Continual Learning with Global Alignment

Summary

This paper tackles continual learning by addressing interference between tasks. For the interference between different tasks, the authors propose a method called ‘global alignment’ to align the data representations using task-specific compositions of pre-trained token representations. Then the authors conduct extensive experiments to verify the effectiveness of the proposed method.

Strengths

1. This paper is well-written and easy to understand. 2. The motivation is clear. From the perspective of cross-task interference, this paper provides some simple but effective methods to avoid the interference of the data representation and classifier. 3. The proposed methods are easy to follow and experiments verify their effectiveness.

Weaknesses

1. The analysis of interference in Section 3.2 does not consider the activation function between the two layers of the network. Multiplying the two linear weight matrices is equivalent to using only one weight matrix, so the analysis of this linear case is different from that of the non-linear case and I’m curious about the inference in the non-linear case. 2. About wiring with neighbor attention: a) The author needs to explain the rationality of neighborhood tokens and the impact of the size of K on the model. b) This method requires matching the most similar K tokens for each token given a sample. This process may require a large amount of computation. c) The author needs to further explain the difference between this method and controlled-LoRA. Both these two methods add a learnable term to the original token representations but adopt different generation strategies. Specifically, this method adds a term composed of neighborhood tokens, while controlled-LoRA uses low-rank matrices. So what are the advantages of this method? 3. About Controlled-LoRA: a) Computational burden and parameter size: As the number of tasks increases, linearly increasing model parameters is not in line with the spirit of CL, and the increased parameter size should be limited even if LoRA only adds small parameters compared to the large model. b) Why does Controlled-LoRA perform worse than other methods? In the case of task-IL, given the task ID for each test sample, we can directly use the best LoRA model.

Questions

Please refer to Weaknesses.

Rating

6

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

NA

Summary

In Continual Learning (CL), the interference caused by the constant modification of the representation is the leading cause of catastrophic forgetting. Motivated by this and the idea that gradients in opposite directions are one cause of this interference, the authors proposed new ways of adding knowledge to a model. The authors motivate the proposal with a toy model that exemplifies where and why the interference occurs, concluding that it comes from the discrepancy between the representations of the different tasks and the relationship between the class vectors. These conclusions lead the authors to propose three proposals that slightly modify the transformer architecture, specifically adding learnable weights in the self-attention layers to generate a task-specific attention representation. The authors also propose using the probing and then fine-tuning approach presented in the past to help initialize class vectors. The results are presented in task and class incremental learning benchmarks in sequences of text datasets.

Strengths

- The authors present a motivation from which they exemplify the problem in a simple way and where they want to aim their solution. This motivation may help the reader understand the context and problem to be solved. - The paper introduces a method that alters the architecture of a transformer model. While the paper does not explicitly state this, the provided motivation helps to understand why certain parts of the self-attention layer (the k matrix) are modified. - The authors must describe why only the K matrix is modified and not the others.

Weaknesses

- What is the actual contribution of applying wiring or C-LoRA? The results show that applying FP can benefit the methods much more than adding the proposals presented. - What are the implications of applying FP to the Adaptation or CL methods? The performance gain from the proposed methods is relatively low (compared to L2P), and this could further decrease with the application of FP. This requires a deeper exploration of the proposed methods and the alternatives. - The writing is unclear and often unnecessarily complex, making reading difficult for someone unfamiliar with the subject. The notation used is only sometimes in line with the literature, which can confuse readers. In addition, there are problems explaining some terms and easily avoidable problems. For example: - Line 46, FP, is mentioned but needs to be correctly defined. However, the document defines the abbreviation FP multiple times after that when only one should be enough. - Line 182, there is an extra “192”.

Questions

- My main concern is the causes of interference that motivate this work. The author hypothesised that a gradient in the same direction can alleviate forgetting between tasks, a concept also adopted by GEM, despite its susceptibility to forgetting. Even with a gradient in the same direction, if it induces significant weight changes, the representations could be compromised, leading to catastrophic forgetting. - How much does having a model pre-trained in tasks similar to those used in CL affect the proposed methods? The gradient will likely be low for these tasks. - Text benchmarks need minor modifications to the model weights due to the similar distribution between the pre-trained model and the new tasks. In cases where higher modifications are needed, can the scaling factor (s) help reduce this effect? Did you explore different values of the scaling factors? - Can this method work on image benchmarks? Images can have a more complex distribution than text, making avoiding interference between tasks difficult. - Is FP applied to CL or adaptation methods in Table 1? - Another technique used in CL is freezing class vectors in the classifier so that only the classes present in the batch during training are modified. This is especially useful when there is no memory. Can this help further reduce interference in the proposed methods? - Fig 2.b, what method of global alignment was used? - Three methods to mitigate interference are presented. Can these methods be complemented by each other? - How many neighbours are used in Wire-Neigh? Did you perform experiments to find the optimal number of neighbours? - Did you train only the classifier for the Task Incremental Learning problem? This method should not forget, and the pre-trained model can have good prior knowledge. - For the CIL experiments, the proposed methods have very similar results to previous methods when applying FT in Yahoo. This somewhat contradicts the claim that the proposed methods mitigate forgetting. Do you have any intuition about why FT+PF works so well? - In Fig 4.b., is the performance with or without FP?

Rating

5

Confidence

4

Soundness

2

Presentation

1

Contribution

2

Limitations

As the authors mention, this method heavily relies on a pre-trained model with a similar distribution to the incoming tasks. This assumption is only sometimes true and can be more complex in scenarios with a more complex input data distribution.

Summary

This paper addresses the problem of Task-Incremental-Learning (TIL) with pre-trained transformer in the context of NLP. The author extended their experiments in the Class-Incremental Learning (CIL) scenario. The authors identify potential forgetting causes as (1) negative correlation between data representations and (2) negative correlation between class prototypes. After theoretically justifying their claim, the authors propose to align the model representations with either a) learning new key matrices for pre-trained queries and values b) learning new key matrices for with additional token chosen as the nearest neighbors of the considered token c) learning new queries and values from the original ones with a low-rank adaptation strategy. The author additionally align prototypes (class vectors) leveraging a probing and fine-tuning strategy. This paper shows superior performances on various dataset in CIL and briefly discuss the effect of various components.

Strengths

- clear and understandable paper - the equations are well derived and clear - the proposed approach has interesting theoretical justifications - The obtained performances on the TIL settings are compelling

Weaknesses

1. While I believe the evaluation to be sufficient, it would improve the paper to include more recent prompt-learning techniques such as CODA [1] which showed stronger performances than L2P. I believe it would equally be interesting to see this approach applied to Computer Vision problems, even though I understand this paper focuses on NLP. 2. Although the proposed approach is theoretically justified, it would be interesting to quantify such alignment through experiments, by computing data representations covariance/correlation between transformed class vectors; for each alignment strategy. 3. I appreciate the effort to include results in CIL scenarios. However, I think more discussions as to how the proposed approach performs poorly compared to ER-ACE could be introduced. 4. What is the link between the findings of this paper and previous work on orthogonal subspaces in continual learning [2, 3]? In these works, amongst others, it seems that the correlation between hidden representation should be zero, and not positive. I think such work should be included in related work section. 5. A discussion regarding the extra computation induced by the alignment strategies and the PF would be welcomed as well, as it seems to increase it considerably. 6. The code is unfortunately not accessible to the reviewers. If the authors can address most of the above points I would happily increase my score. **Typos and presentation** - LoRA l44 and PF l.46 are not defined. Please either define the acronyms or cite the corresponding paper. - l.114 $h_{\tau}$ should be h_{i} - I do not like the use of RHS l.163 and 164. - l. 182: "grounded192"? [1] Smith, James Seale, et al. "Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023. [2] Wang, Xiao, et al. "Orthogonal subspace learning for language model continual learning." Findings of the Association for Computational Linguistics: EMNLP 2023, pages 10658–10671 [3] Chaudhry, Arslan, et al. "Continual learning in low-rank orthogonal subspaces." Advances in Neural Information Processing Systems 33 (2020): 9900-9911.

Questions

See weaknesses.

Rating

7

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

Limitations have been partially addressed in the main draft. I believe a discussion regarding the poor performances on CIL scenarios should be included, as well as a discussion on the potential computation overhead.

Summary

This paper studies the cause of cross-task interference in class-incremental learning of transformer-based language models. The authors disentangle the cause into the correlation (i) between data representations and (ii) between class vectors in the linear classifier. To tackle (i), the authors propose three ways to construct data representations at each layer by learning an attention over the pretrained token representations. To tackle (ii), the authors propose to only train the classifier for a new task (to obtain a good initialization) before jointly training both the classifier and the encoder. The authors perform experiments with the pretrained BERT-base model and various text datasets. Training is in the task-incremental setup (where task labels are provided and the model predicts over in-task classes); the model is evaluated on both task-incremental and class-incremental (where task labels are not provided and the model predicts among all classes) setups. The authors found that their methods, "alignment models", outperform existing adaptation and continual models.

Strengths

1. This paper is well-written and I did not find major technical flaws. 2. I enjoyed reading the motivation of the paper in Sec. 3 where the authors examine the causes of cross-task interference. 3. I find that the initialization of class vectors may influence cross-task interference interesting, although I have some related questions.

Weaknesses

1. While the main goal of the paper is reducing cross-task interference, the main results (Table 1) is using the task-incremental setup, where task labels are known during prediction. The class-incremental learning results are only in Fig. 4, comparing the proposed method only with LoRA and ERACE. 2. I find it a bit hard to infer where the performance improvement comes from from the results. I wonder if it is possible to do some fine-grained ablation analysis that verifies that the proposed method indeed helps by reducing overlap in data representations and in class vectors' features for different tasks. For example, one could measure the accuracy on the task level or look at the confusion matrix, which may reveal some information about cross-task confusions. 3. Is $\Delta W$ shared across tasks? If so, I don't understand why the proposed method does not forget. The authors hypothesize that this is potentially due to referencing pretrained representations. However, since the [CLS] representations are constructed involving $\Delta W$, I'd expect the model to lose some ability to generate good representations for past tasks. Could the authors elaborate on this point?

Questions

1. Could the authors explain the choice of training in TIL while evaluating in CIL, as opposed to training and evaluating in the same setup? 2. (Eq. 2) Since weights are usually initialized from a distribution centered around 0, shouldn't the first term be quite small? 3. (Sec 4.3) Does probing perform softmax over only the new classes? If so, how does it help two class vectors for different tasks to focus on different features, since the loss does not require differentiating between the two classes? 4. (Sec. 4.2) In the wiring models, what is the intuition behind only replacing the key matrix but not the query and value matrices? Minor: 5. (Eq. 1) It seems that interference is eliminated once any of the three components is 0. Is it possible to fix the class vectors to be the canonical basis which are guaranteed to be perpendicular, and only learn the data representations? 6. (Sec. 4.2) In Wire-Neigh, do you need to find the $K$ nearest neighbors for each token? If so, how expensive it is in practice? 7. (L#219) Could you elaborate on what "grounded" means here? 8. Missing references: [1, 2] are two prompting-based CL methods that are shown to perform better than L2P. 9. (L#41-44) I recommend changing the indices to, e.g., "(i), (ii), (iii)", as a different "(2)" is referred to in the next paragraph. 10. (L#53) "by" -> "after" 11. (L#182) "grounded192" -> "grounded" [1] DualPrompt: Complementary Prompting for Rehearsal-free Continual Learning. Wang et al. ECCV 2022. [2] CODA-Prompt: COntinual Decomposed Attention-based Prompting for Rehearsal-Free Continual Learning. Smith et al. CVPR 2023.

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have addressed the limitations clearly in Sec. 6.

Thanks for rebuttal

I thank the authors for answers my questions. My concerns are resolved and I've increased my score to a 6. Regarding W3, I wonder if the reason that a shared $\Delta W$ shows low forgetting could also be due to that it does not contain a lot of parameters. This recent paper [1] shows that PEFT methods enjoy low forgetting when you don't have too many tunable parameters. [1] Thede et al. Reflecting on the State of Rehearsal-free Continual Learning with Pretrained Models. CoLLAs 2024.

Thank you for the rebuttal

I thank the authors for taking the time to answer my question and improve my understanding of the manuscript. **W1** I believe it would make it a fairer comparison to include CODA as a SoTA prompt-learning based method after hyper-parameter, since it outperforms L2P on vision tasks, at least for CIL scenarios. I acknowledge that results might differ for the TIL scenario. I still appreciate the effort and I understand that the limited time for rebuttal is not necessarily sufficient to conduct an extensive hyper-parameter search. **W2** Thank you for the extra experiments, which are convincing. Including them in the main draft could also clarify the impact of the proposed loss on the model alignment. **W3** Thank you for the clarification. I agree that direct comparison to replay-based methods might be unfair. **W4** Thank you for the clarification. I would advise the authors to include such discussion in the related work. **Time consumption** Thank you for the additional information. Such details could be included in the paper or the appendix, as computation is a major focus focus in Continual Learning. Overall, I do not have major concern and I will **increase my score to 7**.

Discussion

Dear Reviewers, This paper has received diverging reviews ranging from 4 to 7. The authors have submitted a rebuttal and the anonymous link to their code https://drive.google.com/file/d/1PNjqBgr-ZUBmarNK-yPZ7GWUdFjNf7t0/view?usp=sharing. Most reviewers have raised their scores after reviewing the authors' rebuttal except one. @Reviewer R2vv: Please read the authors' rebuttal/code and other reviewers' comments. Have the authors' responses addressed your concerns? If so, please reconsider the borderline decision. AC

Thank you for the rebuttal

Thanks for answering my questions. After going through all the other reviews and the given rebuttal, I've increased my score to 6. W1. I understand the difficulty of the non-linearity analysis and thanks for the authors' explanation. W2. The comparison between Wire-Fixed, Wire-Neigh, and C-LoRA is clear, and I suggest adding this part to the main text for better understanding. W3. I'm sorry for the misunderstanding about C-LoRA and my concerns have been addressed.