Dealing with Synthetic Data Contamination in Online Continual Learning

Thank you for your feedback and valuable comments.

Weaknesses

1. We indeed tested our method with a limited set of generative models, due to the computation constraint. Also, it is true that almost all of the existing work in online Continual Learning (CL) is limited to Class-Incremental Learning (CIL) scenarios, where the major task is image classification. We fully agree the scope of the tasks can be beneficial, which we take as an important direction for future work.

2. Following the suggestion, we conducted extra rebuttal experiments as shown in the general rebuttal (1). We use LLaMA to generate more diverse and complex prompts to simulate a more realistic contamination scenario. The experimental result in Table I in the rebuttal material validates the effectiveness of our method when the synthetic images are generated with LLM-enhanced prompts.

3. The entropy criteria is indeed important in our work, and we would like to prove the universality of the entropy metric both empirically and theoretically. Empirically, we expand the experiments with broader settings. The Domain-Incremental Learning setting and LLM enhancement setting in Table I of the rebuttal material further proved the universality of ESRM. Theoretically, classifiers trained on a limited diversity dataset often overfit, resulting in confident predictions on the training data but poor generalization to new data. Such a pattern is universal and not limited to synthetic contamination. Methods like Label Smoothing [4] and Confidence Penalty [5] are proposed to alleviate such problems. In our paper, this limited diversity issue is magnified by synthetic contamination, where the generative model's output diversity remains imperfect for current AI research. With limited diversity, synthetic data are easier for the model's feature extractor to cluster (cf. Fig. 3 in the paper) and for the classifier to classify, leading to more confident (lower entropy) predictions (cf. Fig. 2 in our paper). We will include such experiments and theoretical discussions to further improve our manuscript.

4. We regard computation efficiency and scalability as crucial criteria for our method, especially in online continual learning scenarios. Thus, we included the training time on the CIFAR-100 dataset in Sec. D.7 and Fig. 11 in the appendix. Following the suggestion, to show how the computation of different baselines is scaled with larger datasets, in general rebuttal (4), we include the training time of all baselines on TinyImageNet and ImageNet-100 datasets. As shown in Fig. II of the rebuttal material, the training time of our method is significantly faster than some state-of-the-art methods like OCM and OnPro, while it is on par with the most efficient method. Also, we would like to clarify that the training computation is agnostic with contamination ratio, because in the generation of the contaminated dataset, we replace the real images with synthetic images while not changing the size of the dataset.

Reference:

[4] "Rethinking the inception architecture for computer vision." CVPR 2016

[5] "Regularizing neural networks by penalizing confident output distributions." arXiv:1701.06548 (2017)

Thank you for your feedback and valuable comments.

Weaknesses

1. In our experiments, we prioritize the importance of diffusion-based contamination. There are two reasons: Firstly, training with synthetic images generated with GAN-based or autoencoder-based methods yields catastrophic performance, especially when the contamination is severe. Another reason is that we find GAN-based or autoencoder-based synthetic images are relatively easy to distinguish from their real counterparts, which hinders them from constituting a de facto “contamination”. For example, in our preliminary experiments, we tried to use GALIP [1], a state-of-the-art text-to-image GAN-based generative method to generate images with our text prompt, as shown in the general rebuttal (6). Fig. III in the rebuttal material shows the GALIP is subject to an extremely limited diversity problem. While we prioritize diffusion-based contamination, we regard using LLM-enhanced prompts to diversify GALIP-generated images as an important future direction.

2. In this work, we focus on the online Continual Learning (CL) setting specifically, where the synthetic contamination is a more factual and severe issue because the online setting restricts us from assessing the quality of the training data beforehand. Unfortunately, due to the online restrictions, there is still a salient performance gap between the performance of conventional CL and the state-of-the-art online CL. Also, it is noteworthy that ESRM could achieve a clean performance that is on par with the current state-of-the-art methods (OCM, GSA, and OnPro). We hope our research sheds light on the research field and helps narrow down the gap.

3. Intuitively and factually, the dropping rate in ES is a hyperparameter that is affected by the contamination ratio of the training dataset. There are two reasons to choose a fixed (50%) dropping rate. (a) We assume the continual learner does not have prior information about the contamination ratio of the dataset, which is more realistic in the real world. This is also the reason why we do not and should not perform a hyperparameter search for the dropping rate for each contamination ratio. (b) A drop rate of 50% gives competitive performance even under extreme conditions (Contamination ratio = 95%). We believe that the obtained performances demonstrate that a fixed ES dropping rate is resilient to different conditions (dataset, contamination ratio).

4. There is indeed excellent work on validating synthetic data, for example, UniFD [2], FatFormer [3], etc. We will include such discussions in the related work section in the revised manuscript. Experimentally, we replaced the ESRM's entropy selection strategy with the pretrained UniFD synthetic data detector, as illustrated in the general rebuttal (5). With our test, the accuracy of the UniFD detector is 66.19% on the C100/SDXL dataset. Table II in the rebuttal material shows that UniFD is more effective than random selection, but its performance is still limited. This is due to a distribution mismatch between our dataset and the dataset used to train the detectors. Moreover, due to the online constraint, we had to set the threshold of UniFD at 0.5, because we can not search the optimal parameter without the knowledge of the dataset distribution.

Questions

  Thank you for your suggestions, we will carefully revise the manuscript and update the tables for a better readability of our paper.

Reference:

[1] "GALIP: Generative adversarial clips for text-to-image synthesis." CVPR 2023

[2] "Towards universal fake image detectors that generalize across generative models." CVPR 2023

[3] "Forgery-aware adaptive transformer for generalizable synthetic image detection." CVPR 2024

Thank you for your feedback and valuable comments.

Weaknesses

1. We would like to include more detailed information on the synthetic dataset generation process. For Stable Diffusion and VQDM, we use source code and model snapshots from huggingface, as mentioned in Table 12 in the appendix. For Glide experiments, we use the official implementation and the released model snapshots. Following the recommendation, we use the refiner in Stable Diffusion XL and the upsampler in GLIDE. The diffusion steps and guidance scale hyperparameters we use are as follows:

   Generative Model	Diffusion Steps	Upsample(Refiner) Steps	Guidance Scale
   SD1.4	50	N/A	7.5
   SD2.1	50	N/A	7.5
   SDXL	40	40	5.0
   VQDM	100	N/A	7.5
   GLIDE	100	27	3.0

   For other hyperparameters, we follow the recommendations from Huggingface and GLIDE's official implementation. We use the prompt "An image of a {class_name}." as the text guidance to generate the image and interpolate the generated image to the size of the target dataset (32 for CIFAR, 64 for TinyImageNet, and 224 for ImageNet-100). We will include such detailed information in the revised manuscript to help the reproducibility of our paper. Also, for reproducibility, we will include the source code for synthetic dataset generation in our project codebase and make it publicly available.

2. As suggested in Sec. 3.1, we have two different strategies for simulating synthetic data contamination: (a) using data generated from SDXL only and (b) mixing the data generated from different generative models. In Table 3 of our paper, We have included the t-SNE visualization result on (a) setting to ground the observation 4. For (b) setting, besides the final average accuracy shown in Table 7 in the appendix, we include extra t-SNE visualization results in the rebuttal material, as introduced in the general rebuttal (3). The t-SNE visualization of feature representations given in Fig. I(a) of the rebuttal material shows that even when generating synthetic data with a mixture of generation models, a gap still exists in feature space between generated and non-generated data when training with ER. Such experiments additionally support the claims of Observation 4.

3. Following the suggestion, we extended our experiments from Class-Incremental Learning (CIL) only to include Domain-Incremental Learning (DIL) experiments on the DIL-CIFAR20 dataset. As mentioned in the general rebuttal (2), the DIL results of ESRM outperform other baselines, showing the robustness of ESRM against synthetic contamination.