Paper ID: | 1224 |
---|---|

Title: | Rethinking Generative Mode Coverage: A Pointwise Guaranteed Approach |

This paper proposes to use a mixture of generators to solve the mode collapse problem. Specifically, for a certain generator, it mainly focuses on the training data which is not covered by previously trained generators. By training a sequence of generators, this paper provides a theoretical analysis that the mixture of generators can cover the whole data distribution. There are three concerns. The first concern lies in the second paragraph. It claims that most divergences are global and are not ideal for promoting mode coverage. It lacks supports, such as related works and experimental results. Besides, several existing works claims that KL divergence can largely solve the mode collapse problem [1][2]. I think these related works should be discussed in the main part. The second part is about speed and scalability. Since the mixture of generators is trained sequentially, what is the time complexity compared to training a simple DCGAN for all these experimental settings, such as in fashion-mnist and toy-data. The third concern is about the experimental setting. For the first toy data in Figure 2, whatâ€™s the weight of the minor mode compared to the major mode? In real-world application, will this mode just be treated as noise? On the other hand, I am curious about the likelihood of samples in the minor mode estimated using the kernel density estimator with sufficient enough samples and sufficient small variance. Besides, this paper does not provide any large scale results, such as cifar10 and celebA. [1] Nguyen, et al. "Dual discriminator generative adversarial nets." Advances in Neural Information Processing Systems. 2017. [2] Du, et al. "Learning Implicit Generative Models by Teaching Explicit Ones." arXiv preprint arXiv:1807.03870. # Post-rebuttal comment I have mixed feelings for this paper. On one hand, this paper proposes a new method (i.e., the point-wise convergence) to define the mode collapse problem, which is novel and interesting. And a corresponding algorithm is provided to train a mixture of generators to address the mode collapse problem. On the other hand, the experiments are the main limitation: no experiment on large scale natural images is provided, such as the Cifar10 and celebA. Indeed, the authors address most of my concerns. However, the claim of the Figure 6 in Appendix B is incorrect. If we use the KL divergence, i.e., KL(P||G), rather than the inverse KL, i.e., KL(G||P), the green distribution will be chosen rather than the red one. According to my numerical results, KL(P||N(0, 1)) = 4.6, whereas KL(P||0.25N(-10, 1) + 0.5N(0, 1) + 0.25N(10, 1)) = 0.36. It verifies the zero-avoiding properties of KL divergence[2]. To summarize, I still have an inclination to reject this paper in its current state, because of the limitation of experiments and incorrect claims for the KL divergence. However, based on its novel contribution, I will also be happy if this paper will be accepted.

This paper considers the problem of learning a generator mixture to capture all the data modes under an interesting formulation. Different from existing work which is mainly based on global statistical distance, this work focuses on the local pointwise coverage which also guarantees of the global distance in a loose way. Based on the multiplicative weights update algorithm, the authors proposed an algorithm to learn the generator mixture and theoretically proved its complete mode coverage guarantee. The proposed method works well on both synthetic and real datasets, achieving better mode coverage compared to existing methods while maintaining good learning of the global distribution. On the other hand, the manuscript is not very easy to follow, especially Section 1.1 and Section 3. The introduction of the (\delta,\beta)-cover is not very intuitive even though it would eventually become clear in Section 3.2. Some intuitive discussion of the motivations might be helpful. The proposed method, Algorithm 1, has a flavor of the non-adaptive boosting algorithm where the weight of data points would be reduced if we got thing right and an simple average of the individual weak leaners is outputted. The existing boosted density estimation methods, additive (AdaGAN) or multiplicative ([1,2]), all have a flavor of the adaptive boosting algorithm. It is interesting to see if the adaptive version has certain guarantee in terms of the introduced mode coverage or examples of the failure cases. The empirical results suggest that the proposed method performs better than AdaGAN in terms of both mode coverage (in Table 1) and global distribution matching (in Figure 5). But AdaGAN is directly trying to minimize the global statistical distance while the proposed method only has a loose guarantee in the global sense, it would be better if the authors could provide some insights here. The experiments are relatively weak. The authors proposed a mode coverage metric and checked the global distribution matching all based on trained classifiers which is somewhat ad-hoc, and it would be good to also look at the coverage metric introduced in the AdaGAN work. The precision and recall metrics introduced in [3] are another interesting candidates. There is only mode coverage results for a fixed T value. Plots of the coverage metric against T might shed more light about the learning process. Another question is about the mode bias measure introduced towards the end of Section 4. It could be computed in Figure 4 mainly because the mode itself is known, but it is unclear how to compute this measure in practice where the modes are generally unknown beforehand. [1] Grover, A. and Ermon, S. Boosted generative models. In AAAI, 2018. [2] Cranko, Z., and Nock, R. Boosted density estimation remastered. In ICML, 2019. [3] Sajjadi, M.S.M., Bachem, O., Lucic, M., Bousquet, O., and Gelly, S. Assessing generative models via precision and recall. In NeurIPS, 2018. After rebuttal: Thanks for the detailed rebuttal. I do not fully agree with the argument provided for why AdaGAN can not guarantee full mode coverage. "In the next iteration, the weight of the major 25 mode may still be large, and thus prevents AdaGAN from learning the minor mode". AdaGAN would also diminish the weight for the well-fitted part in an exponential way similar to the proposed method. As pointed out by other reviewers, the quantitative results are limited, especially on real data. But I would maintain my score as 7 since I do think the theoretical contribution is novel and interesting.

The paper describes a new algorithm for training a mixture of generator that can, with some theoretical guarantee, capture multiple modes in the data distribution. One of the main contributions of the paper is a new measure of distance between distributions that explicitly considers point-wise coverage (rather than the traditional global notions of coverage, e.g. f-divergence), optimizing for this metric allows one to construct mixtures that aim to capture multiple modes (even minor ones) in the target distribution. It appears that existing work training generators with modified objective functions and those that ensemble multiple generators do not explicitly train for this type of "point-wise" or "local" coverage, but rather train for some proxy that may indirectly encourage coverage of multiple modes. In this light, the contribution of both the distance measure and the algorithm is significant. Furthermore, the proposed algorithm comes with theoretical guarantees of certain degrees of point-wise coverage (given appropriate choices of the number of expressive generators in the mixtures and a covering threshold). This is valuable as the author demonstrate that theoretical analysis lends insight to how to choose hyper parameters in practice. Finally, the paper presents evidence, in cases where the number of ground truth modes are known, for the effectiveness of the proposed method on synthetic and image datasets, and comparison results with respect to other methods that aim to capture multiple modes. Overall the paper is very well written. The technical exposition is clear and easy to follow. The theoretical analysis is thorough and the central idea is intuitive and appealing.