Train-Attention: Meta-Learning Where to Focus in Continual Knowledge Learning

Summary

This paper proposes a meta-learning framework to dynamically adjust token perplexity weights based on their usefulness to achieve good continuous knowledge learning performance and also provides a new benchmark for CKL.

Strengths

1. It is inspiring to use meta-learning technique for adjusting token weight. 2. The proposed benchmark improves on distinguishing plasticity and stability. 3. The experiments are comprehensive and solid.

Weaknesses

1. The idea of adjusting token weights is not fresh enough and shares some similarity with [1]. However, training a meta-learner to evaluate token importance is still a good try. 2. The analysis about token importance (Figure 6 and its corresponding analysis) is not enough. Analysis about attention pattern on these tokens would be a good supplement. [1] Lin Z, Gou Z, Gong Y, et al. Rho-1: Not all tokens are what you need[J]. arXiv preprint arXiv:2404.07965, 2024.

Questions

1. In Figure 7, I am wondering why the to-learn accuracy drops after 4-th epoch? It is kind of counter-intuitive to me as I expect it will grow along with increasing epoch. 2. Could the authors provide more analysis and rationale behind the phenomena that K-adapter keeps its not-to-forget accuracy basically unchanged?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors addressed the limitations adequately.

Summary

This paper studies the continual knowledge learning (CKL) problem in large language models. The authors notice that the existing methods either apply equal weights to all tokens or re-weight tokens with a trivial consideration that tokens with low confidence are important. They propose a new definition of token importance based on the expected functionality of the token in solving related tasks. In particular, they design a meta-learning strategy to learn token weights in the training phase to facilitate the continual learning of new knowledge while retaining already learned knowledge. A new benchmark, LAMA-CKL, is established. The experimental results demonstrate that the proposed method outperforms other methods in both existing and introduced benchmarks.

Strengths

1. Consider the expected functionality of tokens in solving related tasks as token importance is novel and well-motivated. 2. The proposed meta-learning framework is simple yet effective and can be complementary to other methods. 3. A new benchmark for evaluating continual knowledge learning is established.

Weaknesses

1. Naming token importance as "usefulness" is too broad and does not accurately reflect the motivation of the paper. 2. Meta learning is a well-known method, the description of the algorithm details in Section 3.2 is however confusing and unclear. 3. The proposed method seems to only be applicable to the experimental setup described in the paper. This setup requires pairs of \mathcal{D} and \mathcal{T}_\mathcal{D}, where \mathcal{T}_\mathcal{D} represents a task that can be solved using the information contained in \mathcal{D}. This setup is not general enough and is limited to very specific scenarios.

Questions

Why does the proposed meta learning approach reach performance peaks in fewer epochs compared to other methods? Is there an explanation for this?

Rating

6

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

The discussion of limitations is included in the paper.

Summary

The article introduces Train-Attention Augmented Language Model (TAALM), a novel approach for continual knowledge learning in large language models. TAALM dynamically assigns weights to tokens based on their usefulness, optimizing learning efficiency and minimizing forgetting.

Strengths

- This paper introduces Train-Attention Augmented Language Model (TAALM) to reweight token sequences in tuning LLM to avoid catastrophic forgetting. - It introduces a new benchmark, LAMA-CKL, for clearer learning-retention trade-off assessment. - TAALM demonstrates state-of-the-art performance, compatibility with existing methods, and computational efficiency, advancing the field of CKL.

Weaknesses

- The proposed method is not novel and it is widely used in multi-task learning to adjust task loss weights in the learning process, like "MetaWeighting: Learning to Weight Tasks in Multi-Task Learning". - This paper lacks of explanation of why it could maintain previous knowledge without forgetting.

Questions

- What will be the influence of dropping the lightweight token sequences in training? If only important token sequences matter in learning new knowledge. - Can authors offer more theoretical or experimental explanations for the effectiveness of the trained attention? - I was confused about the relation of the learned meta weight with continual learning. Why dropping the unuseful tokens will help the model learn without forgetting? - The ablation study of the learned meta-weight is missing. I only found the results in Figure 6, however the effectiveness of the learned weight should be further validated.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

N/A

Summary

The paper introduces Train-Attention-Augmented Language Model (TAALM), a novel approach to continual knowledge learning (CKL) in large language models (LLMs). Unlike traditional methods that uniformly apply weight across all tokens, TAALM dynamically predicts and applies weights to tokens based on their importance using a meta-learning framework. This approach aims to enhance learning efficiency by targeting essential knowledge updates and minimizing forgetting. The authors also introduce a new benchmark, LAMA-CKL, to better assess the trade-off between learning new information and retaining existing knowledge. Experimental results show that TAALM significantly outperforms existing CKL methods on both new and established benchmarks.

Strengths

Originality: The paper introduces a novel approach to CKL by using meta-learning to predict token importance, which is a significant departure from traditional methods. Quality: The technical claims are well-supported by both theoretical justifications and empirical evidence. The experiments are comprehensive and rigorously conducted. Clarity: The paper is well-structured and clearly written, with detailed explanations of the methodology and results. Significance: The contributions are substantial, offering both a new method (TAALM) and a new benchmark (LAMA-CKL) that advance the state-of-the-art in CKL research.

Weaknesses

Generalization to Other Tasks: While the paper demonstrates the effectiveness of TAALM in the context of CKL, it would be beneficial to explore its applicability to other types of continual learning tasks beyond language models. Computational Resources: The training of TAALM, especially with large models, requires significant computational resources. A discussion on the scalability and efficiency of the approach for smaller models or resource-constrained environments would be useful. Ablation Studies: While the paper includes comprehensive experiments, additional ablation studies to isolate the impact of different components of TAALM (e.g., the specific meta-learning algorithm used) would strengthen the evaluation.

Questions

Can the authors provide more details on the potential applicability of TAALM to other types of continual learning tasks outside of language models? How does TAALM perform with significantly smaller models or in environments with limited computational resources? Could the authors include more detailed ablation studies to better understand the contribution of each component of TAALM to the overall performance?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

The paper adequately discusses the limitations of the proposed approach, particularly the task-specific nature of Train-Attention and the requirement for data-task pairs for training. The authors also highlight the potential for Train-Attention to evolve and adapt to new tasks, suggesting areas for future exploration. However, a more detailed discussion on the computational requirements and scalability of TAALM would be beneficial.

Response to Rebuttal by Authors

Thanks for your detailed response to my questions. I am impressed with the results you achieved with both larger and smaller pre-trained models. I would like to keep my score.

Respond

Many thanks to the author for the responses. They address my concerns and I will maintain a positive score.

Discussion

Dear Reviewers, Thanks for the reviews. The authors have uploaded their responses to your comments, please check if the rebuttal address your concerns and if you have further questions/comments to discuss with the authors. If the authors have addressed your concerns, please adjust your rating accordingly or vice versa. Regards, AC

Sincere thanks

We thank again for your constructive review. We hope that our responses have adequately addressed your previous concerns, and we would be happy to address any further questions/suggestions that might come up until the end of the discussion period.