WHICH RESTRAINS FEW-SHOT CLASS-INCREMENTAL LEARNING, FORGETTING OR FEW-SHOT LEARNING?

1. First of all, I must highlight that the claim point in the introduction (Figure 1) may not be true for all cases. In the introduction, the authors evaluate the performance of old and new classes on several methods and find the accuracy of new classes is much lower than that of old classes. However, this claim point may be biased since (almost) all few-shot class-incremental learning methods tend to freeze the backbone and extract prototypes for new classes (now a common practice). Hence, the old classes’ performance will be maintained as the model is frozen, restricting its plasticity. We can assume a typical baseline to update the model without freezing any parameters with only few-shot instances, where we definitely will find the accuracy of new classes will be larger than old ones due to catastrophic forgetting. In other words, the analysis in Figure 1 may not be a sufficient enplanement for the conclusion that “few-shot instances restrain few-shot class-incremental learning”.
2. Secondly, the main idea of this paper is somehow incremental due to the overlapping of another typical work in the few-shot learning field [1]. In [1], the authors find similar classes have similar mean and variance. Hence, they utilize the similarity-based weight to calibrate the distribution of new classes with old classes. The core idea is almost the same as the current method (which only changes the similarity-based calibration into OT-based calibration). Hence, highlighting the difference to [1] requires careful discussion and ablation. I would also like to see the results when using [1] to sample new class features in the current method.
3. In my opinion, Eq. 7 calibrates the prototypes of new classes with old classes and helps the learning of new classes. However, it still has a possible negative impact on the old classes. Taking a 2-class classification problem for an example, if you are moving the prototype of class 2 nearer to class 1, the classification accuracy of class 1 will correspondingly decrease since the new prototype pushes the decision boundary. Discussing such negative impacts and why the final performance will improve would be better. Additionally, showing the accuracy of old and new classes will clarify this.
4. The performance comparison in Table 2 shows that the main improvement comes from the base task training. Specifically, compared to NC-FSCIL, “Ours (Prototype-based learning)” has better performance in the former stages while having lower performance in the latter stages. Hence, the main improvement in the “Avg Acc” column comes from the improvement of the base session instead of the capability to resist forgetting. Such a learning protocol seems to avoid the inherent goal of class-incremental learning, leading to biased works that work on pre-training the model with complex tricks instead of designing useful training techniques.

It is understandable that "...performance of new classes is much lower than that of old classes..." (page 2) because the images of the old classes significantly outnumber the images of the new classes. Hence, this observation is somewhat trivial.

The main contribution is exactly the same as "Distribution Calibration (DC)" which has already been widely adopted in few-shot learning. If this work simply borrows the same idea from few-shot learning to FSCIL, then the contribution is quite limited.

1. The paper could strengthen its claims by testing the method across a wider variety of datasets and incremental learning scenarios to ensure the generalizability of the results.
2. While the paper compares the proposed method with state-of-the-art approaches, a more detailed analysis of how the method performs relative to a broader range of existing methods could provide a clearer picture of its advantages and limitations.