NeurIPS 2020

GAN Memory with No Forgetting

Review 1

Summary and Contributions: Update: Having seen all the other reviews and the author responses, I continue to support acceptance of this paper. This paper presents a replay-based approach for lifelong/incremental learning. In particular, it leverages a generative model to produce samples corresponding to previous tasks/data when adapting the learning framework to a new task/data. A GAN model is used as the framework for generating data, which in itself is "modulated" as and when new data is seen, inspired by style-transfer techniques. Two techniques, namely FiLM and AdaFM, are used to achieve this modulation step, which allow a base GAN model to adapt. This approach is evaluated on the image classification task, and compared with related work.

Strengths: * The paper presents an interesting idea for incremental learning. The idea of adapting the generative model also incrementally is worth noting. * The approach is reasonaly well explained, barring a few details that are missing in the main paper (see below). * The approach is evaluated at different stages, including the data generation steps, with several ablation studies. * Source code (which I have not had a chance to test) is included along with the submission.

Weaknesses: * At several points the paper refers to a well-behaved GAN. This remains vague. How does the initial GAN model trained on CelebA become a well-behaved GAN? What is the impact of learning the initial GAN model on other datasets? How sensitive does the adaptation process become in this case? There are some of the questions that need clarification in the paper. * Section 4.2 defines properties of GAN memory. I do wonder how dependent some of these properties are on the initial "well-behaved" GAN model? Especially, how generic are the observations made about FC to B6 blocks. On a related note, it may be useful to show these blocks in a schematic figure, at least in the supplementary material. * An important comparison in the context of incremental learning is showing the upper bound, i.e., the performance when the model is trained on all the tasks/data jointly. This needs to be in the paper.

Correctness: Some of the claims need clarity (see above).

Clarity: More or less, but some technical details are missing in the main paper (see additional comments below).

Relation to Prior Work: Yes

Reproducibility: No

Additional Feedback: * Important technical details do need to be in the main paper. For example, the approach and the protocol for incremental learning with replay should be in the main paper, at least briefly. The current level of detail for this is insufficient (some of this is in the supplementary material, but it is really needed in the main paper). * Another aspect I was wondering about is the impact related tasks/data have on the generative model. For example, if the sequential tasks were based on data from cat, dog, horse, and cow classes, how would the generated samples be? Is the model robust to such data? * On a philosophical level, should the transfer process be treated as content transfer instead of style? I would associate style with things like a painting, or filters. What is changing here, atleast in the examples (Celeb -> flowers -> buildings -> cat) is the content of the image. * Other comments: - The paper needs to be improved in terms of writing. There are also several typos and grammatical errors. On a related note, the abstract of the paper is very cryptic. - The availability of source code should be mentioned in the paper.

Review 2

Summary and Contributions: This paper adapts feature modulation to continually learn new image generation tasks by only learning new style modulation parameters per task. The backbone GAN remains fixed, which results in no forgetting. The authors adapt FiLM and AdaFM to the continual learning setting. To save memory in the new parameters, they can be compressed with a low rank decomposition. The method is evaluated for continual image generation and classification in a challenging setting with six tasks.

Strengths: - No forgetting by design, while having high quality in generated images. - The approach is sound, and the evaluation too. - The method is scalable to many tasks and efficient (relatively few parameters per task). - The paper provides interesting insights on how style parameters adapt the image generator to other domains.

Weaknesses: - Heavily depends on a powerful source model, which in turn depends on the source domain. - The capacity grows every new task (in contrast to MerGAN, although it requires an additional copy of the generator). - The technical novelty still remains largely incremental. Using style modulation to adapt GANs to other domains was already proposed in BSA[54] and AdaFM. Similarly, mFiLM and mAdaFM are minor tweaks of FiLM and AdaFM.

Correctness: Yes, they are correct.

Clarity: Yes, easy to follow and understand.

Relation to Prior Work: Yes.

Reproducibility: Yes

Additional Feedback: For better understanding, please include in Table 1 the number of new parameters without compression, and the total number of parameters (including the fixed source model).

Review 3

Summary and Contributions: The paper suggests a way to train GANs on multiple datasets sequentially (lifelong learning) using a modulation trick from style transfer literature. They demonstrate that it is possible to reuse the same frozen GAN to generate very different data like faces, flowers and even brain scans. It opens an opportunity to use such a model for generative replay in lifelong learning scenarios. Because the main model is frozen it demonstrates no forgetting. However, it comes at a cost of added style parameters that can take about 20% of the main parameter count. The authors also suggest a trick to compress them further.

Strengths: It is an interesting work that presents a counter-intuitive discovery that GANs seem to learn very general principles of image decomposition which can be reused to completely different datasets with style transfer techniques. It is especially interesting because usually style transfer can't alter image so much to make a brain scan out of a flower.

Weaknesses: Evaluation can be stronger. While the authors demonstrate that GAN can learn multiple datasets, but there are no experiments measuring how well they do this compared to GANs learning these datasets from scratch. I consider this an important part of any lifelong learning paper because very few would want to use tricky generative replay techniques if they underperform significantly. After all compute is getting cheaper every day. The authors claim that their model can be used on streaming datasets (L191), but don't explain it well. In my opinion streaming dataset explicitly doesn't have task boundaries and may manifest very smooth variation of images. It is difficult to capture it with task-specific parameters. Perhaps it is worth mentioning what scenario they had in mind.

Correctness: I believe the method and the claims are correct.

Clarity: The paper is well-written and easy to read.

Relation to Prior Work: The authors compare their work with MeRGAN, relatively recent memory replay GAN, and demonstrate better performance. Related work can be a bit longer, because it is a well-studied task after all. In general lifelong learning is a problem for classification networks, even more so for generative ones.

Reproducibility: Yes

Additional Feedback: In L272 there is a mention that Figure 5 compares the suggested method with MeRGAN, but the Figure doesn't have MeRGAN samples. I suggest also to find a catchy name for their work. It is much easier to cite it in the text if it is *-GAN or something like that. The exact type of GAN used in this work is mentioned only in supplementary. I consider it quite important info, there is no such thing as "just a GAN" nowadays. POST REBUTTAL: I found the rebuttal satisfying and decided to keep positive rating.

Review 4

Summary and Contributions: This paper proposes to learn a GAN on several dataset sequentially by affine transforming its parameters. The main contributions are: 1) Demonstrate that by affine transforming the parameters of a pre-trained GAN, one can adapt it to new domains. 2) Utilize 1) to learn a GAN on several dataset of different image classes sequentially. 3) Propose a method to compress the parameters in the affine transformation process.

Strengths: The paper is well written and easy to follow. The idea of affine transforming the parameters of a network so that it can be adapted to new domains is interesting. The paper includes detailed analysis and experiments for the affine transformation. The results look promising compared to baselines.

Weaknesses: The term style transfer in the paper is misleading. Style transfer refers to the specific task of transferring the style of an image to another image, existing methods achieve so by normalizing and transforming the intermediate features. However, in this paper, the GAN is adapted to the new domain by affine transformation on its parameters rather than learned features. So the concept of style transfer does not seem fit here. Also, it is not clear the major difference of this paper compared to [95]. The core novelty of this paper lies in equation (2) and (3), which are also explored by [95]. What's the major novelty in method compared to [95]?

Correctness: Yes.

Clarity: Yes.

Relation to Prior Work: Please see the weakness section, that this work may add more discussion of its novelty on methods compared to [95].

Reproducibility: Yes

Additional Feedback: Generally the paper is well written, but some parts could be a bit confusing. For instance, what's the definition of a well-behaved GAN? Also, line 114 says W_hat is used to convolve with input feature maps, but in Eq. (3) it is used to convolve with the Conv filter. The citation of [95] misses publication conference.