NeurIPS 2020

Robust Disentanglement of a Few Factors at a Time using rPU-VAE

Review 1

Summary and Contributions: This work provides a novel disentanglement method by recursively learning to account for different variations in the dataset.

Strengths: - The intuition and implementation of the recursive strategy seem sound. - Figure 5 is very convincing that different factors are indeed learned sequentially - Empirical results (such as Figure 4) are convincing, with the low variance quite desirable.

Weaknesses: - In Figure 2, the size of the sprite often varies along with the xy position, which suggests lack of disentanglement to me. - In Figure 1, the right panel has a method simply labeled [14] with no name. - S2.2, what kind of object "eval" and "step" are is never explicitly defined. - The font of Figure 4 is very hard to read - While I know that the multi-image panels such as in Figure 5 are common for this field, these are a bit small to see accurately: even digitally but especially in print. I would suggest reducing the number of images, or stacking the pajnels vertically to increase the width. Very Minor/Finnicky: L152, present tense is acceptable for describing experiments, but "We now test" is a bit too narrative.

Correctness: To the best of my knowledge, but I am not very knowledgeable about disentanglement. From a high-level standpoint, the concept of iteratively reducing a problem is akin to some fundamental methods in linear algebra or other fields. Since the classification of the factors is a well-posed problem given the deterministic nature of the dataset, that implies that with s

Clarity: Overall yes. Despite my unfamiliarity with the field, it was quite easy to follow. That said, I think that neither the text and visual presentations are well-polished. The authors communicate well, but the presentation is relatively lacking. Specifically, figures are fairly consistently undersized and code variables in the text are often confusing.

Relation to Prior Work: Yes. I believe S2 provides a good overview. I would suggest a short paragraph on similar recursion-based methods. Gram-Schmidt is a classical example that is very analogous .

Reproducibility: Yes

Additional Feedback: Capitalization on section headers is inconsistent In Figure 4 (particularly 4b), the skew of the distributions seems consistent across other methods, but markedly different for your methods (Besides just low variance). Do you have any insight for why this might be? To summarize: I think that this paper does seem to present a new SotA and explains it well. The communication is effective, but not well-polished from a visual standpoint. The method is fairly complex, but the intuition is simple. =============== Update after rebuttal: My concerns re: presentation were addressed as expected; it's hard for an author to fully assuage concerns like that in a 1-page rebuttal to multiple reviewers. Seeing other reviews and discussion had made me appreciate the results and method even more. I'm tentative in my assessment due to my unfamiliarity with the field but I'm increasing my rating up to a 7 and recommending acceptance for this paper.

Review 2

Summary and Contributions: **** After author feedback **** Having read the other reviews and the author feedback, I feel very happy about the paper. I am going to keep my score but will also happily argue for the paper to be accepted. ****** This paper proposes a novel approach for robust training of disentangled VAE-based approaches in an unsupervised manner, thus addressing one of the biggest problems with the existing methods for unsupervised disentangled representation learning. The approach involves a recurrent process which trains beta-TCVAE models to disentangle a subset of dimensions at a time. The discovered dimensions are then removed from the dataset through an automatic process, and another set of beta-TCVAE models are trained to discover some of the remaining disentangled dimensions. This process is repeated until all disentangled dimensions are discovered. Each iteration uses PBT for hyper parameter tuning, combined with the recently proposed unsupervised UDR score for model evaluation.

Strengths: The authors are able to demonstrate convincing results on two datasets (dSprites and Shapes3D), achieving high disentanglement and low variance in final model performances. The paper is very well written and systematically evaluates how each step of the proposed algorithm performs. The main strength of this paper IMO is that it proposes a pipeline that addresses one of the biggest shortcomings of unsupervised VAE-based disentangling approaches - that of high variability in the quality of the learnt representation even when trained with the same hyperparameters, and the fact that these approaches often tend to learn only a subset of the generative facts. These two limitations create a problem for applying unsupervised disentangled representation learning in practice. By proposing an automated pipleline that removes these limitations, this paper is likely to pave way for a large body of applied disentanglement work.

Weaknesses: The weakness of this paper is that it doesn't necessarily create any theoretical contributions to the field. It describes a highly engineered approach instead, and this approach relies on a lot of computing power (due to relying to training numerous models through PBT+UDR). This, however, should not affect the chances of this paper of being accepted to NeurIPS, since I believe it will still be of high interest to a wide range of practitioners interested in applying disentangled representation learning to their work.

Correctness: Yes

Clarity: Yes

Relation to Prior Work: Yes

Reproducibility: Yes

Additional Feedback: Given that this paper is primarily engineering/experimental, I would have liked to see additional results to verify the success of the proposed pipeline: 1. Applying this method to more complex datasets that don't necessarily have good labels for semi-supervised disentanglement (at least CelebA) 3. Latent traversals for the final models trained to disentangle all the generative factors (on dSprites and Shapes3D, as well as the more complex dataset suggested above).

Review 3

Summary and Contributions: The paper explores using population-based training (PBT) for improving disentangled VAEs. The paper shows that applying PBT with (semi-)supervised evaluation function significantly reduces the variance of the results. In addition, the paper proposes an unsupervised approach by incorporating UDR as the evaluation function. To further improve the result, the paper proposes rPU-VAE which recursively labels the dataset with UDR and finally uses the labels to conduct another iteration of PBT-S-VAE. The result shows that this unsupervised training approach achieves better disentanglement performance than some baselines.

Strengths: Although PBT is not a new method, demonstrating the effectiveness of PBT on training disentangled VAEs is still valuable. The most interesting part is the proposed recursive factor removing algorithm. Beyond the evaluation in the paper, I can see that this idea might be useful for disentangling more challenging datasets (e.g. celeba) where it is hard to disentangle all factors by one training pass.

Weaknesses: I have some doubts about the evaluation of the (semi-)supervised part. The presentation of some important details is also obscure. This makes it hard for me to judge the soundness of the algorithm and evaluation. Please explain these details in the rebuttal and I'll adjust my score accordingly. - Section 3.2: I don't fully understand how you train "supervised" and "semi-supervised" PBT-S-VAE. Do you mean that in the supervised setting, you use *all* labeled images to evaluate MIG, whereas in the semi-supervised setting, you use only 1000 labeled images to evaluate MIG? If that's the case, I think the comparison in Figure 1 is unfair and insufficient, because beta-TCVAE doesn't use any labeled data, but (1000-)PBT-S-VAE uses. Besides comparing with [15], a more natural baseline would be to train a set of beta-TCVAE models and then use the labeled datasets (either the entire dataset or 1000 samples) to do a supervised model selection, as PBT implicitly does this model selection as well. The score of the beta-TCVAE with supervised model selection will definitely be higher, and I would imagine that the variance will also be much smaller. I am not sure how much benefit PBT has in this (semi-)supervised setting. - Line 113: "beta-VAE metric as well as the FactorVAE require the true underlying generative model and are therefore not suited". What do you mean by "underlying generative model"? If you mean a set of images with ground truth labels, I think MIG and DCI also need that. In particular, the publicly released dSprites dataset contains all the information you need for computing MIG, DCI, beta-VAE, and FactorVAE score. - The description around line 240 about leaf-run is hard to understand. From the description, it seems like in each metaEpoch, you do multiple leaf-runs. But Algorithm 1 shows that leaf-run is in the outer loop. I don't understand what leaf-run is and what's its purpose. Also, how do you intersect the intervals of z_active? Which intervals are intersected? What is "n"? How does each metaEpoch generate more surrogate labels as you are doing argmax in line 207? - Figure 3: what's the meaning of latent-encoding-value? I don't think the latent codes of VAEs can be such large. Minor issues: - Line 130: what do you mean by "upper limit"? - Figure 1(b): [14] -> [15]

Correctness: Generally yes.

Clarity: Generally yes.

Relation to Prior Work: Yes.

Reproducibility: Yes

Additional Feedback: ----- Thank the authors for the detailed response! The response answers my questions clearly, and therefore I increase the score. However, I would recommend the authors to polish the writing regarding these questions in the revison. Especially, the intention of (semi-)supervised experiments should be highlighted better. The current writing and experimental comparisons are misleading. ----- I suggest the author clarify the above points in the rebuttal.

Review 4

Summary and Contributions: In this paper, the authors propose population based training for variational training with a goal of achieving more consistent disentanglement of factors. To validate the effectiveness of the proposed method, the authors compare the extent of disentanglement for different datasets and demonstrate improved performance over the baselines. ================================================================Update after rebuttal========================================================== In my initial review, I felt the work is incremental because it is a direct application of population based training to variational training and the authors failed to provide enough evidence supporting stability of the proposed approach and the evaluation was mostly on simple datasets. After reading the author response, I am convinced of the effectiveness and the stability of the proposed method based on the model performance on CelebA datasets provided in the author response. Additionally, the authors try to address the questions as much as they could in a one page rebuttal. Based on the author response, I would like the paper to get accepted (trusting that the authors would make clarity based revisions promised in the author response). Therefore, I have increased my score from 4 to 6.

Strengths: - The paper presents the idea of population based training for VAE in multiple settings, i.e., supervised, semi-supervised, and unsupervised. The evolution of the model and highlighting the challenges and effectiveness of each settings is interesting. - The empirical evaluation highlights the potential of population based training for disentanglement in unsupervised learning.

Weaknesses: I think the contribution of the work is incremental. The model simply applies population based training to variational training. One interesting aspect of the paper is to address hyperparameter sensitivity issue. While the authors also claim hyperparameter sensitivity as one of the goals that they wish to achieve, the paper fails to provide thorough analysis and study of this aspect. The authors only briefly highlight of in Supplementary but I would like to see more analysis. These are the two main reasons I think the paper falls short of NeurIPS standards. Details oriented questions: - The motivation of rPU-VAE, i.e., recurrent version of the model in unsupervised setting is not entirely clear to me. Given that the major improvement is in the first run, what is the model achieving in the future runs? - The model description also lacks explanation on what is formal definition of active latent factors from each run? Are they determined based on variations in the latent factors? Also, discussion in the results highlighting what the model actually learns as active latent factors. - Typo: plot in Figure 1(b): [14] -> [15]

Correctness: The claims and method are technically sound.

Clarity: I think the paper needs some more clarifications on the various variants of the population based training models presented in the paper.

Relation to Prior Work: Prior work has been discussed clearly.

Reproducibility: Yes

Additional Feedback: