Asymptotically exact inference in differentiable generative models

Many generative models can be expressed as a differentiable function of random inputs drawn from some simple probability density. This framework includes both deep generative architectures such as Variational Autoencoders and a large class of procedurally defined simulator models. We present a method for performing efficient MCMC inference in such models when conditioning on observations of the model output. For some models this offers an asymptotically exact inference method where Approximate Bayesian Computation might otherwise be employed. We use the intuition that inference corresponds to integrating a density across the manifold corresponding to the set of inputs consistent with the observed outputs. This motivates the use of a constrained variant of Hamiltonian Monte Carlo which leverages the smooth geometry of the manifold to coherently move between inputs exactly consistent with observations. We validate the method by performing inference tasks in a diverse set of models.

[1]  Tianqi Chen,et al.  Stochastic Gradient Hamiltonian Monte Carlo , 2014, ICML.

[2]  Jean-Michel Marin,et al.  Approximate Bayesian computational methods , 2011, Statistics and Computing.

[3]  R. Wilkinson Approximate Bayesian computation (ABC) gives exact results under the assumption of model error , 2008, Statistical applications in genetics and molecular biology.

[4]  T. O’Neil Geometric Measure Theory , 2002 .

[5]  M. Plummer,et al.  CODA: convergence diagnosis and output analysis for MCMC , 2006 .

[6]  Robert Leenders,et al.  Hamiltonian ABC , 2015, UAI.

[7]  Iain Murray,et al.  Pseudo-Marginal Slice Sampling , 2015, AISTATS.

[8]  Michael Betancourt,et al.  The Fundamental Incompatibility of Scalable Hamiltonian Monte Carlo and Naive Data Subsampling , 2015, ICML.

[9]  Max Welling,et al.  Optimization Monte Carlo: Efficient and Embarrassingly Parallel Likelihood-Free Inference , 2015, NIPS.

[10]  S. Duane,et al.  Hybrid Monte Carlo , 1987 .

[11]  Ryan P. Adams,et al.  Gradient-based Hyperparameter Optimization through Reversible Learning , 2015, ICML.

[12]  Benedict Leimkuhler,et al.  Efficient molecular dynamics using geodesic integration and solvent–solute splitting , 2016, Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences.

[13]  B. Leimkuhler,et al.  Symplectic Numerical Integrators in Constrained Hamiltonian Systems , 1994 .

[14]  Eric Barth,et al.  Algorithms for constrained molecular dynamics , 1995, J. Comput. Chem..

[15]  Radford M. Neal MCMC Using Hamiltonian Dynamics , 2011, 1206.1901.

[16]  Michael J. Black,et al.  Pose-conditioned joint angle limits for 3D human pose reconstruction , 2015, 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).

[17]  Philip Rabinowitz,et al.  Numerical methods for nonlinear algebraic equations , 1970 .

[18]  Raquel Urtasun,et al.  A Family of MCMC Methods on Implicitly Defined Manifolds , 2012, AISTATS.

[19]  Hugo Larochelle,et al.  A Deep and Tractable Density Estimator , 2013, ICML.

[20]  H. C. Andersen Rattle: A “velocity” version of the shake algorithm for molecular dynamics calculations , 1983 .

[21]  P. Diaconis,et al.  Sampling From A Manifold , 2012, 1206.6913.

[22]  John Salvatier,et al.  Theano: A Python framework for fast computation of mathematical expressions , 2016, ArXiv.

[23]  Yee Whye Teh,et al.  Bayesian Learning via Stochastic Gradient Langevin Dynamics , 2011, ICML.

[24]  Christophe Andrieu,et al.  Model criticism based on likelihood-free inference, with an application to protein network evolution , 2009, Proceedings of the National Academy of Sciences.

[25]  Iain Murray,et al.  Differentiation of the Cholesky decomposition , 2016, ArXiv.

[26]  Daan Wierstra,et al.  Stochastic Backpropagation and Approximate Inference in Deep Generative Models , 2014, ICML.

[27]  D. Balding,et al.  Approximate Bayesian computation in population genetics. , 2002, Genetics.

[28]  Yoshua Bengio,et al.  NICE: Non-linear Independent Components Estimation , 2014, ICLR.

[29]  Radford M. Neal Slice Sampling , 2003, The Annals of Statistics.

[30]  Edward Meeds,et al.  Optimization Monte Carlo: Efficient and Embarrassingly Parallel Likelihood-Free Inference , 2015, Neural Information Processing Systems.

[31]  Paul Marjoram,et al.  Markov chain Monte Carlo without likelihoods , 2003, Proceedings of the National Academy of Sciences of the United States of America.

[32]  Ernst Hairer,et al.  Simulating Hamiltonian dynamics , 2006, Math. Comput..

[33]  Gabriel Stoltz,et al.  Langevin dynamics with constraints and computation of free energy differences , 2010, Math. Comput..

[34]  Carsten Hartmann,et al.  An Ergodic Sampling Scheme for Constrained Hamiltonian Systems with Applications to Molecular Dynamics , 2008 .

[35]  Christian P. Robert,et al.  Model choice versus model criticism , 2009, Proceedings of the National Academy of Sciences.

[36]  M. Girolami,et al.  Geodesic Monte Carlo on Embedded Manifolds , 2013, Scandinavian journal of statistics, theory and applications.

[37]  Yoshua Bengio,et al.  Generative Adversarial Nets , 2014, NIPS.

[38]  Max Welling,et al.  Auto-Encoding Variational Bayes , 2013, ICLR.

[39]  Pieter Abbeel,et al.  Gradient Estimation Using Stochastic Computation Graphs , 2015, NIPS.

[40]  G. Ciccotti,et al.  Numerical Integration of the Cartesian Equations of Motion of a System with Constraints: Molecular Dynamics of n-Alkanes , 1977 .

[41]  Carsten Hartmann,et al.  A constrained hybrid Monte‐Carlo algorithm and the problem of calculating the free energy in several variables , 2005 .