Knowledge Swapping via Learning and Unlearning

We sincerely thank Reviewer ab73 for the valuable comments. Reviewer ab73 notes that our claims are well-supported by experimental evidence and technical details, acknowledges that Tables 1–3 demonstrate superior results for our proposed method compared to the reverse approach, commends our two-stage approach as well-motivated by our analysis of how learning and forgetting affect different layers of neural networks, and highlights the generalizability of our method across various vision problems. We address the main concerns as follows:

Additional results using other architectures, such as CNNs.

Good point. We conduct additional experiments using the ResNet-18 architecture to learn five new classes and forget five original classes. These additional results consistently support the key insight of "Learning before Forgetting.”

More ablation studies about individual components, such as the impact of different regularization strengths and the importance of sparse constraints.

Thank you for this suggestion. The table below summarizes the performance metrics on the CUB dataset when excluding sparse constraints. Metrics are provided under various conditions (default settings: $\alpha_l = \alpha_f = 0.01$):

More ablation studies about hyperparameters ($\alpha$, $\beta$, BND).

Below we provide detailed ablation results illustrating the impact of various hyperparameters on detection performance using the CUB dataset as the learning set.

Effect of BND (default: BND=15)

Effect of $\beta_{forget}$ (default: $\beta_{forget}=0.2$)

Effect of $\alpha$ (default: $\alpha_l = \alpha_f = 0.01$)

Computational efficiency discussion.

Thank you for highlighting this aspect. The following table provides a comparative analysis of training time, inference speed per image, and the number of trainable parameters for various model architectures. It demonstrates the trade-offs among training duration, inference efficiency, and parameter efficiency:

Should further clarification or additional details be necessary, we welcome further discussion during the rebuttal period.

We sincerely thank Reviewer tJYm for the insightful and constructive comments. Reviewer tJYm finds that "Knowledge Swapping is a practical and novel task", “the experimental design is overall valid”, and "learning before forgetting would be a substantial contribution". We address the main concerns below.

Reviewer tJYm think (1) weight norms in Fig. 2 do not convincingly validate the key claim; (2) more metrics like CKA are suggested to make a stronger argument.

Thanks. (1) We would like to clarify potential confusion regarding Fig. 2. Specifically, in the $L^W→F$ (Fig.2(a)), the shallow layer weight norms clearly increase significantly during the learning phase compared to those in the $F^W→L$ (Fig.2(b)). Given that we initialize the weights uniformly using Kaiming initialization, resulting in an average initial norm of approximately 1.12, the observed increase in shallow-layer norms in Fig. 2(a) implies substantial updates occurring at shallow layers during learning (thereby subsequently impacting the middle-to-deep layers). Conversely, the elevated norms observed in Fig. 2(b), when the forgetting phase precedes learning, occur primarily in middle-to-deep layers. This difference indicates that when learning precedes forgetting, modifications mainly affect higher-level semantic information, whereas in the forgetting-first scenario, substantial updates are more prominent in shallow layers. We hope this clarification adequately addresses the reviewer's concern.

(2) Good suggestion! We have conducted additional validation experiments using the CKA metric as recommended. The detailed results are available (Fig. S1 and Fig. S2) at the anonymous link (https://anonymous.4open.science/r/rebuttal-C764/figures.pdf) and further corroborate our original findings.

Should additional clarification be required, we would gladly engage in further discussion during the rebuttal period.

Reviewer tJYm suggests more sequential $L$ and $F$ phase.

Good point. We have explored sequential phases of learning and forgetting as reported in Tab.2 (F→L→F and L→F→L). We have now extended these experiments by adding further cycles, as shown in the updated results below:

The expanded experimental results continue to support our claim: concluding with F→L makes forgetting significantly more challenging, whereas ending with L→F effectively alleviate this. Additionally, multiple cycles appear to contribute to incremental performance improvements.

Reviewer tJYm questions whether the observed forgetting results reflect genuine forgetting or merely the model's ability to nullify LoRA weights.

Thanks. We respectfully provide an alternative interpretation based on the following considerations: (1) Our "knowledge swapping" task explicitly requires the model to forget knowledge originally learned during pretraining. Merely nullifying LoRA parameters (setting them to 0) would return the model to its pretraining state without achieving actual forgetting of previously learned knowledge.

(2) Additionally, we conduct further experiments using CNN-based full tuning (ResNet-18). We set the model to learn 5 new classes and forget 5 original classes. These results still demonstrate the validity of our key insight. (Please see the first response for R#ab73).

Should further clarification be necessary, we would welcome additional discussion during the rebuttal period.

Reviewer tJYm suggests to explore more results about the forgetting of emergent knowledge not included in the pretraining data.

Thanks. We additionally conduct segmentation experiments on the COCO dataset, where "l5f1" denotes learning 5 new classes and forgetting 1 new class (with other settings similar), none of which are included in the pretraining. Interestingly, the accuracy for the forgotten class dropped to 0. This indicates that forgetting emergent knowledge (previously unknown during pretraining) can be readily achieved by nullifying specific LoRA parameters. However, our primary scenario involves forgetting previous knowledge (distinct from "emergent knowledge" as defined in Sec. 3.1). For detailed results and discussions, please refer to the previously response.

We thank Reviewer 3jJx for the valuable comments. Reviewer 3jJx appreciates that "Knowledge Swapping is a good task," "is interesting and novel," "well-structured," represents a "novel and intriguing concept," and is "robust." Below, we provide detailed responses addressing the remaining concerns.

Reviewer 3jJx asks for more results in Fig.5 & Fig.7.

Thank you for this suggestion. We have included additional related results (Fig. S3 and Fig. S4) in the supplementary material accessible via the anonymous link (https://anonymous.4open.science/r/rebuttal-C764/figures.pdf). These results will also be included explicitly in our final manuscript.

Reviewer 3jJx suggests providing further discussion about the limitations of the current approach, specifically regarding the varying difficulty in learning and forgetting different types of knowledge, and outlining potential future directions.

We greatly appreciate Reviewer 3jJx’s insightful recommendation to elaborate on limitations. Our experiments have demonstrated that the difficulty in acquiring new knowledge and the ease of forgetting existing knowledge indeed vary significantly across different categories. Investigating these variations presents a meaningful future research direction. Specifically, exploring and characterizing the complexity associated with different knowledge categories can reveal critical insights. One promising approach is to incorporate uncertainty-based assessment methods, as discussed by [R1], to better evaluate model confidence and quantify these complexities. Employing uncertainty estimation can further elucidate the underlying mechanisms influencing learning and forgetting within our proposed framework. Ultimately, this line of inquiry may foster the development of more robust, targeted, and efficient strategies in future research.

References:

[R1] Gawlikowski, Jakob, et al. "A survey of uncertainty in deep neural networks." Artificial Intelligence Review, 2023.