Recall-Oriented Continual Learning with Generative Adversarial Meta-Model
Authors: Haneol Kang, Dong-Wan Choi
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through our experiments, we show that our framework not only effectively learns new knowledge without any disruption but also achieves high stability of previous knowledge in both task-aware and task-agnostic learning scenarios. Our code is available at: https://github.com/bigdata-inha/recall-orientedcl-framework. |
| Researcher Affiliation | Academia | Haneol Kang, Dong-Wan Choi* Department of Computer Science and Engineering, Inha University, South Korea haneol0415@gmail.com, dchoi@inha.ac.kr |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available at: https://github.com/bigdata-inha/recall-orientedcl-framework. |
| Open Datasets | Yes | MNIST (Le Cun et al. 1998), SVHN (Netzer et al. 2011), and CIFAR-10 (Krizhevsky, Hinton et al. 2009) datasets, equally consisting of 10 classes. Split CIFAR-10 is constructed by dividing the original CIFAR-10 dataset into 5 tasks, each with 2 classes. Similarly, Split CIFAR-100 consists of 10 tasks, each with 10 classes. PMNIST is a variant of MNIST, where each task has a different random permutation on image pixels. 5-Datasets consists of five different datasets: CIFAR-10, MNIST, SVHN, FMNIST (Xiao, Rasul, and Vollgraf 2017), and not MNIST (Bulatov 2011), each of which is learned as a single task. |
| Dataset Splits | No | The paper mentions "training data D t train and test data D t test" for each task, but does not provide specific percentages, sample counts, or explicit instructions for validation splits within these tasks. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts) used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiment. |
| Experiment Setup | Yes | We use a simple GAN architecture for GAMM, where the generator G consists of 100 input units followed by two fully-connected (FC) layers of 200 units and the discriminator D has one FC layer of 256 units followed by a binary output head. The output dimensionality of G varies depending on the chunk size. For inference models to make prediction, GAMM generates lightweight neural networks, which are reduced versions of the original architectures (e.g., 0.15 times the original size of Res Net-32) that have been commonly used in the literature. |