ICML 2018 Papers

Minimax optimization plays a key role in adversarial training of machine learning algorithms, such as learning generative models, domain adaptation, privacy preservation, and robust learning. In this paper, we demonstrate the failure of alternating gradient descent in minimax optimization problems due to the discontinuity of solutions of the inner maximization. To address this, we propose a new $\epsilon$-subgradient descent algorithm that addresses this problem by simultaneously tracking $K$ candidate solutions. Practically, the algorithm can find solutions that previous saddle-point algorithms cannot find, with only a sublinear increase of complexity in $K$. We analyze the conditions under which the algorithm converges to the true solution in detail. A significant improvement in stability and convergence speed of the algorithm is observed in simple representative problems, GAN training, and domain-adaptation problems.

We seek to automate the design of molecules based on specific chemical properties. In computational terms, this task involves continuous embedding and generation of molecular graphs. Our primary contribution is the direct realization of molecular graphs, a task previously approached by generating linear SMILES strings instead of graphs. Our junction tree variational autoencoder generates molecular graphs in two phases, by first generating a tree-structured scaffold over chemical substructures, and then combining them into a molecule with a graph message passing network. This approach allows us to incrementally expand molecules while maintaining chemical validity at every step. We evaluate our model on multiple tasks ranging from molecular generation to optimization. Across these tasks, our model outperforms previous state-of-the-art baselines by a significant margin.

A new generative adversarial network is developed for joint distribution matching.Distinct from most existing approaches, that only learn conditional distributions, the proposed model aims to learn a joint distribution of multiple random variables (domains). This is achieved by learning to sample from conditional distributions between the domains, while simultaneously learning to sample from the marginals of each individual domain.The proposed framework consists of multiple generators and a single softmax-based critic, all jointly trained via adversarial learning.From a simple noise source, the proposed framework allows synthesis of draws from the marginals, conditional draws given observations from a subset of random variables, or complete draws from the full joint distribution. Most examples considered are for joint analysis of two domains, with examples for three domains also presented.

Inference models are a key component in scaling variational inference to deep latent variable models, most notably as encoder networks in variational auto-encoders (VAEs). By replacing conventional optimization-based inference with a learned model, inference is amortized over data examples and therefore more computationally efficient. However, standard inference models are restricted to direct mappings from data to approximate posterior estimates. The failure of these models to reach fully optimized approximate posterior estimates results in an amortization gap. We aim toward closing this gap by proposing iterative inference models, which learn to perform inference optimization through repeatedly encoding gradients. Our approach generalizes standard inference models in VAEs and provides insight into several empirical findings, including top-down inference techniques. We demonstrate the inference optimization capabilities of iterative inference models and show that they outperform standard inference models on several benchmark data sets of images and text.

Recent work suggests that controlling the entire distribution of Jacobian singular values is an important design consideration in deep learning. Motivated by this, we study the distribution of singular values of the Jacobian of the generator in Generative Adversarial Networks. We find that this Jacobian generally becomes ill-conditioned at the beginning of training. Moreover, we find that the average (across the latent space) conditioning of the generator is highly predictive of two other ad-hoc metrics for measuring the “quality” of trained GANs: the Inception Score and the Frechet Inception Distance. We then test the hypothesis that this relationship is causal by proposing a “regularization” technique (called Jacobian Clamping) that softly penalizes the condition number of the generator Jacobian. Jacobian Clamping improves the mean score for nearly all datasets on which we tested it. It also greatly reduces inter-run variance of the aforementioned scores, addressing (at least partially) one of the main criticisms of GANs.

What makes humans so good at solving seemingly complex video games? Unlike computers, humans bring in a great deal of prior knowledge about the world, enabling efficient decision making. This paper investigates the role of human priors for solving video games. Given a sample game, we conduct a series of ablation studies to quantify the importance of various priors on human performance. We do this by modifying the video game environment to systematically mask different types of visual information that could be used by humans as priors. We find that removal of some prior knowledge causes a drastic degradation in the speed with which human players solve the game, e.g. from 2 minutes to over 20 minutes. Furthermore, our results indicate that general priors, such as the importance of objects and visual consistency, are critical for efficient game-play. Videos and the game manipulations are available at https://rach0012.github.io/humanRL_website/

An interesting approach to analyzing neural networks that has received renewed attention is to examine the equivalent kernel of the neural network. This is based on the fact that a fully connected feedforward network with one hidden layer, a certain weight distribution, an activation function, and an infinite number of neurons can be viewed as a mapping into a Hilbert space. We derive the equivalent kernels of MLPs with ReLU or Leaky ReLU activations for all rotationally-invariant weight distributions, generalizing a previous result that required Gaussian weight distributions. Additionally, the Central Limit Theorem is used to show that for certain activation functions, kernels corresponding to layers with weight distributions having $0$ mean and finite absolute third moment are asymptotically universal, and are well approximated by the kernel corresponding to layers with spherical Gaussian weights. In deep networks, as depth increases the equivalent kernel approaches a pathological fixed point, which can be used to argue why training randomly initialized networks can be difficult. Our results also have implications for weight initialization.

The interpretation of deep learning models is a challenge due to their size, complexity, and often opaque internal state. In addition, many systems, such as image classifiers, operate on low-level features rather than high-level concepts. To address these challenges, we introduce Concept Activation Vectors (CAVs), which provide an interpretation of a neural net’s internal state in terms of human-friendly concepts. The key idea is to view the high-dimensional internal state of a neural net as an aid, not an obstacle. We show how to use CAVs as part of a technique, Testing with CAVs (TCAV), that uses directional derivatives to quantify the degree to which a user-defined concept is important to a classification result–for example, how sensitive a prediction of “zebra” is to the presence of stripes. Using the domain of image classification as a testing ground, we describe how CAVs may be used to explore hypotheses and generate insights for a standard image classification network as well as a medical application.

One important task of topic modeling for text analysis is interpretability. By discovering structured topics one is able to yield improved interpretability as well as modeling accuracy. In this paper, we propose a novel topic model with a deep structure that explores both inter-topic and intra-topic structures informed by word embeddings. Specifically, our model discovers inter topic structures in the form of topic hierarchies and discovers intra topic structures in the form of sub-topics, each of which is informed by word embeddings and captures a fine-grained thematic aspect of a normal topic. Extensive experiments demonstrate that our model achieves the state-of-the-art performance in terms of perplexity, document classification, and topic quality. Moreover, with topic hierarchies and sub-topics, the topics discovered in our model are more interpretable, providing an illuminating means to understand text data.

In this paper, we derive bounds on the mutual information of the empirical risk minimization (ERM) procedure for both 0-1 and strongly-convex loss classes. We prove that under the Axiom of Choice, the existence of an ERM learning rule with a vanishing mutual information is equivalent to the assertion that the loss class has a finite VC dimension, thus bridging information theory with statistical learning theory. Similarly, an asymptotic bound on the mutual information is established for strongly-convex loss classes in terms of the number of model parameters. The latter result rests on a central limit theorem (CLT) that we derive in this paper. In addition, we use our results to analyze the excess risk in stochastic convex optimization and unify previous works. Finally, we present two important applications. First, we show that the ERM of strongly-convex loss classes can be trivially scaled to big data using a naive parallelization algorithm with provable guarantees. Second, we propose a simple information criterion for model selection and demonstrate experimentally that it outperforms the popular Akaike’s information criterion (AIC) and Schwarz’s Bayesian information criterion (BIC).

Amortized inference allows latent-variable models trained via variational learning to scale to large datasets. The quality of approximate inference is determined by two factors: a) the capacity of the variational distribution to match the true posterior and b) the ability of the recognition network to produce good variational parameters for each datapoint. We examine approximate inference in variational autoencoders in terms of these factors. We find that divergence from the true posterior is often due to imperfect recognition networks, rather than the limited complexity of the approximating distribution. We show that this is due partly to the generator learning to accommodate the choice of approximation. Furthermore, we show that the parameters used to increase the expressiveness of the approximation play a role in generalizing inference rather than simply improving the complexity of the approximation.

Latent prediction models, exemplified by multi-layer networks, employ hidden variables that automate abstract feature discovery. They typically pose nonconvex optimization problems and effective semi-definite programming (SDP) relaxations have been developed to enable global solutions (Aslan et al., 2014).However, these models rely on nonparametric training of layer-wise kernel representations, and are therefore restricted to transductive learning which slows down test prediction. In this paper, we develop a new inductive learning framework for parametric transfer functions using matching losses. The result for ReLU utilizes completely positive matrices, and the inductive learner not only delivers superior accuracy but also offers an order of magnitude speedup over SDP with constant approximation guarantees.

Alternating direction method of multiplier (ADMM) is a popular method used to design distributed versions of a machine learning algorithm, whereby local computations are performed on local data with the output exchanged among neighbors in an iterative fashion. During this iterative process the leakage of data privacy arises. A differentially private ADMM was proposed in prior work (Zhang & Zhu, 2017) where only the privacy loss of a single node during one iteration was bounded, a method that makes it difficult to balance the tradeoff between the utility attained through distributed computation and privacy guarantees when considering the total privacy loss of all nodes over the entire iterative process. We propose a perturbation method for ADMM where the perturbed term is correlated with the penalty parameters; this is shown to improve the utility and privacy simultaneously. The method is based on a modified ADMM where each node independently determines its own penalty parameter in every iteration and decouples it from the dual updating step size. The condition for convergence of the modified ADMM and the lower bound on the convergence rate are also derived.

The Gaussian mechanism is an essential building block used in multitude of differentially private data analysis algorithms. In this paper we revisit the Gaussian mechanism and show that the original analysis has several important limitations. Our analysis reveals that the variance formula for the original mechanism is far from tight in the high privacy regime ($\varepsilon \to 0$) and it cannot be extended to the low privacy regime ($\varepsilon \to \infty$). We address these limitations by developing an optimal Gaussian mechanism whose variance is calibrated directly using the Gaussian cumulative density function instead of a tail bound approximation. We also propose to equip the Gaussian mechanism with a post-processing step based on adaptive estimation techniques by leveraging that the distribution of the perturbation is known. Our experiments show that analytical calibration removes at least a third of the variance of the noise compared to the classical Gaussian mechanism, and that denoising dramatically improves the accuracy of the Gaussian mechanism in the high-dimensional regime.

Sign random projections (SRP) is a technique which allows the user to quickly estimate the angular similarity and inner products between data. We propose using additional information to improve these estimates which is easy to implement and cost efficient. We prove that the variance of our estimator is lower than the variance of SRP. Our proposed method can also be used together with other modifications of SRP, such as Super-Bit LSH (SBLSH). We demonstrate the effectiveness of our method on the MNIST test dataset and the Gisette dataset. We discuss how our proposed method can be extended to random projections or even other hashing algorithms.

There is growing evidence that converting targets to soft targets in supervised learning can provide considerable gains in performance. Much of this work has considered classification, converting hard zero-one values to soft labels—such as by adding label noise, incorporating label ambiguity or using distillation. In parallel, there is some evidence from a regression setting in reinforcement learning that learning distributions can improve performance. In this work, we investigate the reasons for this improvement, in a regression setting. We introduce a novel distributional regression loss, and similarly find it significantly improves prediction accuracy. We investigate several common hypotheses, around reducing overfitting and improved representations. We instead find evidence for an alternative hypothesis: this loss is easier to optimize, with better behaved gradients, resulting in improved generalization. We provide theoretical support for this alternative hypothesis, by characterizing the norm of the gradients of this loss.

Many loss functions in representation learning are invariant under a continuous symmetry transformation. For example, the loss function of word embeddings (Mikolov et al., 2013) remains unchanged if we simultaneously rotate all word and context embedding vectors. We show that representation learning models for time series possess an approximate continuous symmetry that leads to slow convergence of gradient descent. We propose a new optimization algorithm that speeds up convergence using ideas from gauge theory in physics. Our algorithm leads to orders of magnitude faster convergence and to more interpretable representations, as we show for dynamic extensions of matrix factorization and word embedding models. We further present an example application of our proposed algorithm that translates modern words into their historic equivalents.

Nearest neighbor search using random projection trees has recently been shown to achieve superior performance, in terms of better accuracy while retrieving less number of data points, compared to locality sensitive hashing based methods. However, to achieve acceptable nearest neighbor search accuracy for large scale applications, where number of data points and/or number of features can be very large, it requires users to maintain, store and search through large number of such independent random projection trees, which may be undesirable for many practical applications. To address this issue, in this paper we present different search strategies to improve nearest neighbor search performance of a single random projection tree. Our approach exploits properties of single and multiple random projections, which allows us to store meaningful auxiliary information at internal nodes of a random projection tree as well as to design priority functions to guide the search process that results in improved nearest neighbor search performance. Empirical results on multiple real world datasets show that our proposed method improves the search accuracy of a single tree compared to baseline methods.

The representation and learning benefits of methods based on graph Laplacians, such as Laplacian smoothing or harmonic function solution for semi-supervised learning (SSL), are empirically and theoretically well supported. Nonetheless, the exact versions of these methods scale poorly with the number of nodes $n$ of the graph. In this paper, we combine a spectral sparsification routine with Laplacian learning. Given a graph $G$ as input, our algorithm computes a sparsifier in a distributed way in $O(n\log^3(n))$ time, $O(m\log^3(n))$ work and $O(n\log(n))$ memory, using only $\log(n)$ rounds of communication. Furthermore, motivated by the regularization often employed in learning algorithms, we show that constructing sparsifiers that preserve the spectrum of the Laplacian only up to the regularization level may drastically reduce the size of the final graph. By constructing a spectrally-similar graph, we are able to bound the error induced by the sparsification for a variety of downstream tasks (e.g., SSL). We empirically validate the theoretical guarantees on Amazon co-purchase graph and compare to the state-of-the-art heuristics.