The paper paper applies Deep Kernel Learning [DKL, 1] to Semi-Supervised Regression. DKL is a combination of a Gaussian Process and a Deep Neural Network (DNN). The idea is to use DNN as a feature transformer inside the kernel of a Gaussian Process (GP). In other words, the the GP operates on the outputs of the DNN. Both the GP and the DNN can be trained using SGD in end-to-end fashion.

In particular, the method estimates the depth of the 2D keypoints of the source images using information from both images, and the method estimates the 3D-to-2D affine transform from the source to the target. With this transformation, a traditional keypoint-based face warping (implemented in OpenGL) algorithm and CycleGAN are used to map the source image to the target image. Note that the estimation of the depth and affine transform can either depends on only the 2D keypoints or both the keypoints and images.

This leads to a much more general analysis than earlier work in SSL, both considering misspecification of the mixture and more than 2 classes. They propose several methods to recover the true mapping of decision regions to classes, for which they show both the sample complexity and show empirical results of the probability of correct recovery in three example simulations.

Summary: The paper presents a semi-supervised approach to learning optical flow using a generative adversarial network (GAN) on flow warp errors. Rather than using a handcrafted loss (e.g., deviation of brightness constancy deviation from smoothness) the paper explores the use of a GAN applied to flow warp errors. Strengths: novel semi-supervised approach to learning; some concerns on the novelty in the light of [21] generally written well Weaknesses: - some key evaluations missing Comments: Supervised (e.g., [8]) and unsupervised (e.g., [39]) approaches to optical flow prediction have previously been investigated, the type of semi-supervised supervision proposed here appears novel. The main contribution is in the introduction of an adversarial loss for training rather than the particulars of the flow prediction architecture. As discussed in Sec. 2, [21] also proposes an adversarial scheme.

This paper presents a factorized hierarchical variational autoencoder applied to unsupervised sequence modeling. The main claim of the paper is that the proposed model can disentangle the sequence representation into frame-level and sequence-level components. The sequence-level representation can be used for applications such as speaker verification, without any supervision in learning the representation, and they show that it is better than competitive unsupervised baselines such as using i-vector representations. The model is a mostly straightforward adaptation of sequential VAEs, with the addition of a discriminative regularizer that encourages sequence-level features to be able to predict sequence indices. Does this mean the actual index of a sequence in the training set?

Update based on author rebuttal: The authors address some of my criticisms and promise to improve some of the motivation in subsequent drafts. This paper proposes a system for correcting sequences, with a target application of fixing buggy source code. They use a sequence-to-sequence model within a GAN framework, which allows the model to be trained without paired source/target data. Some additional new tricks are proposed to make the model output consistent translations of the input. The model is tested on two synthetic tasks and a source code correction benchmark.

This paper presents a method to transfer graph structures learned on unlabeled data to downstream tasks, which is a conceptual shift from existing research that aims to transfer features (e.g., embeddings). The method consists of jointly training a feature and graph predictor using an unsupervised objective (which are decoupled) and then extracting only the output of the graph predictor for downstream tasks, where it is multiplicatively applied to arbitrary features. The method yields small improvements on a variety of NLP and vision tasks, and the qualitative analysis of the learned graphs does not convince me that it learns "meaningful" substructures. Overall, however, the paper has a compelling and promising idea (graph transfer), and it seems like there is room to improve on its results, so I'm a weak accept. Detailed comments: - Is "unsupervisedly" a word? It sounds weird... - The objective function in eq 3 is interesting and could have potential uses outside of just graph induction, as it seems especially powerful from the ablations in table 2...

I maintain my original review and think the paper should be accepted. To get around the ambiguity of shape and pose, the authors propose to have an ensemble of pose predictors, which they distill post-training into a single model. I am inclined to accept the paper. The method is a solid solution to an interesting problem and the paper is well-written. In more detail: a) This is clearly a novel solution to an interesting but, so far, poorly explored problem.

This paper addresses the problem of view-invariant action representation within an unsupervised learning framework. In particular, the unsupervised learning task is the prediction of 3D motion from different viewpoints. The proposed model comprises four modules: "encoder", "cross-view decoder", "reconstruction decoder" and "view classifier". For training purposes, a loss function defined as the linear combination of three task-specific losses is proposed. Given an encoding, the cross-view decoder is in charge of estimating the 3D flow in a target view different of the source one.

To circumvent this problem, the authors propose to use a simple variant of InfoGAN [1] with discrete variables.