Autoencoder-Based Hybrid Replay for Class-Incremental Learning
Authors: Milad Khademi Nori, Il Min Kim, Guanghui Wang
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We provide comprehensive experiments to demonstrate the strong performance of AHR: we conduct our experiments across five benchmarks and ten baselines to showcase the effectiveness of AHR utilizing HAE and RFA while operating with the same memory and compute budgets. ... Table 2: Empirical evaluation of AHR against a suite of CIL baselines. Accuracies and the SEMs. |
| Researcher Affiliation | Academia | 1Department of Computer Science, Toronto Metropolitan University, Toronto, Ontario, Canada 2Electrical and Computer Engineering, Queen s University, Kingston, Ontario, Canada. Correspondence to: Milad Khademi Nori <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 Autoencoder-Based Hybrid Replay Algorithm 2 CCE Placement Algorithm 3 HAE Training Algorithm 4 Memory Population |
| Open Source Code | Yes | Implementation is available at github.com/miladkhademinori/autoencoderhybrid-replay-cil. |
| Open Datasets | Yes | We have MNIST(5/2) (Le Cun et al., 2010), Balanced SVHN(5/2) (Netzer et al., 2011), CIFAR-10(5/2) (Krizhevsky et al., 2009), CIFAR-100(10/10) (Krizhevsky et al., 2009), and mini Image Net(20/5) (Vinyals et al., 2016) benchmarks. |
| Dataset Splits | Yes | The series of tasks for CIL are constructed according to (Masana et al., 2020; Ven et al., 2021; Zaj ac et al., 2023), where the popular image classification datasets are split up such that each task presents data pertaining to a subset of classes, in a non-overlapping manner. For naming benchmarks, we follow (Masana et al., 2020), where dataset D is divided into T tasks with C classes for each task. Hence, a benchmark is named as D(T/C). |
| Hardware Specification | No | No specific hardware details (e.g., GPU/CPU models, memory) are provided in the paper. |
| Software Dependencies | No | The paper mentions using 'Adam (Kingma & Ba, 2014) as the optimizer' and 'Res Net-32' for network architecture, but does not provide specific version numbers for software libraries or frameworks (e.g., Python, PyTorch, TensorFlow, CUDA). |
| Experiment Setup | No | The main text states: 'The learning rates, batch sizes, and strategy-dependent hyperparameters are detailed in Appendix B in the supplementary material.' However, specific concrete values for these hyperparameters are not provided in the main text of the paper. |