NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:2434
Title:On Adversarial Mixup Resynthesis

Reviewer 1

This paper proposes a method for enhancing the latent space learned by auto-encoders, so that the learned latent space produces meaningful features useful for downstream tasks. The proposed approach considers interpolations in the latent space and encourages the reconstructions from these interpolations to be similar to the data using adversarial learning. The learned latent space is shown to capture useful feature via experiments on MNIST, KMNIST and SVHN. The paper presents some promising preliminary experiments. However, there are many issues in the experimental setup Why is the quality of features measured during training ? It is more sensible to learn the features, fix them, and then train the classifier on top of the features. Measuring the classifier accuracy during the training introduces some confounding factors, which makes me question the validity of the results. Why is one method not consistently better than the other ? For instance, in Table 1 mixup(3) is best on MNIST and mixup(2) is best on KMNIST and SVHN. It is unclear to me whether the baselines considered here are comprehensive, significant or strong enough. I appreciate the authors presenting some discussion about this in 211-213. I don’t understand the point of showing the reconstructions from the interpolation (Eg: section 4.3). Many prior work have demonstrated controlling attributes in image generation. Going beyond MNIST, SVHN and evaluating the approach on more complex/real-world datasets would make the paper more compelling. The paper serves as a preliminary exploration on some interesting ideas. However, the experiments need to be performed on real-world datasets against strong baselines with systematic evaluation to demonstrate the benefits of the approach. Questions Is the label based interpolation (section 2.2) used in the quantitative experiments ? Eq (7): Parts of the AC-GAN loss seem to be missing. Other remarks 64: The term ARAE has been used in Zhao et al., Adversarially Regularized Autoencoders. I would suggest using a different term. Make figures 3 and 4 bigger Eq (7) x1, y2 should be x1, y1 ? 151 - 160: I don’t know if the biological motivation needs to be emphasized

Reviewer 2

Significance (6/10) ------------------- Minor insights like the fact that Bernoulli mixup seems to perform worse than linear mixup can be gained from this paper. Based on the results, I don't expect widespread adoption of adversarial mixup as a regularizer but the paper will nevertheless be interesting to some. Originality (5/10) ------------------ The approach seems straightforward based on the ideas that exist in the literature. Quality (6/10) -------------- The quantitative experiments seem reasonable and well executed. What's missing are comparisons with further baselines / better benchmarks. For example, it is not clear from the paper whether I'd want to use any form of adversarial mixing to regularize my classifier compared to or in addition to other data augmentation techniques or dropout, for example. The interpolation results (Figure 1) are surprisingly poor, with not much semantic interpolation and a lot of the kind of ghosting artefacts expected from linear interpolation. The authors claim that a "fundamental difference" between mixup and VAEs is that they "impose no constraint, at least not in the probabilistic sense." I disagree with this statement. In the extreme case of picking each latent dimension from the representation of a different image, you'd be generating independent coefficients just like in a VAE, except you'd sample from an empirical distribution instead of a Gaussian. That is, Bernoulli mixup introduces statistical independence assumptions, even if the authors don't present them as such. It would enhance the paper if the authors could formalize a connection between Bernoulli mixup and nonlinear ICA, and perhaps also explore the probabilistic interpretation of linear mixup. Clarity (7/10) -------------- The paper is mostly well written and clear. Please explain how the parameters of p = embed(y) are trained. Since the image is dependent on y only through the binary mask m ~ p, and this Bernoulli sampling step is non-differentiable, it is not clear to me how this embedding is trained. The authors write that they "collect the highest accuracy on the validation set over the entire course of training". Please make explicit that you used separate validation and test sets to eliminate any doubt that the results are biased. Before Equation 7, the authors write that these losses are optimized "in addition to their unsupervised losses described in Equation 3". The "min_F" in front of the loss in Equation 7 suggests that only these two terms are optimized with respect to F. This should be clearer.

Reviewer 3

This paper proposed a new regularization method to learn better representations through auto-encoders. Intuitively, the proposed method encourages mix in the latent representations so that a combination of latent representations is able to produce realistic images (achieved through an adversarial loss). The high-level idea and formulation is similar to [1], but the proposed method differs in that it is not limited by the specific form of mixing functions, while [1] seems requiring a linear combination of two examples. As also mentioned in the paper, this is a difference between JSGAN and least squares GAN, and the proposed method is able to be freely combined with various mixing functions. In addition to the simplest mixup between two latent representations, this paper proposed and studied two additional mixing functions: (1). an element-wise mask and sum (which is like selection from features maps), and (2). Mixing with K examples (while the experiment only studies K=3). I think the investigation of different mixing functions is original, interesting, and important. The method is mainly evaluated from downstream classification task and qualitative interpolation examples (in the supplementary material). The proposed method outperforms the baselines when the bottleneck dimensionality is 32. Particularly in low data regimes mixing with more than 2 examples plays an important role. However, I think there are several weaknesses of the paper: (1) methodologically, the main difference and advantage of the proposed method over ACAI lie in the flexibility of mixing functions, but in the full training data setting (Table 1) it seems the simplest mixup performs the best. This suggests that it may not be necessary to use different mixing functions. While the proposed method still outperforms ACAI in this case, the gain is more from a different GAN variant under the ACAI framework than a different mixing function, which is less novel and interesting. (2) When the bottleneck dimension is 256, it looks ACAI mixup(3) greatly underperforms ACAI numbers from the original paper. Thus I doubt that the authors’ reimplementation of ACAI is more or less problematic, as mentioned in line 211. Why not include numbers from ACAI mixup(2) from your implementation ? That should be a more fair comparison to the numbers from the original paper to judge whether your re-implementation is correct or not. (3) Given that AMR is much worse than ACAI when dimension is 256 on SVHN dataset, I suspect it might be the same case for other relatively complex datasets. Does this mean the proposed method is inferior when working with high-complexity dataset ? (4) I read all the qualitative examples in the supplementary material, and I cannot tell whether the proposed AMR is better than ACAI or not. [1] Berthelot et al. Understanding and Improving Interpolation in Autoencoders via an Adversarial Regularizer. ICLR 2019 After rebuttal: Authors responded to my points 1-3. Regarding 1, the added CIFAR10 experiments and SVHN with mixup(4) mitigate my concern about the usefulness of varying mixup functions. Regarding 2, it was a miscommunication caused by a typo in the submission and it is now clear. Regarding 3, it seems that AMR only outperforms the authors' re-implementation of ACAI while underperforming the quoted ACAI results on complex settings (e.g. SVHN (256) and CIFAR10 (1024)). I agree that it has less confounders under the same codebase for comparison but I think it is still important to figure out why your implementation fails to reproduce the results. There might be important (but seemingly ignorable) details to the implementation. Overall, I think the authors addressed part of my concerns and I will increase my score accordingly.