arXiv subjects

Jonathan Taylor

Publications and source records attributed to Jonathan Taylor.

At least 55 records · Page 3Linked to original sources

arXiv2017

Bootstrap inference after using multiple queries for model selection

In this work, we provide a refinement of the selective CLT result of Tian and Taylor (2015), which allows for selective inference in non-parametric settings by adjusting for the asymptotic Gaussian limit for selection. Under some regularity assumptions on the density of the randomization, including heavier tails than Gaussian satisfied by e.g. logistic distribution, we prove the selective CLT holds without any assumptions on the underlying parameter, allowing for rare selection events. We also show a selective CLT result for Gaussian randomization, though the quantitative results are qualitatively different for the Gaussian randomization as compared to the heavier tailed results. Furthermore, we propose a bootstrap version of this test statistic, which is provably asymptotically pivotal uniformly across a family of non-parametric distributions. This result can be interpreted as resolving the impossibility results of Leeb and Potscher (2006). We describe several sampling methods involving the projected Langevin Monte Carlo to compute the bootstrapped test statistic and the corresponding confidence intervals valid after selection. The applications of our work include valid inferential and sampling tools after running various model selection algorithms including their combinations into multiple views/queries framework. We also present a way to do data carving, providing more powerful tests than classical data splitting by reusing the information in the data from the first stage.

stat.ME↗

arXiv2017

Scalable methods for Bayesian selective inference

Modeled along the truncated approach in Panigrahi (2016), selection-adjusted inference in a Bayesian regime is based on a selective posterior. Such a posterior is determined together by a generative model imposed on data and the selection event that enforces a truncation on the assumed law. The effective difference between the selective posterior and the usual Bayesian framework is reflected in the use of a truncated likelihood. The normalizer of the truncated law in the adjusted framework is the probability of the selection event; this is typically intractable and it leads to the computational bottleneck in sampling from such a posterior. The current work lays out a primal-dual approach of solving an approximating optimization problem to provide valid post-selective Bayesian inference. The selection procedures are posed as data-queries that solve a randomized version of a convex learning program which have the advantage of preserving more left-over information for inference. We propose a randomization scheme under which the optimization has separable constraints that result in a partially separable objective in lower dimensions for many commonly used selective queries to approximate the otherwise intractable selective posterior. We show that the approximating optimization under a Gaussian randomization gives a valid exponential rate of decay for the selection probability on a large deviation scale. We offer a primal-dual method to solve the optimization problem leading to an approximate posterior; this allows us to exploit the usual merits of a Bayesian machinery in both low and high dimensional regimes where the underlying signal is effectively sparse. We show that the adjusted estimates empirically demonstrate better frequentist properties in comparison to the unadjusted estimates based on the usual posterior, when applied to a wide range of constrained, convex data queries.

stat.ME↗

arXiv2017

Inferactive data analysis

We describe inferactive data analysis, so-named to denote an interactive approach to data analysis with an emphasis on inference after data analysis. Our approach is a compromise between Tukey's exploratory (roughly speaking "model free") and confirmatory data analysis (roughly speaking classical and "model based"), also allowing for Bayesian data analysis. We view this approach as close in spirit to current practice of applied statisticians and data scientists while allowing frequentist guarantees for results to be reported in the scientific literature, or Bayesian results where the data scientist may choose the statistical model (and hence the prior) after some initial exploratory analysis. While this approach to data analysis does not cover every scenario, and every possible algorithm data scientists may use, we see this as a useful step in concrete providing tools (with frequentist statistical guarantees) for current data scientists. The basis of inference we use is selective inference [Lee et al., 2016, Fithian et al., 2014], in particular its randomized form [Tian and Taylor, 2015a]. The randomized framework, besides providing additional power and shorter confidence intervals, also provides explicit forms for relevant reference distributions (up to normalization) through the {\em selective sampler} of Tian et al. [2016]. The reference distributions are constructed from a particular conditional distribution formed from what we call a DAG-DAG -- a Data Analysis Generative DAG. As sampling conditional distributions in DAGs is generally complex, the selective sampler is crucial to any practical implementation of inferactive data analysis. Our principal goal is in reviewing the recent developments in selective inference as well as describing the general philosophy of selective inference.

math.ST↗

arXiv2017

An MCMC-free approach to post-selective inference

We develop a Monte Carlo-free approach to inference post output from randomized algorithms with a convex loss and a convex penalty. The pivotal statistic based on a truncated law, called the selective pivot, usually lacks closed form expressions. Inference in these settings relies upon standard Monte Carlo sampling techniques at a reference parameter followed by an exponential tilting at the reference. Tilting can however be unstable for parameters that are far off from the reference parameter. We offer in this paper an alternative approach to construction of intervals and point estimates by proposing an approximation to the intractable selective pivot. Such an approximation solves a convex optimization problem in |E| dimensions, where |E| is the size of the active set observed from selection. We empirically show that the confidence intervals obtained by inverting the approximate pivot have valid coverage.

stat.ME↗

arXiv2017

Kinematic Formula for Heterogeneous Gaussian Related Fields

We provide a generalization of the Gaussian Kinematic Formula (GKF) in Taylor(2006) for multivariate, heterogeneous Gaussian-related fields. The fields under consideration are non-Gaussian fields built out of smooth, independent Gaussian fields with heterogeneity in distribution amongst the individual building blocks. Our motivation comes from potential applications in the analysis of Cosmological Data (CMB). Specifically, future CMB experiments will be focusing on polarization data, typically modeled as isotropic vector-valued Gaussian related fields with independent, but non-identically distributed Gaussian building blocks; this necessitates such a generalization. Extending results of Taylor(2006) to these more general Gaussian relatives with distributional heterogeneity, we present a generalized Gaussian Kinematic Formula (GKF). The GKF in this paper decouples the expected Euler characteristic of excursion sets into Lipschitz Killing Curvatures (LKCs) of the underlying manifold and certain Gaussian Minkowski Functionals (GMFs). These GMFs arise from Gaussian volume expansions of ellipsoidal tubes as opposed to the usual tubes in the Euclidean volume of classical tube formulae. The GMFs form a main contribution of this work that identifies this tubular structure and a corresponding volume of tubes expansion in which the GMFs appear.

math.PR↗

arXiv2017

Optimal Inference After Model Selection

To perform inference after model selection, we propose controlling the selective type I error; i.e., the error rate of a test given that it was performed. By doing so, we recover long-run frequency properties among selected hypotheses analogous to those that apply in the classical (non-adaptive) context. Our proposal is closely related to data splitting and has a similar intuitive justification, but is more powerful. Exploiting the classical theory of Lehmann and Scheffé (1955), we derive most powerful unbiased selective tests and confidence intervals for inference in exponential family models after arbitrary selection procedures. For linear regression, we derive new selective z-tests that generalize recent proposals for inference after model selection and improve on their power, and new selective t-tests that do not require knowledge of the error variance.

math.ST↗

arXiv2016

Summary - TerpreT: A Probabilistic Programming Language for Program Induction

We study machine learning formulations of inductive program synthesis; that is, given input-output examples, synthesize source code that maps inputs to corresponding outputs. Our key contribution is TerpreT, a domain-specific language for expressing program synthesis problems. A TerpreT model is composed of a specification of a program representation and an interpreter that describes how programs map inputs to outputs. The inference task is to observe a set of input-output examples and infer the underlying program. From a TerpreT model we automatically perform inference using four different back-ends: gradient descent (thus each TerpreT model can be seen as defining a differentiable interpreter), linear program (LP) relaxations for graphical models, discrete satisfiability solving, and the Sketch program synthesis system. TerpreT has two main benefits. First, it enables rapid exploration of a range of domains, program representations, and interpreter models. Second, it separates the model specification from the inference algorithm, allowing proper comparisons between different approaches to inference. We illustrate the value of TerpreT by developing several interpreter models and performing an extensive empirical comparison between alternative inference algorithms on a variety of program models. To our knowledge, this is the first work to compare gradient-based search over program space to traditional search-based alternatives. Our key empirical finding is that constraint solvers dominate the gradient descent and LP-based formulations. This is a workshop summary of a longer report at arXiv:1608.04428

cs.LG↗

arXiv2016

High-dimensional regression adjustments in randomized experiments

We study the problem of treatment effect estimation in randomized experiments with high-dimensional covariate information, and show that essentially any risk-consistent regression adjustment can be used to obtain efficient estimates of the average treatment effect. Our results considerably extend the range of settings where high-dimensional regression adjustments are guaranteed to provide valid inference about the population average treatment effect. We then propose cross-estimation, a simple method for obtaining finite-sample-unbiased treatment effect estimates that leverages high-dimensional regression adjustments. Our method can be used when the regression model is estimated using the lasso, the elastic net, subset selection, etc. Finally, we extend our analysis to allow for adaptive specification search via cross-validation, and flexible non-parametric regression adjustments with machine learning methods such as random forests or neural networks.

stat.ME↗

arXiv2016

Post-selection inference for L1-penalized likelihood models

We present a new method for post-selection inference for L1 (lasso)-penalized likelihood models, including generalized regression models. Our approach generalizes the post-selection framework presented in Lee et al (2014). The method provides p-values and confidence intervals that are asymptotically valid, conditional on the inherent selection done by the lasso. We present applications of this work to (regularized) logistic regression, Cox's proportional hazards model and the graphical lasso.

stat.ME↗

arXiv2016

Selective sampling after solving a convex problem

We consider the problem of selective inference after solving a (randomized) convex statistical learning program in the form of a penalized or constrained loss function. Our first main result is a change-of-measure formula that describes many conditional sampling problems of interest in selective inference. Our approach is model-agnostic in the sense that users may provide their own statistical model for inference, we simply provide the modification of each distribution in the model after the selection. Our second main result describes the geometric structure in the Jacobian appearing in the change of measure, drawing connections to curvature measures appearing in Weyl-Steiner volume-of-tubes formulae. This Jacobian is necessary for problems in which the convex penalty is not polyhedral, with the prototypical example being group LASSO or the nuclear norm. We derive explicit formulae for the Jacobian of the group LASSO. To illustrate the generality of our method, we consider many examples throughout, varying both the penalty or constraint in the statistical learning problem as well as the loss function, also considering selective inference after solving multiple statistical learning programs. Penalties considered include LASSO, forward stepwise, stagewise algorithms, marginal screening and generalized LASSO. Loss functions considered include squared-error, logistic, and log-det for covariance matrix estimation. Having described the appropriate distribution we wish to sample from through our first two results, we outline a framework for sampling using a projected Langevin sampler in the (commonly occuring) case that the distribution is log-concave.

math.ST↗

arXiv2016

TerpreT: A Probabilistic Programming Language for Program Induction

We study machine learning formulations of inductive program synthesis; given input-output examples, we try to synthesize source code that maps inputs to corresponding outputs. Our aims are to develop new machine learning approaches based on neural networks and graphical models, and to understand the capabilities of machine learning techniques relative to traditional alternatives, such as those based on constraint solving from the programming languages community. Our key contribution is the proposal of TerpreT, a domain-specific language for expressing program synthesis problems. TerpreT is similar to a probabilistic programming language: a model is composed of a specification of a program representation (declarations of random variables) and an interpreter describing how programs map inputs to outputs (a model connecting unknowns to observations). The inference task is to observe a set of input-output examples and infer the underlying program. TerpreT has two main benefits. First, it enables rapid exploration of a range of domains, program representations, and interpreter models. Second, it separates the model specification from the inference algorithm, allowing like-to-like comparisons between different approaches to inference. From a single TerpreT specification we automatically perform inference using four different back-ends. These are based on gradient descent, linear program (LP) relaxations for graphical models, discrete satisfiability solving, and the Sketch program synthesis system. We illustrate the value of TerpreT by developing several interpreter models and performing an empirical comparison between alternative inference algorithms. Our key empirical finding is that constraint solvers dominate the gradient descent and LP-based formulations. We conclude with suggestions for the machine learning community to make progress on program synthesis.

cs.LG↗

arXiv2016

Asymptotics of selective inference

In this paper, we seek to establish asymptotic results for selective inference procedures removing the assumption of Gaussianity. The class of selection procedures we consider are determined by affine inequalities, which we refer to as affine selection procedures. Examples of affine selection procedures include post-selection inference along the solution path of the LASSO, as well as post-selection inference after fitting the LASSO at a fixed value of the regularization parameter. We also consider some tests in penalized generalized linear models. Our method of proof adapts a method of Chatterjee (2005).

math.ST↗

arXiv2016

MAGIC: a general, powerful and tractable method for selective inference

Selective inference is a recent research topic that tries to perform valid inference after using the data to select a reasonable statistical model. We propose MAGIC, a new method for selective inference that is general, powerful and tractable. MAGIC is a method for selective inference after solving a convex optimization problem with smooth loss and $\ell_1$ penalty. Randomization is incorporated into the optimization problem to boost statistical power. Through reparametrization, MAGIC reduces the problem into a sampling problem with simple constraints. MAGIC applies to many $\ell_1$ penalized optimization problem including the Lasso, logistic Lasso and neighborhood selection in graphical models, all of which we consider in this paper.

math.ST↗

arXiv2015

Selective Sequential Model Selection

Many model selection algorithms produce a path of fits specifying a sequence of increasingly complex models. Given such a sequence and the data used to produce them, we consider the problem of choosing the least complex model that is not falsified by the data. Extending the selected-model tests of Fithian et al. (2014), we construct p-values for each step in the path which account for the adaptive selection of the model path using the data. In the case of linear regression, we propose two specific tests, the max-t test for forward stepwise regression (generalizing a proposal of Buja and Brown (2014)), and the next-entry test for the lasso. These tests improve on the power of the saturated-model test of Tibshirani et al. (2014), sometimes dramatically. In addition, our framework extends beyond linear regression to a much more general class of parametric and nonparametric model selection problems. To select a model, we can feed our single-step p-values as inputs into sequential stopping rules such as those proposed by G'Sell et al. (2013) and Li and Barber (2015), achieving control of the familywise error rate or false discovery rate (FDR) as desired. The FDR-controlling rules require the null p-values to be independent of each other and of the non-null p-values, a condition not satisfied by the saturated-model p-values of Tibshirani et al. (2014). We derive intuitive and general sufficient conditions for independence, and show that our proposed constructions yield independent p-values.

stat.ME↗

arXiv2015

A general framework for estimation and inference from clusters of features

Applied statistical problems often come with pre-specified groupings to predictors. It is natural to test for the presence of simultaneous group-wide signal for groups in isolation, or for multiple groups together. Classical tests for the presence of such signals rely either on tests for the omission of the entire block of variables (the classical F-test) or on the creation of an unsupervised prototype for the group (either a group centroid or first principal component) and subsequent t-tests on these prototypes. In this paper, we propose test statistics that aim for power improvements over these classical approaches. In particular, we first create group prototypes, with reference to the response, hopefully improving on the unsupervised prototypes, and then testing with likelihood ratio statistics incorporating only these prototypes. We propose a (potentially) novel model, called the "prototype model", which naturally models the two-step prototype-then-test procedure. Furthermore, we introduce an inferential schema detailing the unique considerations for different combinations of prototype formation and univariate/multivariate testing models. The prototype model also suggests new applications to estimation and prediction. Prototype formation often relies on variable selection, which invalidates classical Gaussian test theory. We use recent advances in selective inference to account for selection in the prototyping step and retain test validity. Simulation experiments suggest that our testing procedure enjoys more power than do classical approaches.

stat.AP↗

arXiv2015

Exact Post-Selection Inference for Sequential Regression Procedures

We propose new inference tools for forward stepwise regression, least angle regression, and the lasso. Assuming a Gaussian model for the observation vector y, we first describe a general scheme to perform valid inference after any selection event that can be characterized as y falling into a polyhedral set. This framework allows us to derive conditional (post-selection) hypothesis tests at any step of forward stepwise or least angle regression, or any step along the lasso regularization path, because, as it turns out, selection events for these procedures can be expressed as polyhedral constraints on y. The p-values associated with these tests are exactly uniform under the null distribution, in finite samples, yielding exact type I error control. The tests can also be inverted to produce confidence intervals for appropriate underlying regression parameters. The R package "selectiveInference", freely available on the CRAN repository, implements the new inference tools described in this paper.

stat.ME↗

arXiv2015

A Selective Approach to Internal Inference

A common goal in modern biostatistics is to form a biomarker signature from high dimensional gene expression data that is predictive of some outcome of interest. After learning this biomarker signature, an important question to answer is how well it predicts the response compared to classical predictors. This is challenging, because the biomarker signature is an internal predictor -- one that has been learned using the same dataset on which we want to evaluate it's significance. We propose a new method for approaching this problem based on the technique of selective inference. Simulations show that our method is able to properly control the level of the test, and that in certain settings we have more power than sample splitting.

math.ST↗

arXiv2015

Gaussian measures on the of space of Riemannian metrics

We introduce Gaussian-type measures on the manifold of all metrics with a fixed volume form on a compact Riemannian manifold of dimension $\geq 3$. For this random model we compute the characteristic function for the $L^2$ (Ebin) distance to the reference metric. In the Appendix, we study Lipschitz-type distance between Riemannian metrics and give applications to the diameter, eigenvalue and volume entropy functionals.

math.DG↗