AISTATS 2019 Papers

An important problem in machine learning and statistics is to identify features that causally affect the outcome. This is often impossible to do from purely observational data, and a natural relaxation is to identify features that are correlated with the outcome even conditioned on all other observed features. For example, we want to identify that smoking really is correlated with cancer conditioned on demographics. The knockoff procedure is a recent breakthrough in statistics that, in theory, can identify truly correlated features while guaranteeing that false discovery rate is controlled. The idea is to create synthetic data-knockoffs-that capture correlations among the features. However, there are substantial computational and practical challenges to generating and using knockoffs. This paper makes several key advances that enable knockoff application to be more efficient and powerful. We develop an efficient algorithm to generate valid knockoffs from Bayesian Networks. Then we systematically evaluate knockoff test statistics and develop new statistics with improved power. The paper combines new mathematical guarantees with systematic experiments on real and synthetic data.

We investigate penalized maximum log-likelihood estimation for exponential family distributions whose natural parameter resides in a reproducing kernel Hilbert space. Key to our approach is a novel technique, doubly dual embedding, that avoids computation of the partition function. This technique also allows the development of a flexible sampling strategy that amortizes the cost of Monte-Carlo sampling in the inference stage. The resulting estimator can be easily generalized to kernel conditional exponential families. We establish a connection between kernel exponential family estimation and MMD-GANs, revealing a new perspective for understanding GANs. Compared to the score matching based estimators, the proposed method improves both memory and time efficiency while enjoying stronger statistical properties, such as fully capturing smoothness in its statistical convergence rate while the score matching estimator appears to saturate. Finally, we show that the proposed estimator empirically outperforms state-of-the-art methods in both kernel exponential family estimation and its conditional extension.

We present new architectures for feedforward neural networks built from products of learned or random low-dimensional rotations that offer substantial space compression and computational speedups in comparison to the unstructured baselines. Models using them are also competitive with the baselines and often, due to imposed orthogonal structure, outperform baselines accuracy-wise. We propose to use our architectures in two settings. We show that in the non-adaptive scenario (random neural networks) they lead to asymptotically more accurate, space-efficient and faster estimators of the so-called PNG-kernels (for any activation function defining the PNG). This generalizes several recent theoretical results about orthogonal estimators (e.g. orthogonal JLTs, orthogonal estimators of angular kernels and more). In the adaptive setting we propose efficient algorithms for learning products of low-dimensional rotations and show how our architectures can be used to improve space and time complexity of state of the art reinforcement learning (RL) algorithms (e.g. PPO, TRPO). Here they offer up to 7x compression of the network in comparison to the unstructured baselines and outperform reward-wise state of the art structured neural networks offering similar computational gains and based on low displacement rank matrices.

Crowdsourcing platforms emerged as popular venues for purchasing human intelligence at low cost for large volume of tasks. As many low-paid workers are prone to give noisy answers, a common practice is to add redundancy by assigning multiple workers to each task and then simply average out these answers. However, to fully harness the wisdom of the crowd, one needs to learn the heterogeneous quality of each worker. We resolve this fundamental challenge in crowdsourced regression tasks, i.e., the answer takes continuous labels, where identifying good or bad workers becomes much more non-trivial compared to a classification setting of discrete labels. In particular, we introduce a Bayesian iterative scheme and show that it provably achieves the optimal mean squared error. Our evaluations on synthetic and real-world datasets support our theoretical results and show the superiority of the proposed scheme.

We present a method for feature interpretation that makes use of recent advances in autoregressive density estimation models to invert model representations. We train generative inversion models to express a distribution over input features conditioned on intermediate model representations. Insights into the invariances learned by supervised models can be gained by viewing samples from these inversion models. In addition, we can use these inversion models to estimate the mutual information between a model’s inputs and its intermediate representations, thus quantifying the amount of information preserved by the network at different stages. Using this method we examine the types of information preserved at different layers of convolutional neural networks, and explore the invariances induced by different architectural choices. Finally we show that the mutual information between inputs and network layers initially increases and then decreases over the course of training, supporting recent work by Shwartz-Ziv and Tishby (2017) on the information bottleneck theory of deep learning.

We study the problem of learning conditional average treatment effects (CATE) from observational data with unobserved confounders. The CATE function maps baseline covariates to individual causal effect predictions and is key for personalized assessments. Recent work has focused on how to learn CATE under unconfoundedness, i.e., when there are no unobserved confounders. Since CATE may not be identified when unconfoundedness is violated, we develop a functional interval estimator that predicts bounds on the individual causal effects under realistic violations of unconfoundedness. Our estimator takes the form of a weighted kernel estimator with weights that vary adversarially. We prove that our estimator is sharp in that it converges exactly to the tightest bounds possible on CATE when there may be unobserved confounders. Further, we study personalized decision rules derived from our estimator and prove that they achieve optimal minimax regret asymptotically. We assess our approach in a simulation study as well as demonstrate its application in the case of hormone replacement therapy by comparing conclusions from a real observational study and clinical trial.

Research in both machine learning and psychology suggests that salient examples can help humans to interpret learning models. To this end, we take a novel look at black box interpretation of test predictions in terms of training examples. Our goal is to ask “which training examples are most responsible for a given set of predictions”? To answer this question, we make use of Fisher kernels as the defining feature embedding of each data point, combined with Sequential Bayesian Quadrature (SBQ) for efficient selection of examples. In contrast to prior work, our method is able to seamlessly handle any sized subset of test predictions in a principled way. We theoretically analyze our approach, providing novel convergence bounds for SBQ over discrete candidate atoms. Our approach recovers the application of influence functions for interpretability as a special case yielding novel insights from this connection. We also present applications of the proposed approach to three use cases: cleaning training data, fixing mislabeled examples and data summarization.

In many prediction tasks such as medical diagnostics, sequential decisions are crucial to provide optimal individual treatment. Budget in real-life applications is always limited, and it can represent any limited resource such as time, money, or side effects of medications. In this contribution, we develop a POMDP-based framework to learn cost-sensitive heterogeneous cascading systems. We provide both the theoretical support for the introduced approach and the intuition behind it. We evaluate our novel method on some standard benchmarks, and we discuss how the learned models can be interpreted by human experts.

Matching methods are heavily used in the social and health sciences due to their interpretability. We aim to create the highest possible quality of treatment-control matches for categorical data in the potential outcomes framework. The method proposed in this work aims to match units on a weighted Hamming distance, taking into account the relative importance of the covariates; the algorithm aims to match units on as many relevant variables as possible. To do this, the algorithm creates a hierarchy of covariate combinations on which to match (similar to downward closure), in the process solving an optimization problem for each unit in order to construct the optimal matches. The algorithm uses a single dynamic program to solve all of the units’ optimization problems simultaneously. Notable advantages of our method over existing matching procedures are its high-quality interpretable matches, versatility in handling different data distributions that may have irrelevant variables, and ability to handle missing data by matching on as many available covariates as possible.

Comparing probability distributions is a fundamental problem in data sciences. Simple norms and divergences such as the total variation and the relative entropy only compare densities in a point-wise manner and fail to capture the geometric nature of the problem. In sharp contrast, Maximum Mean Discrepancies (MMD) and Optimal Transport distances (OT) are two classes of distances between measures that take into account the geometry of the underlying space and metrize the convergence in law. This paper studies the Sinkhorn divergences, a family of geometric divergences that interpolates between MMD and OT. Relying on a new notion of geometric entropy, we provide theoretical guarantees for these divergences: positivity, convexity and metrization of the convergence in law. On the practical side, we detail a numerical scheme that enables the large scale application of these divergences for machine learning: on the GPU, gradients of the Sinkhorn loss can be computed for batches of a million samples.

Motivated by the pursuit of a systematic computational and algorithmic understanding of Generative Adversarial Networks (GANs), we present a simple yet unified non-asymptotic local convergence theory for smooth two-player games, which subsumes several discrete-time gradient-based saddle point dynamics. The analysis reveals the surprising nature of the off-diagonal interaction term as both a blessing and a curse. On the one hand, this interaction term explains the origin of the slow-down effect in the convergence of Simultaneous Gradient Ascent (SGA) to stable Nash equilibria. On the other hand, for the unstable equilibria, exponential convergence can be proved thanks to the interaction term, for four modified dynamics proposed to stabilize GAN training: Optimistic Mirror Descent (OMD), Consensus Optimization (CO), Implicit Updates (IU) and Predictive Method (PM). The analysis uncovers the intimate connections among these stabilizing techniques, and provides detailed characterization on the choice of learning rate. As a by-product, we present a new analysis for OMD proposed in Daskalakis, Ilyas, Syrgkanis, and Zeng [2017] with improved rates.

Methods based on Bayesian decision tree ensembles have proven valuable in constructing high-quality predictions, and are particularly attractive in certain settings because they encourage low-order interaction effects. Despite adapting to the presence of low-order interactions for prediction purpose, we show that Bayesian decision tree ensembles are generally anti-conservative for the purpose of conducting interaction detection. We address this problem by introducing Dirichlet process forests (DP-Forests), which leverage the presence of low-order interactions by clustering the trees so that trees within the same cluster focus on detecting a specific interaction. We show on both simulated and benchmark data that DP-Forests perform well relative to existing interaction detection techniques for detecting low-order interactions, attaining very low false-positive and false-negative rates while maintaining the same performance for prediction using a comparable computational budget.

Machine learning has witnessed tremendous success in solving tasks depending on a single hyperparameter. When considering simultaneously a finite number of tasks, multi-task learning enables one to account for the similarities of the tasks via appropriate regularizers. A step further consists of learning a continuum of tasks for various loss functions. A promising approach, called Parametric Task Learning, has paved the way in the continuum setting for affine models and piecewise-linear loss functions. In this work, we introduce a novel approach called Infinite Task Learning whose goal is to learn a function whose output is a function over the hyperparameter space. We leverage tools from operator-valued kernels and the associated vector-valued RKHSs that provide an explicit control over the role of the hyperparameters, and also allows us to consider new type of constraints. We provide generalization guarantees to the suggested scheme and illustrate its efficiency in cost-sensitive classification, quantile regression and density level set estimation.

Modeling how individuals evolve over time is a fundamental problem in the natural and social sciences. However, existing datasets are often cross-sectional with each individual observed only once, making it impossible to apply traditional time-series methods. Motivated by the study of human aging, we present an interpretable latent-variable model that learns temporal dynamics from cross-sectional data. Our model represents each individual’s features over time as a nonlinear function of a low-dimensional, linearly-evolving latent state. We prove that when this nonlinear function is constrained to be order-isomorphic, the model family is identifiable solely from cross-sectional data provided the distribution of time-independent variation is known. On the UK Biobank human health dataset, our model reconstructs the observed data while learning interpretable rates of aging associated with diseases, mortality, and aging risk factors.

The Model-X knockoff procedure has recently emerged as a powerful approach for feature selection with statistical guarantees. The advantage of knockoffs is that if we have a good model of the features X, then we can identify salient features without knowing anything about how the outcome Y depends on X. An important drawback of knockoffs is its instability: running the procedure twice can result in very different selected features, potentially leading to different conclusions. Addressing this instability is critical for obtaining reproducible and robust results. Here we present a generalization of the knockoff procedure that we call simultaneous multi-knockoffs. We show that multi-knockoffs guarantee false discovery rate (FDR) control, and are substantially more stable and powerful compared to the standard (single) knockoffs. Moreover we propose a new algorithm based on entropy maximization for generating Gaussian multi-knockoffs. We validate the improved stability and power of multi-knockoffs in systematic experiments. We also illustrate how multi-knockoffs can improve the accuracy of detecting genetic mutations that are causally linked to phenotypes.

We present an improved Bayesian framework for performing inference of affine transformations of constrained functions. We focus on quadrature with nonnegative functions, a common task in Bayesian inference. We consider constraints on the range of the function of interest, such as nonnegativity or boundedness. Although our framework is general, we derive explicit approximation schemes for these constraints, and argue for the use of a log transformation for functions with high dynamic range such as likelihood surfaces. We propose a novel method for optimizing hyperparameters in this framework: we optimize the marginal likelihood in the original space, as opposed to in the transformed space. The result is a model that better explains the actual data. Experiments on synthetic and real-world data demonstrate our framework achieves superior estimates using less wall-clock time than existing Bayesian quadrature procedures.

We present a new approach for graph based semi-supervised learning based on a multi-component extension to the Gaussian MRF model. This approach models the observations on the vertices as jointly Gaussian with an inverse covariance matrix that is a weighted linear combination of multiple matrices. Building on randomized matrix trace estimation and fast Laplacian solvers, we develop fast and efficient algorithms for computing the best-fit (maximum likelihood) model and the predicted labels using gradient descent. Our model is considerably simpler, with just tens of parameters, and a single hyperparameter, in contrast with state-of-the-art approaches using deep learning techniques. Our experiments on benchmark citation networks show that the best-fit model estimated by our algorithm leads to significant improvements on all datasets compared to baseline models. Further, our performance compares favorably with several state-of-the-art methods on these datasets, and is comparable with the best performances.

Kernels are powerful and versatile tools in machine learning and statistics. Although the notion of universal kernels and characteristic kernels has been studied, kernel selection still greatly influences the empirical performance. While learning the kernel in a data driven way has been investigated, in this paper we explore learning the spectral distribution of kernel via implicit generative models parametrized by deep neural networks. We called our method Implicit Kernel Learning (IKL). The proposed framework is simple to train and inference is performed via sampling random Fourier features. We investigate two applications of the proposed IKL as examples, including generative adversarial networks with MMD (MMD GAN) and standard supervised learning. Empirically, MMD GAN with IKL outperforms vanilla predefined kernels on both image and text generation benchmarks; using IKL with Random Kitchen Sinks also leads to substantial improvement over existing state-of-the-art kernel learning algorithms on popular supervised learning benchmarks. Theory and conditions for using IKL in both applications are also studied as well as connections to previous state-of-the-art methods.

We study the problem of offline learning in automated decision systems under the contextual bandits model. We are given logged historical data consisting of contexts, (randomized) actions, and (nonnegative) rewards. A common goal is to evaluate what would happen if different actions were taken in the same contexts, so as to optimize the action policies accordingly. The typical approach to this problem, inverse probability weighted estimation (IPWE), requires logged action probabilities, which may be missing in practice due to engineering complications. Even when available, small action probabilities cause large uncertainty in IPWE, rendering the corresponding results insignificant. To solve both problems, we show how one can use policy improvement (PIL) objectives, regularized by policy imitation (IML). We motivate and analyze PIL as an extension to Clipped-IPWE, by showing that both are lower-bound surrogates to the vanilla IPWE. We also formally connect IML to IPWE variance estimation and natural policy gradients. Without probability logging, our PIL-IML interpretations justify and improve, by reward-weighting, the state-of-art cross-entropy (CE) loss that predicts the action items among all action candidates available in the same contexts. With probability logging, our main theoretical contribution connects IML-underfitting to the existence of either confounding variables or model misspecification. We show the value and accuracy of our insights by simulations based on Simpson’s paradox, standard UCI multiclass-to-bandit conversions and on the Criteo counterfactual analysis challenge dataset.