## 第7回 統計的機械学習セミナー (2012.7.2)

第7回 統計的機械学習セミナー／The 7th Statistical Machine Learning Seminar

（主催：統計数理研究所 統計的機械学習NOE・統計的機械学習研究センター)

**日時** 2012年7月2日（月） 15:00-18:00

**会場** 統計数理研究所 セミナー室5（3階D313, D314）

1. Matthew Parry (Otago University, New Zealand)

title: **The entropy of scoring rules**

abstract: A scoring rule is a principled way of assessing a

probabilistic statement. As such, it finds uses in forecasting and

statistical inference. The key requirement of a scoring rule is that

it rewards honest statements of ones beliefs.

Associated with each scoring rule is a concave entropy. Conversely, we

may (almost) think of each concave entropy as generating a scoring

rule. The obvious question is then what features of the entropy are

transferred to the scoring rule. I report on recent work on extensive

entropies and local entropies. Local entropies are particularly

interesting in that they give rise to scoring rules that can assess

probability models whose normalization is unknown or is not feasible

to compute. I will discuss an application to Bayesian inference for

doubly intractable distributions.

2. Ben Calderhead (University College London, UK)

title: **A Sample of Differential Geometric MCMC Methods**

abstract: Markov chain Monte Carlo methods enable samples to be drawn

from arbitrary probability distributions, and advances in such

algorithms have fuelled the rapid expansion in the use of Bayesian

methodology over the last 20 years. However, one of the enduring

challenges in MCMC methodology is the development of proposal

mechanisms that make moves distant from the current point, yet are

accepted with high probability and at low computational cost.

In this talk I will introduce locally adaptive MCMC methods that

exploit the natural underlying Riemannian geometry of many statistical

models [1].Such algorithms automatically adapt to the local

correlation structure of the model parameters when simulating paths

across the manifold, providing highly efficient convergence and

exploration of the target density for many classes of models. I will

provide examples of Bayesian inference using these methods on a

variety of models including logistic regression, log-Gaussian Cox

point processes, stochastic volatility models and Bayesian estimation

of dynamical systems described by nonlinear differential equations.

I will then discuss some very recent research in this area, which

extends the applicability of Riemannian Manifold MCMC methods to

statistical models that do not admit an analytically computable metric

tensor. [1] demonstrate the application of this algorithm for

inferring the parameters of a realistic system of highly nonlinear

ordinary differential equations using a biologically motivated robust

Student-t error model, for which the expected Fisher Information is

analytically intractable.

I will conclude with an overview of the outstanding opportunities and

challenges that lie ahead at this vibrant intersection between

differential geometry and Monte Carlo methodology.

[1] M. Girolami and B. Calderhead, Riemann Manifold Langevin and

Hamiltonian Monte Carlo Methods (with discussion), Journal of the

Royal Statistical Society: Series B (Statistical Methodology),

73:123-214, 2011