CLASS-INCREMENTAL LEARNING USING GENERATIVE EXPERIENCE REPLAY BASED ON TIME-AWARE REGULARIZATION

The authors should describe the experimental tasks in more detail, since the benchmarks mentioned can be implemented in many different ways. For example: - Are output units added incrementally for new classes, or are the same original output units mapped to multiple classes? - Permuted-MNIST is typically used for task-incremental continual learning. The authors mention 10 tasks, but it is not clear whether that refers to 10 random permutations (each with its 10 digits), or to a single permutation (with the tasks being the individual digits). If the benchmark is used in a task-incremental way (10 permutations each with the full 10 digits available for the same permutation), then the EWC result with an accuracy of just 27% is puzzling, since the original paper -in that setting- achieves over 97-98% accuracy. - Same for CIFAR-100; it is not clear what the 10 tasks refer to. Is incremental classification performed on 10 tasks at a time? The authors should be clear about it.

Minor: - some typos, e.g., "The idea is to stor and the", or "laten space" in Fig. 6(b), or section 8.2 missing capital letter at the beginning of the sentence "our method is only applicable".

1. The rationale for the scheduling in Section 4.4 is unclear. Although the feedback concept aligns with models like ACGAN, the authors' specific scheduling approach lacks theoretical or intuitive foundations. Displaying $\alpha(\widehat{y})$, $\beta(\widehat{y})$, and their ratio, coupled with a discussion, might offer better insight.
2. Figure 3 reveals the proposed method's diminishing plasticity over time, particularly in tasks 9 & 10. The unaddressed plasticity-stability dilemma limits its real-world applicability. Also, the introduction should cover the plasticity stability dilemma.
3. The FID score is redundant; better stability metrics like BWT exist~\cite{diaz2018don}.
4. Section 6 can benefit from presentation polishing.
5. Section 7 is not an analytical analysis.
6. The extended neuroscience connection can be condensed to make room for implementation details, enhancing reproducibility.
7. The study can also benefit from comparisons with more recent SOTA methods as BI-R is 2020 and EWC is even older.

1. The paper's novelty appears limited, as similar ideas have been explored in prior works. For instance, BioSLAM [1] employs a time-decay factor to reduce the importance of older samples (see section 5.B.3 in [1]), resembling the time-aware regularization introduced in this paper. Could you clarify the distinctions and unique contributions of this paper in comparison to the time-decay importance mechanism in [1]?

2. The structure of the time-aware regularization method presented in section 4.4 seems tricky. It would be helpful to conduct ablation studies on different time-aware regularization functions, e.g., linear decay, polynomial decay, and exponential decay (as demonstrated in the paper).

3. The scope of the experiments is somewhat limited. It might be beneficial to expand the experimental setup to include different numbers of tasks, such as a 5-task or 20-task configuration on CIFAR-100 datasets, as discussed in [2].

4. For datasets with high complexity and resolution, such as ImageNet, the quality of sample generation using VAE might not meet acceptable standards, which could potentially restrict the performance of the proposed method. It would be helpful to conduct experiments on large datasets like ImageNet, as discussed in [3].

[1] P. Yin, et al. "BioSLAM: A Bioinspired Lifelong Memory System for General Place Recognition." IEEE Transactions on Robotics (2023).
[2] J. James, et al. "Always be dreaming: A new approach for data-free class-incremental learning." ICCV. 2021.
[3] X. Hu, et al. "Distilling causal effect of data in class-incremental learning." CVPR, 2021.

• My biggest concern about this paper is the scenario and the restrictions the authors are using. On one hand, they motivate the idea of not using a memory buffer due to privacy concerns. However, for a generative model to work in this scenario, it is a requirement that it generates data as similar as possible to past distribution, creating an issue about the data replicability. On the other hand, they motivate the idea of using a generative model to have a fixed amount of memory allocated, something that is normally done and easily achieved with memory-based methods.
  • I understand the first motivation if the generative model would work in a latent space. With a generated vector should be unlikely to replicate sensitive information.
• Assuming that generative-based methods can be used in some scenarios where memory-based methods can not. The limitations on the dataset that can be represented with generative methods limit them to particular problems.
  • I recommend finding an application, example or scenario where generating the input can help to present a more stimulating idea.
• Figure 1 needs to be clarified. I don't know what each color of the arrows represents.
• The example in section 4.3 needs to be clarified. Class 0 was seen in the first task?
  • Can we say that "t" is like the task ID?
  • What do you mean by "predicted numeral class label"?
• I appreciate works that do ablation studies. However, it isn't easy to understand what the colors and figures represent.