Dealing with Synthetic Data Contamination in Online Continual Learning

Summary

This paper investigates the impact of AI-generated images on the performance of online continual learning (CL) models. It introduces a novel method called Entropy Selection with Real-synthetic similarity Maximization (ESRM) to mitigate the negative effects of synthetic data contamination. ESRM leverages entropy-based sample selection and a contrastive learning approach to align the feature embeddings of real and synthetic data, thereby enhancing the robustness of CL models against the degradation caused by synthetic data.

Strengths

- The authors clearly articulate the problem, methodology, and results, enhancing the paper's accessibility and understanding for a broad readership. - The paper uniquely identifies the issue of synthetic data contamination in online continual learning, a significant challenge for the future of this field. The work has substantial implications for the ML community, offering a pioneering approach to maintaining the integrity of continual learning models in the presence of synthetic data. - This paper proposes ESRM, an innovative method that combines entropy selection and contrastive learning to mitigate the negative effects of synthetic data, demonstrating creativity in addressing this new problem. - The paper is underpinned by robust technical approaches, ensuring the quality and reliability of the proposed solution through comprehensive experimental validation.

Weaknesses

- The paper evaluates the impact of synthetic data using a limited set of generative models. And the experiments are primarily conducted on image classification datasets. Expanding this to include a broader range of tasks, particularly truly text-to-images, text-to-text, could strengthen the findings. - The method for generating synthetic datasets is straightforward, using simple prompts. Incorporating more complex and diverse prompts, potentially using large language models to simulate user queries, could better reflect real-world synthetic data contamination. - The reliance on entropy as a key metric for distinguishing real from synthetic data might not be universally applicable. Further exploration of this metric's effectiveness across different domains and its theoretical underpinnings could bolster the method's credibility. - While the paper provides insights into the method's effectiveness, there is limited discussion on its computational efficiency and scalability, especially when dealing with large-scale datasets or high contamination ratios. Addressing these aspects could be crucial for practical applications where computational resources and time are critical constraints.

Questions

Please refer to the weaknesses above.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have discussed the limitations in the paper.

Summary

Image generation has been showing promising achievements in the last few years, but the generative models may not be able to keep up with the distribution of the real samples. Due to low diversity, synthetic images perform poorly in downstream learning tasks. This paper tackles this problem by diversifying the sample selection using the proposed ESRM framework, which prioritizes the high-entropy samples in the experience replay memory with an RM loss function. The authors first performed experimental analysis to build their motivation in section 4 and proposed their solution to the problem in section 5. The experiments are diversified with several datasets, benchmarks, and extensive ablation studies.

Strengths

1. The paper is nicely written, with clear language and sufficient details. 3. Section 4's experimental analysis of existing CL methods clearly motivates the proposed model. 4. The authors propose a novel loss function to address the issue of low-entropy samples in CL. This approach builds upon existing supervised constructive loss techniques. Their method aims to maximize the cosine similarity between embeddings of real and synthetic samples while simultaneously minimizing the impact of low-entropy samples on the learning process. 4. In the ablation study, the authors studied different components of the proposed framework that show the superiority of the overall ESRM.

Weaknesses

1. The paper only deals with diffusion generative models. I understand that diffusion models are the most powerful image generators but it would be good to experiment other generative models. 2. The accuracy scores are overall low in CL. Does the proposed method 3. Why is a fixed 50% dropping rate selected for ES? 4. Downstream ML tasks have been explored with synthetic data to validate synthetic datasets, and related works clearly miss that angle. The authors should also discuss them and maybe add some of them to benchmarks.

Questions

Authors can refer to the weaknesses section for major questions. Here, I listed minor points for authors reference: 1. It seems there is a typo in the line 231. 2. It could be good to highlight the best results in Tables 3 and 4, the same as Table 2.

Rating

7

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

I agree with the limitation the authors brought up.

Summary

This paper investigates the negative impact of synthetic data contamination on existing online continual learning methods. An entropy selection with the real-synthetic similarity maximization method is proposed to alleviate the performance deterioration.

Strengths

1. Detailed analysis of synthetic data contamination and its influence on continual learning. 2. This paper is technically clear and easy to follow.

Weaknesses

1. The creation of simulated data is the cornerstone of this research, but there is a lack of detailed explanation on how to generate these data in the main paper. 2. For Observation 4 "With the limited diversity of synthetic data", how about mixing the synthetic data from different generation models to increase the diversity, since the synthetic images on the internet also form different models? While the observation remains unchanged in this case? 3. In order to simulate real data collection more realistically, in addition to synthetic data, new open-domain real data should also be incorporated.

Questions

Please refer to the weaknesses.

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

The limitations are discussed while the potential negative societal impact is not discussed. However, for this work, I think it is not necessary to discuss this.

Thank you for your rebuttal

Thank you for the rebuttal. This paper is very insteresting and promising. After considering the comments from other reviewers and the rebuttal, I decided to raise my score.

Thanks for the rebuttal.

The authors have addressed most of my concerns. I decide to keep my initial positive rating.