100 as regularization strength weight for all the scenarios. For GEM we used a memory for each task, composed of 500 samples for CIFAR10 and 1000 for the other experiments. [...] Regarding our approach, the support set contains 100 images from the training set of each task, and we set the penalty weight λ to 0.1 for CIFAR10, 0.75 for CIFAR100 and Tiny Image Net; regarding the CIL scenarios, we used a fixed size memory of 500 for each scenario. [...] Models and Training: for each dataset we use Res Net20 (He et al., 2016) architecture, trained using SGD with learning rate set to 0.01 and momentum to 0.9. For CIFAR10-100, we trained the model for 10 epochs on each task, while for Tiny Imagenet we used 30 epochs; [...] We repeat each experiment 5 times; each time the seed of the experiment is changed in an incremental way (starting from 0). [...] Also, we used the following augmentation schema for the proposed datasets: the images are standardized, randomly flipped with probability 50%, and then a random portion of the image is cropped and resized to match the original size. |