Summary
To address the issues of insufficient representation capability and class confusion caused by distribution drift in existing prototype-based classification methods in class incremental learning, this paper proposes a method named CREATE. Based on a replay strategy, this method introduces an autoencoder module for each class to learn compact manifold structures. Simultaneously, to prevent class confusion, contrastive learning is employed in the latent space across autoencoders of all classes to increase inter-class distances. During classification, the reconstruction loss of sample features on each autoencoder is calculated as the basis for classification. The lightweight nature of the auto-encoder modules also makes this method feasible for use in resource-constrained environments, potentially broadening the applicability of CIL methods in the real world. Experiments demonstrate the effectiveness of the proposed method.
Strengths
The paper proposes an innovative solution to the class confusion problem in class incremental learning (CIL) by introducing a CREATE classifier. This approach is original in that it leverages class-specific auto-encoders to disentangle class manifolds, which is a novel application in the context of CIL. Unlike traditional methods that primarily rely on prototypes or parameter-based strategies to combat catastrophic forgetting, CREATE uses a lightweight auto-encoder model to capture compact manifolds for each class, addressing the issue of class overlap caused by distribution drift. This paper is well-written, logically clear, and expresses the research objectives intuitively through visualization.
Weaknesses
Absence of Analysis on Inter-Class Similarity: The model uses class-specific auto-encoders and separation loss to reduce class-wise confusion, but it does not address how well it performs when classes are highly similar. In real-world applications, some classes may share overlapping features, leading to higher confusion. Providing analysis or visualizations of how the model handles such cases would strengthen the claim that CREATE effectively reduces class-wise confusion.
Class-Specific vs. Unified Reconstructor: The paper lacks a comparison between class-specific and unified feature reconstructors. Clarifying this difference, along with comparisons to [1] and [2], would better highlight CREATE’s unique advantages.
[1] Zhou, Y., Yao, J., Hong, F., Zhang, Y., & Wang, Y. (2024). Balanced destruction-reconstruction dynamics for memory-replay class incremental learning. IEEE Transactions on Image Processing.
[2] Zhai, J. T., Liu, X., Bagdanov, A. D., Li, K., & Cheng, M. M. (2023). Masked autoencoders are efficient class incremental learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 19104-19113).
Questions
(1). When replay samples are involved in training, does the feature shift remain as severe? Additionally, since the feature reconstructor for previous tasks is updated while training on the current task, is there a risk of catastrophic forgetting?
(2). The reconstruction module currently uses only simple convolutional and activation layers. If more complex layers were introduced in this module, would the performance improve?
(3). Does the performance improve if the contrastive loss term L_CR is used independently, without the reconstruction module?
(4). There is a typo in Table 1. The best result of Inc 5 experiment on CIFAR100 B50 dataset is obtained by DSGD (63.58) instead of CREATE (63.53).
(5). During ablation experiments, how are the prototypes in NCM obtained? Do they update with tasks? If only prototypes obtained from previous tasks are used without any updating strategy, severe forgetting is inevitable.