Task-Free Continual Generation and Representation Learning via Dynamic Expansionable Memory Cluster
Authors: Fei Ye, Adrian G. Bors
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that the proposed Continual Variational Autoencoder (CAA) can produce diverse images across all learnt data domains over time, which is consistent with our theoretical results. |
| Researcher Affiliation | Academia | Fei Ye1,2, Adrian G. Bors1,2 1Department of Computer Science, University of York, York YO10 5GH, UK 2Machine Learning Department, Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE fy689@york.ac.uk, adrian.bors@york.ac.uk |
| Pseudocode | Yes | Algorithm Implementation We provide the algorithm for learning the proposed CAA, which can be summarized into three steps in a training time: Step 1. At a training time Ti, the memory buffer is updated using the Reservoir approach (Vitter 1985). The generator and inference model are updated on the memory buffer using Lg from Eq. (5) and then the discriminator using Ld from Eq. (3). Step 2. The inference model is updated using Linf from Eq. (6). Step 3 (Memory expansion). If the temporary memory buffer is full, |Mt i| = |Mt i|max and Md is empty, the first memory cluster Md[1] is created using Mt i. This is used for checking the memory expansion, and if Eq. (8) is satisfied, we perform the sample selection and transfer the temporary memory buffer to a new memory cluster Md[k + 1] using Eq. (11) while clearing up the temporary memory buffer to avoid learning samples which are statistically similar to those already stored. |
| Open Source Code | Yes | The source code and supplementary material (SM) are available at https://github.com/dtuzi123/DEMC. |
| Open Datasets | Yes | For the class-incremental learning paradigm, we adopt the standard datasets, including Split MNIST (Le Cun et al. 1998), Split SVHN (Netzer et al. 2011), Split Fashion MNIST (Fashion) (Xiao, Rasul, and Vollgraf 2017) and Split CIFAR10 (Krizhevsky and Hinton 2009) (See details in Appendix-D2 from SM). |
| Dataset Splits | No | The paper mentions training and test datasets for evaluation but does not explicitly provide details on validation dataset splits (e.g., percentages, sample counts, or specific configurations for a validation set) in the main text. |
| Hardware Specification | No | The paper does not provide any specific hardware details such as GPU or CPU models, memory specifications, or types of computing resources used for the experiments. |
| Software Dependencies | No | The paper does not specify software dependencies with version numbers (e.g., 'Python 3.8, PyTorch 1.9') required to replicate the experiments. |
| Experiment Setup | Yes | We test the expansion threshold λ in Eq. (8) for values between 5 to 30. Then, we consider λ for Split MNIST, Split Fashion, Split SVHN and Split CIFAR10, Split MSC, Celeb A and Image Net under the generation task, as 28, 30, 29, 29, 21, 27 and 26, respectively. We also empirically find that employing λ2 = 1 in Eq. (5) performs well. For the classification task, the final λ for Split MNIST, Split CIFAR10, Split CIFAR100, Split MImage Net and Permuted MNIS is 20, 22, 21, 25 and 25, respectively. Each memory cluster in the evolved memory buffer Md can store up to 64 samples and the batch size used during each training time is 64. |