arXiv subjects

Michael Betancourt

Publications and source records attributed to Michael Betancourt.

28 records · Page 2Linked to original sources

arXiv2018

On Symplectic Optimization

Accelerated gradient methods have had significant impact in machine learning -- in particular the theoretical side of machine learning -- due to their ability to achieve oracle lower bounds. But their heuristic construction has hindered their full integration into the practical machine-learning algorithmic toolbox, and has limited their scope. In this paper we build on recent work which casts acceleration as a phenomenon best explained in continuous time, and we augment that picture by providing a systematic methodology for converting continuous-time dynamics into discrete-time algorithms while retaining oracle rates. Our framework is based on ideas from Hamiltonian dynamical systems and symplectic integration. These ideas have had major impact in many areas in applied mathematics, but have not yet been seen to have a relationship with optimization.

stat.CO↗

arXiv2018

The Convergence of Markov chain Monte Carlo Methods: From the Metropolis method to Hamiltonian Monte Carlo

From its inception in the 1950s to the modern frontiers of applied statistics, Markov chain Monte Carlo has been one of the most ubiquitous and successful methods in statistical computing. In that time its development has been fueled by increasingly difficult problems and novel techniques from physics. In this article I will review the history of Markov chain Monte Carlo from its inception with the Metropolis method to today's state-of-the-art in Hamiltonian Monte Carlo. Along the way I will focus on the evolving interplay between the statistical and physical perspectives of the method.

stat.ME↗

arXiv2017

The prior can generally only be understood in the context of the likelihood

A key sticking point of Bayesian analysis is the choice of prior distribution, and there is a vast literature on potential defaults including uniform priors, Jeffreys' priors, reference priors, maximum entropy priors, and weakly informative priors. These methods, however, often manifest a key conceptual tension in prior modeling: a model encoding true prior information should be chosen without reference to the model of the measurement process, but almost all common prior modeling techniques are implicitly motivated by a reference likelihood. In this paper we resolve this apparent paradox by placing the choice of prior into the context of the entire Bayesian analysis, from inference to prediction to model evaluation.

stat.ME↗

arXiv2017

Hierarchical inference of the relationship between Concentration and Mass in Galaxy Groups and Clusters

Mass is a fundamental property of galaxy groups and clusters. In theory weak gravitational lensing will enable an approximately unbiased measurement of mass, but parametric methods for extracting cluster masses from data require the additional knowledge of concentration. Measurements of both mass and concentration are limited by the degeneracy between the two parameters, particularly in low mass, high redshift systems where the signal-to-noise is low. In this paper we develop a hierarchical model of mass and concentration for mass inference we test our method on toy data and then apply it to a sample of galaxy groups and poor clusters down to masses of $\sim$ 1e13 M$_\odot$. Our fit and model gives a relationship among masses, concentrations and redshift that allow prediction of these parameters from incomplete and noisy future measurements. Additionally the underlying population can be used to infer an observationally based concentration-mass relation. Our method is equivalent to a quasi- stacking approach with the degree of stacking set by the data. We also demonstrate that mass and concentration derived from pure stacking can be offset from the population mean with differing values depending on the method of stacking.

astro-ph.CO↗

arXiv2016

Diagnosing Suboptimal Cotangent Disintegrations in Hamiltonian Monte Carlo

When properly tuned, Hamiltonian Monte Carlo scales to some of the most challenging high-dimensional problems at the frontiers of applied statistics, but when that tuning is suboptimal the performance leaves much to be desired. In this paper I show how suboptimal choices of one critical degree of freedom, the cotangent disintegration, manifest in readily observed diagnostics that facilitate the robust application of the algorithm.

stat.ME↗

arXiv2016

Identifying the Optimal Integration Time in Hamiltonian Monte Carlo

By leveraging the natural geometry of a smooth probabilistic system, Hamiltonian Monte Carlo yields computationally efficient Markov Chain Monte Carlo estimation. At least provided that the algorithm is sufficiently well-tuned. In this paper I show how the geometric foundations of Hamiltonian Monte Carlo implicitly identify the optimal choice of these parameters, especially the integration time. I then consider the practical consequences of these principles in both existing algorithms and a new implementation called \emph{Exhaustive Hamiltonian Monte Carlo} before demonstrating the utility of these ideas in some illustrative examples.

stat.ME↗

arXiv2015

The Stan Math Library: Reverse-Mode Automatic Differentiation in C++

As computational challenges in optimization and statistical inference grow ever harder, algorithms that utilize derivatives are becoming increasingly more important. The implementation of the derivatives that make these algorithms so powerful, however, is a substantial user burden and the practicality of these algorithms depends critically on tools like automatic differentiation that remove the implementation burden entirely. The Stan Math Library is a C++, reverse-mode automatic differentiation library designed to be usable, extensive and extensible, efficient, scalable, stable, portable, and redistributable in order to facilitate the construction and utilization of such algorithms. Usability is achieved through a simple direct interface and a cleanly abstracted functional interface. The extensive built-in library includes functions for matrix operations, linear algebra, differential equation solving, and most common probability functions. Extensibility derives from a straightforward object-oriented framework for expressions, allowing users to easily create custom functions. Efficiency is achieved through a combination of custom memory management, subexpression caching, traits-based metaprogramming, and expression templates. Partial derivatives for compound functions are evaluated lazily for improved scalability. Stability is achieved by taking care with arithmetic precision in algebraic expressions and providing stable, compound functions where possible. For portability, the library is standards-compliant C++ (03) and has been tested for all major compilers for Windows, Mac OS X, and Linux.

cs.MS↗

arXiv2015

A Unified Treatment of Predictive Model Comparison

The predictive performance of any inferential model is critical to its practical success, but quantifying predictive performance is a subtle statistical problem. In this paper I show how the natural structure of any inferential problem defines a canonical measure of relative predictive performance and then demonstrate how approximations of this measure yield many of the model comparison techniques popular in statistics and machine learning.

stat.ME↗

arXiv2014

Unsupervised Transient Light Curve Analysis Via Hierarchical Bayesian Inference

Historically, light curve studies of supernovae (SNe) and other transient classes have focused on individual objects with copious and high signal-to-noise observations. In the nascent era of wide field transient searches, objects with detailed observations are decreasing as a fraction of the overall known SN population, and this strategy sacrifices the majority of the information contained in the data about the underlying population of transients. A population level modeling approach, simultaneously fitting all available observations of objects in a transient sub-class of interest, fully mines the data to infer the properties of the population and avoids certain systematic biases. We present a novel hierarchical Bayesian statistical model for population level modeling of transient light curves, and discuss its implementation using an efficient Hamiltonian Monte Carlo technique. As a test case, we apply this model to the Type IIP SN sample from the Pan-STARRS1 Medium Deep Survey, consisting of 18,837 photometric observations of 76 SNe, corresponding to a joint posterior distribution with 9,176 parameters under our model. Our hierarchical model fits provide improved constraints on light curve parameters relevant to the physical properties of their progenitor stars relative to modeling individual light curves alone. Moreover, we directly evaluate the probability for occurrence rates of unseen light curve characteristics from the model hyperparameters, addressing observational biases in survey methodology. We view this modeling framework as an unsupervised machine learning technique with the ability to maximize scientific returns from data to be collected by future wide field transient searches like LSST. \smallskip

astro-ph.IM↗

arXiv2011

The Geometry of Hamiltonian Monte Carlo

With its systematic exploration of probability distributions, Hamiltonian Monte Carlo is a potent Markov Chain Monte Carlo technique; it is an approach, however, ultimately contingent on the choice of a suitable Hamiltonian function. By examining both the symplectic geometry underlying Hamiltonian dynamics and the requirements of Markov Chain Monte Carlo, we construct the general form of admissible Hamiltonians and propose a particular choice with potential application in Bayesian inference.

stat.ME↗