Construction of nonparametric Bayesian models from parametric Bayes equations

We consider the general problem of constructing nonparametric Bayesian models on infinite-dimensional random objects, such as functions, infinite graphs or infinite permutations. The problem has generated much interest in machine learning, where it is treated heuristically, but has not been studied in full generality in non-parametric Bayesian statistics, which tends to focus on models over probability distributions. Our approach applies a standard tool of stochastic process theory, the construction of stochastic processes from their finite-dimensional marginal distributions. The main contribution of the paper is a generalization of the classic Kolmogorov extension theorem to conditional probabilities. This extension allows a rigorous construction of nonparametric Bayesian models from systems of finite-dimensional, parametric Bayes equations. Using this approach, we show (i) how existence of a conjugate posterior for the nonparametric model can be guaranteed by choosing conjugate finite-dimensional models in the construction, (ii) how the mapping to the posterior parameters of the nonparametric model can be explicitly determined, and (iii) that the construction of conjugate models in essence requires the finite-dimensional models to be in the exponential family. As an application of our constructive framework, we derive a model on infinite permutations, the nonparametric Bayesian analogue of a model recently proposed for the analysis of rank data.

Effects of preferential transport in turbulent bluff-body-stabilized lean premixed CH 4/air flames

Preferential species diffusion is known to have important effects on local flame structure in turbulent premixed flames, and differential diffusion of heat and mass can have significant effects on both local flame structure and global flame parameters, such as turbulent flame speed. However, models for turbulent premixed combustion normally assume that atomic mass fractions are conserved from reactants to fully burnt products. Experiments reported here indicate that this basic assumption may be incorrect for an important class of turbulent flames. Measurements of major species and temperature in the near field of turbulent, bluff-body stabilized, lean premixed methane-air flames (Le=0.98) reveal significant departures from expected conditional mean compositional structure in the combustion products as well as within the flame. Net increases exceeding 10% in the equivalence ratio and the carbon-to-hydrogen atom ratio are observed across the turbulent flame brush. Corresponding measurements across an unstrained laminar flame at similar equivalence ratio are in close agreement with calculations performed using Chemkin with the GRI 3.0 mechanism and multi-component transport, confirming accuracy of experimental techniques. Results suggest that the large effects observed in the turbulent bluff-body burner are cause by preferential transport of H 2 and H 2O through the preheat zone ahead of CO 2 and CO, followed by convective transport downstream and away from the local flame brush. This preferential transport effect increases with increasing velocity of reactants past the bluff body and is apparently amplified by the presence of a strong recirculation zone where excess CO 2 is accumulated. © 2011 The Combustion Institute.

Forecasting high-frequency futures returns using online langevin dynamics

Forecasting the returns of assets at high frequency is the key challenge for high-frequency algorithmic trading strategies. In this paper, we propose a jump-diffusion model for asset price movements that models price and its trend and allows a momentum strategy to be developed. Conditional on jump times, we derive closed-form transition densities for this model. We show how this allows us to extract a trend from high-frequency finance data by using a Rao-Blackwellized variable rate particle filter to filter incoming price data. Our results show that even in the presence of transaction costs our algorithm can achieve a Sharpe ratio above 1 when applied across a portfolio of 75 futures contracts at high frequency. © 2011 IEEE.

Combining information sources for confidence estimation with CRF models

Obtaining accurate confidence measures for automatic speech recognition (ASR) transcriptions is an important task which stands to benefit from the use of multiple information sources. This paper investigates the application of conditional random field (CRF) models as a principled technique for combining multiple features from such sources. A novel method for combining suitably defined features is presented, allowing for confidence annotation using lattice-based features of hypotheses other than the lattice 1-best. The resulting framework is applied to different stages of a state-of-the-art large vocabulary speech recognition pipeline, and consistent improvements are shown over a sophisticated baseline system. Copyright © 2011 ISCA.

Gaussian Process Vine Copulas for Multivariate Dependence

Copulas allow to learn marginal distributions separately from the multivariate dependence structure (copula) that links them together into a density function. Vine factorizations ease the learning of high-dimensional copulas by constructing a hierarchy of conditional bivariate copulas. However, to simplify inference, it is common to assume that each of these conditional bivariate copulas is independent from its conditioning variables. In this paper, we relax this assumption by discovering the latent functions that specify the shape of a conditional copula given its conditioning variables We learn these functions by following a Bayesian approach based on sparse Gaussian processes with expectation propagation for scalable, approximate inference. Experiments on real-world datasets show that, when modeling all conditional dependencies, we obtain better estimates of the underlying copula of the data.

Complex chemistry DNS of n-heptane spray autoignition at high pressure and intermediate temperature conditions

Direct Numerical Simulations (DNS) of turbulent n-heptane sprays autoigniting at high pressure (P=24bar) and intermediate air temperature (Tair=1000K) have been performed to investigate the physical mechanisms present under conditions where low-temperature chemistry is expected to be important. The initial turbulence in the carrier gas, the global equivalence ratio in the spray region, and the initial droplet size distribution of the spray were varied. Results show that spray ignition exhibits a spotty nature, with several kernels developing independently in those regions where the mixture fraction is close to its most reactive value ξMR (as determined from homogeneous reactor calculations) and the scalar dissipation rate is low. Turbulence reduces the ignition delay time as it promotes mixing between air and the fuel vapor, eventually resulting in lower values of scalar dissipation. High values of the global equivalence ratio are responsible for a larger number of ignition kernels, due to the higher probability of finding regions where ξ=ξMR. Spray polydispersity results in the occurrence of ignition over a wider range of mixture fraction values. This is a consequence of the inhomogeneities in the mixing field that characterize these sprays, where poorly mixed rich spots are seen to alternate with leaner ones which are well-mixed. The DNS simulations presented in this work have also been used to assess the applicability of the Conditional Moment Closure (CMC) method to the simulation of spray combustion. CMC is found to be a valid method for capturing spray autoignition, although care should be taken in the modelling of the unclosed terms appearing in the CMC equations. © 2013 The Combustion Institute.

Approximate Bayesian Computation for Smoothing

We consider a method for approximate inference in hidden Markov models (HMMs). The method circumvents the need to evaluate conditional densities of observations given the hidden states. It may be considered an instance of Approximate Bayesian Computation (ABC) and it involves the introduction of auxiliary variables valued in the same space as the observations. The quality of the approximation may be controlled to arbitrary precision through a parameter ε > 0. We provide theoretical results which quantify, in terms of ε, the ABC error in approximation of expectations of additive functionals with respect to the smoothing distributions. Under regularity assumptions, this error is, where n is the number of time steps over which smoothing is performed. For numerical implementation, we adopt the forward-only sequential Monte Carlo (SMC) scheme of [14] and quantify the combined error from the ABC and SMC approximations. This forms some of the first quantitative results for ABC methods which jointly treat the ABC and simulation errors, with a finite number of data and simulated samples. © Taylor & Francis Group, LLC.

M-type channels selectively control bursting in rat dopaminergic neurons

Midbrain dopaminergic neurons in the substantia nigra, pars compacta and ventral tegmental area are critically important in many physiological functions. These neurons exhibit firing patterns that include tonic slow pacemaking, irregular firing and bursting, and the amount of dopamine that is present in the synaptic cleft is much increased during bursting. The mechanisms responsible for the switch between these spiking patterns remain unclear. Using both in-vivo recordings combined with microiontophoretic or intraperitoneal drug applications and in-vitro experiments, we have found that M-type channels, which are present in midbrain dopaminergic cells, modulate the firing during bursting without affecting the background low-frequency pacemaker firing. Thus, a selective blocker of these channels, 10,10-bis(4-pyridinylmethyl)-9(10H)- anthracenone dihydrochloride, specifically potentiated burst firing. Computer modeling of the dopamine neuron confirmed the possibility of a differential influence of M-type channels on excitability during various firing patterns. Therefore, these channels may provide a novel target for the treatment of dopamine-related diseases, including Parkinson's disease and drug addiction. Moreover, our results demonstrate that the influence of M-type channels on the excitability of these slow pacemaker neurons is conditional upon their firing pattern. © 2010 Federation of European Neuroscience Societies and Blackwell Publishing Ltd.

Bayesian Structured Prediction Using Gaussian Processes

We introduce a conceptually novel structured prediction model, GPstruct, which is kernelized, non-parametric and Bayesian, by design. We motivate the model with respect to existing approaches, among others, conditional random fields (CRFs), maximum margin Markov networks (M3N), and structured support vector machines (SVMstruct), which embody only a subset of its properties. We present an inference procedure based on Markov Chain Monte Carlo. The framework can be instantiated for a wide range of structured objects such as linear chains, trees, grids, and other general graphs. As a proof of concept, the model is benchmarked on several natural language processing tasks and a video gesture segmentation task involving a linear chain structure. We show prediction accuracies for GPstruct which are comparable to or exceeding those of CRFs and SVMstruct.

A confidence-based approach for improving keyword hypothesis scores

The task in keyword spotting (KWS) is to hypothesise times at which any of a set of key terms occurs in audio. An important aspect of such systems are the scores assigned to these hypotheses, the accuracy of which have a significant impact on performance. Estimating these scores may be formulated as a confidence estimation problem, where a measure of confidence is assigned to each key term hypothesis. In this work, a set of discriminative features is defined, and combined using a conditional random field (CRF) model for improved confidence estimation. An extension to this model to directly address the problem of score normalisation across key terms is also introduced. The implicit score normalisation which results from applying this approach to separate systems in a hybrid configuration yields further benefits. Results are presented which show notable improvements in KWS performance using the techniques presented in this work. © 2013 IEEE.

Scalar gradient behaviour in MILD combustion

The results of three-dimensional Direct Numerical Simulation (DNS) of Moderate, Intense Low-oxygen Dilution (MILD) and conventional premixed turbulent combustion conducted using a skeletal mechanism including the effects of non-unity Lewis numbers and temperature dependent transport properties are analysed to investigate combustion characteristics using scalar gradient information. The DNS data is also used to synthesise laser induced fluorescence (LIF) signals of OH, CH2O, and CHO. These signals are analysed to verify if they can be used to study turbulent MILD combustion and it has been observed that at least two (OH and CH2O) LIF signals are required since the OH increase across the reaction zone is smaller in MILD combustion compared to premixed combustion. The scalar gradient PDFs conditioned on the reaction rate obtained from the DNS data and synthesised LIF signals suggests a strong gradient in the direction normal to the MILD reaction zone with moderate reaction rate implying flamelet combustion. However, the PDF of the normal gradient is as broad as for the tangential gradient when the reaction rate is high. This suggests a non-flamelet behaviour, which is due to interaction of reaction zones. The analysis of the conditional PDFs for the premixed case confirms the expected behaviour of scalar gradient in flamelet combustion. It has been shown that the LIF signals synthesised using 2D slices of DNS data also provide very similar insights. These results demonstrate that the so-called flameless combustion is not an idealised homogeneous reactive mixture but has common features of conventional combustion while containing distinctive characteristics. © 2013 The Combustion Institute.

Influence of turbulence-chemistry interaction for n-heptane spray combustion under diesel engine conditions with emphasis on soot formation and oxidation

The influence of the turbulence-chemistry interaction (TCI) for n-heptane sprays under diesel engine conditions has been investigated by means of computational fluid dynamics (CFD) simulations. The conditional moment closure approach, which has been previously validated thoroughly for such flows, and the homogeneous reactor (i.e. no turbulent combustion model) approach have been compared, in view of the recent resurgence of the latter approaches for diesel engine CFD. Experimental data available from a constant-volume combustion chamber have been used for model validation purposes for a broad range of conditions including variations in ambient oxygen (8-21% by vol.), ambient temperature (900 and 1000 K) and ambient density (14.8 and 30 kg/m3). The results from both numerical approaches have been compared to the experimental values of ignition delay (ID), flame lift-off length (LOL), and soot volume fraction distributions. TCI was found to have a weak influence on ignition delay for the conditions simulated, attributed to the low values of the scalar dissipation relative to the critical value above which auto-ignition does not occur. In contrast, the flame LOL was considerably affected, in particular at low oxygen concentrations. Quasi-steady soot formation was similar; however, pronounced differences in soot oxidation behaviour are reported. The differences were further emphasised for a case with short injection duration: in such conditions, TCI was found to play a major role concerning the soot oxidation behaviour because of the importance of soot-oxidiser structure in mixture fraction space. Neglecting TCI leads to a strong over-estimation of soot oxidation after the end of injection. The results suggest that for some engines, and for some phenomena, the neglect of turbulent fluctuations may lead to predictions of acceptable engineering accuracy, but that a proper turbulent combustion model is needed for more reliable results. © 2014 Taylor & Francis.

Classification using log Gaussian Cox processes

McCullagh and Yang (2006) suggest a family of classification algorithms based on Cox processes. We further investigate the log Gaussian variant which has a number of appealing properties. Conditioned on the covariates, the distribution over labels is given by a type of conditional Markov random field. In the supervised case, computation of the predictive probability of a single test point scales linearly with the number of training points and the multiclass generalization is straightforward. We show new links between the supervised method and classical nonparametric methods. We give a detailed analysis of the pairwise graph representable Markov random field, which we use to extend the model to semi-supervised learning problems, and propose an inference method based on graph min-cuts. We give the first experimental analysis on supervised and semi-supervised datasets and show good empirical performance.

Quantum master-equation approach to quantum transport through mesoscopic systems

For quantum transport through mesoscopic systems, a quantum master-equation approach is developed in terms of compact expressions for the transport current and the reduced density matrix of the system. The present work is an extension of Gurvitz's approach for quantum transport and quantum measurement, namely, to finite temperature and arbitrary bias voltage. Our derivation starts from a second-order cumulant expansion of the tunneling Hamiltonian; then follows the conditional average over the electrode reservoir states. As a consequence, in the usual weak-tunneling regime, the established formalism is applicable for a wide range of transport problems. The validity of the formalism and its convenience in application are well illustrated by a number of examples.

Spontaneous relaxation of a charge qubit under electrical measurement

In this work we first derive a generalized conditional master equation for quantum measurement by a mesoscopic detector, then study the readout characteristics of qubit measurement where a number of remarkable new features are found. The work would, in particular, highlight the qubit spontaneous relaxation effect induced by the measurement itself rather than an external thermal bath.