arXiv subjects

Radford M. Neal

Publications and source records attributed to Radford M. Neal.

34 records · Page 2Linked to original sources

arXiv2010

Slice Sampling with Adaptive Multivariate Steps: The Shrinking-Rank Method

The shrinking rank method is a variation of slice sampling that is efficient at sampling from multivariate distributions with highly correlated parameters. It requires that the gradient of the log-density be computable. At each individual step, it approximates the current slice with a Gaussian occupying a shrinking-dimension subspace. The dimension of the approximation is shrunk orthogonally to the gradient at rejected proposals, since the gradients at points outside the current slice tend to point towards the slice. This causes the proposal distribution to converge rapidly to an estimate of the longest axis of the slice, resulting in states that are less correlated than those generated by related methods. After describing the method, we compare it to two other methods on several distributions and obtain favorable results.

stat.CO↗

arXiv2010

Covariance-Adaptive Slice Sampling

We describe two slice sampling methods for taking multivariate steps using the crumb framework. These methods use the gradients at rejected proposals to adapt to the local curvature of the log-density surface, a technique that can produce much better proposals when parameters are highly correlated. We evaluate our methods on four distributions and compare their performance to that of a non-adaptive slice sampling method and a Metropolis method. The adaptive methods perform favorably on low-dimensional target distributions with highly-correlated parameters.

stat.CO↗

arXiv2007

A Method for Compressing Parameters in Bayesian Models with Application to Logistic Sequence Prediction Models

Bayesian classification and regression with high order interactions is largely infeasible because Markov chain Monte Carlo (MCMC) would need to be applied with a great many parameters, whose number increases rapidly with the order. In this paper we show how to make it feasible by effectively reducing the number of parameters, exploiting the fact that many interactions have the same values for all training cases. Our method uses a single ``compressed'' parameter to represent the sum of all parameters associated with a set of patterns that have the same value for all training cases. Using symmetric stable distributions as the priors of the original parameters, we can easily find the priors of these compressed parameters. We therefore need to deal only with a much smaller number of compressed parameters when training the model with MCMC. The number of compressed parameters may have converged before considering the highest possible order. After training the model, we can split these compressed parameters into the original ones as needed to make predictions for test cases. We show in detail how to compress parameters for logistic sequence prediction models. Experiments on both simulated and real data demonstrate that a huge number of parameters can indeed be reduced by our compression method.

stat.ML↗

arXiv2007

Nonlinear Models Using Dirichlet Process Mixtures

We introduce a new nonlinear model for classification, in which we model the joint distribution of response variable, y, and covariates, x, non-parametrically using Dirichlet process mixtures. We keep the relationship between y and x linear within each component of the mixture. The overall relationship becomes nonlinear if the mixture contains more than one component. We use simulated data to compare the performance of this new approach to a simple multinomial logit (MNL) model, an MNL model with quadratic terms, and a decision tree model. We also evaluate our approach on a protein fold classification problem, and find that our model provides substantial improvement over previous methods, which were based on Neural Networks (NN) and Support Vector Machines (SVM). Folding classes of protein have a hierarchical structure. We extend our method to classification problems where a class hierarchy is available. We find that using the prior information regarding the hierarchical structure of protein folds can result in higher predictive accuracy.

math.ST↗

arXiv2007

A Method for Avoiding Bias from Feature Selection with Application to Naive Bayes Classification Models

For many classification and regression problems, a large number of features are available for possible use - this is typical of DNA microarray data on gene expression, for example. Often, for computational or other reasons, only a small subset of these features are selected for use in a model, based on some simple measure such as correlation with the response variable. This procedure may introduce an optimistic bias, however, in which the response variable appears to be more predictable than it actually is, because the high correlation of the selected features with the response may be partly or wholely due to chance. We show how this bias can be avoided when using a Bayesian model for the joint distribution of features and response. The crucial insight is that even if we forget the exact values of the unselected features, we should retain, and condition on, the knowledge that their correlation with the response was too small for them to be selected. In this paper we describe how this idea can be implemented for ``naive Bayes'' models of binary data. Experiments with simulated data confirm that this method avoids bias due to feature selection. We also apply the naive Bayes model to subsets of data relating gene expression to colon cancer, and find that correcting for bias from feature selection does improve predictive performance.

math.ST↗

arXiv2006

Puzzles of Anthropic Reasoning Resolved Using Full Non-indexical Conditioning

I consider the puzzles arising from four interrelated problems involving `anthropic' reasoning, and in particular the `Self-Sampling Assumption' (SSA) - that one should reason as if one were randomly chosen from the set of all observers in a suitable reference class. The problem of Freak Observers might appear to force acceptance of SSA if any empirical evidence is to be credited. The Sleeping Beauty problem arguably shows that one should also accept the `Self-Indication Assumption' (SIA) - that one should take one's own existence as evidence that the number of observers is more likely to be large than small. But this assumption produces apparently absurd results in the Presumptuous Philosopher problem. Without SIA, however, a definitive refutation of the counterintuitive Doomsday Argument seems difficult. I show that these problems are satisfyingly resolved by applying the principle that one should always condition on all evidence - not just on the fact that you are an intelligent observer, or that you are human, but on the fact that you are a human with a specific set of memories. This `Full Non-indexical Conditioning' (FNC) approach usually produces the same results as assuming both SSA and SIA, with a sufficiently broad reference class, while avoiding their ad hoc aspects. I argue that the results of FNC are correct using the device of hypothetical ``companion'' observers, whose existence clarifies what principles of reasoning are valid. I conclude by discussing how one can use FNC to infer how densely we should expect intelligent species to occur, and by examining recent anthropic arguments in inflationary and string theory cosmology.

math.ST↗

arXiv2006

Gene Function Classification Using Bayesian Models with Hierarchy-Based Priors

We investigate the application of hierarchical classification schemes to the annotation of gene function based on several characteristics of protein sequences including phylogenic descriptors, sequence based attributes, and predicted secondary structure. We discuss three Bayesian models and compare their performance in terms of predictive accuracy. These models are the ordinary multinomial logit (MNL) model, a hierarchical model based on a set of nested MNL models, and a MNL model with a prior that introduces correlations between the parameters for classes that are nearby in the hierarchy. We also provide a new scheme for combining different sources of information. We use these models to predict the functional class of Open Reading Frames (ORFs) from the E. coli genome. The results from all three models show substantial improvement over previous methods, which were based on the C5 algorithm. The MNL model using a prior based on the hierarchy outperforms both the non-hierarchical MNL model and the nested MNL model. In contrast to previous attempts at combining these sources of information, our approach results in a higher accuracy rate when compared to models that use each data source alone. Together, these results show that gene function can be predicted with higher accuracy than previously achieved, using Bayesian models that incorporate suitable prior information.

q-bio.GN↗

arXiv2005

Estimating Ratios of Normalizing Constants Using Linked Importance Sampling

Ratios of normalizing constants for two distributions are needed in both Bayesian statistics, where they are used to compare models, and in statistical physics, where they correspond to differences in free energy. Two approaches have long been used to estimate ratios of normalizing constants. The `simple importance sampling' (SIS) or `free energy perturbation' method uses a sample drawn from just one of the two distributions. The `bridge sampling' or `acceptance ratio' estimate can be viewed as the ratio of two SIS estimates involving a bridge distribution. For both methods, difficult problems must be handled by introducing a sequence of intermediate distributions linking the two distributions of interest, with the final ratio of normalizing constants being estimated by the product of estimates of ratios for adjacent distributions in this sequence. Recently, work by Jarzynski, and independently by Neal, has shown how one can view such a product of estimates, each based on simple importance sampling using a single point, as an SIS estimate on an extended state space. This `Annealed Importance Sampling' (AIS) method produces an exactly unbiased estimate for the ratio of normalizing constants even when the Markov transitions used do not reach equilibrium. In this paper, I show how a corresponding `Linked Importance Sampling' (LIS) method can be constructed in which the estimates for individual ratios are similar to bridge sampling estimates. I show empirically that for some problems, LIS estimates are much more accurate than AIS estimates found using the same computation time, although for other problems the two methods have similar performance. Linked sampling methods similar to LIS are useful for other purposes as well.

math.ST↗

arXiv2005

Improving Classification When a Class Hierarchy is Available Using a Hierarchy-Based Prior

We introduce a new method for building classification models when we have prior knowledge of how the classes can be arranged in a hierarchy, based on how easily they can be distinguished. The new method uses a Bayesian form of the multinomial logit (MNL, a.k.a. ``softmax'') model, with a prior that introduces correlations between the parameters for classes that are nearby in the tree. We compare the performance on simulated data of the new method, the ordinary MNL model, and a model that uses the hierarchy in different way. We also test the new method on a document labelling problem, and find that it performs better than the other methods, particularly when the amount of training data is small.

math.ST↗

arXiv2005

The Short-Cut Metropolis Method

I show how one can modify the random-walk Metropolis MCMC method in such a way that a sequence of modified Metropolis updates takes little computation time when the rejection rate is outside a desired interval. This allows one to effectively adapt the scale of the Metropolis proposal distribution, by performing several such "short-cut" Metropolis sequences with varying proposal stepsizes. Unlike other adaptive Metropolis schemes, this method converges to the correct distribution in the same fashion as the standard Metropolis method.

math.ST↗

arXiv2005

Taking Bigger Metropolis Steps by Dragging Fast Variables

I show how Markov chain sampling with the Metropolis-Hastings algorithm can be modified so as to take bigger steps when the distribution being sampled from has the characteristic that its density can be quickly recomputed for a new point if this point differs from a previous point only with respect to a subset of 'fast' variables. I show empirically that when using this method, the efficiency of sampling for the remaining 'slow' variables can approach what would be possible using Metropolis updates based on the marginal distribution for the slow variables.

math.ST↗

arXiv2004

Improving Asymptotic Variance of MCMC Estimators: Non-reversible Chains are Better

I show how any reversible Markov chain on a finite state space that is irreducible, and hence suitable for estimating expectations with respect to its invariant distribution, can be used to construct a non-reversible Markov chain on a related state space that can also be used to estimate these expectations, with asymptotic variance at least as small as that using the reversible chain (typically smaller). The non-reversible chain achieves this improvement by avoiding (to the extent possible) transitions that backtrack to the state from which the chain just came. The proof that this modification cannot increase the asymptotic variance of an MCMC estimator uses a new technique that can also be used to prove Peskun's (1973) theorem that modifying a reversible chain to reduce the probability of staying in the same state cannot increase asymptotic variance. A non-reversible chain that avoids backtracking will often take little or no more computation time per transition than the original reversible chain, and can sometime produce a large reduction in asymptotic variance, though for other chains the improvement is slight. In addition to being of some practical interest, this construction demonstrates that non-reversible chains have a fundamental advantage over reversible chains for MCMC estimation. Research into better MCMC methods may therefore best be focused on non-reversible chains.

math.PR↗

arXiv2003

Markov Chain Sampling for Non-linear State Space Models Using Embedded Hidden Markov Models

I describe a new Markov chain method for sampling from the distribution of the state sequences in a non-linear state space model, given the observation sequence. This method updates all states in the sequence simultaneously using an embedded Hidden Markov model (HMM). An update begins with the creation of a ``pool'' of K states at each time, by applying some Markov chain update to the current state. These pools define an embedded HMM whose states are indexes within this pool. Using the forward-backward dynamic programming algorithm, we can then efficiently choose a state sequence at random with the appropriate probabilities from the exponentially large number of state sequences that pass through states in these pools. I show empirically that when states at nearby times are strongly dependent, embedded HMM sampling can perform better than Metropolis methods that update one state at a time.

math.PR↗

arXiv2000

Slice Sampling

Markov chain sampling methods that automatically adapt to characteristics of the distribution being sampled can be constructed by exploiting the principle that one can sample from a distribution by sampling uniformly from the region under the plot of its density function. A Markov chain that converges to this uniform distribution can be constructed by alternating uniform sampling in the vertical direction with uniform sampling from the horizontal `slice' defined by the current vertical position, or more generally, with some update that leaves the uniform distribution over this slice invariant. Variations on such `slice sampling' methods are easily implemented for univariate distributions, and can be used to sample from a multivariate distribution by updating each variable in turn. This approach is often easier to implement than Gibbs sampling, and more efficient than simple Metropolis updates, due to the ability of slice sampling to adaptively choose the magnitude of changes made. It is therefore attractive for routine and automated use. Slice sampling methods that update all variables simultaneously are also possible. These methods can adaptively choose the magnitudes of changes made to each variable, based on the local properties of the density function. More ambitiously, such methods could potentially allow the sampling to adapt to dependencies between variables by constructing local quadratic approximations. Another approach is to improve sampling efficiency by suppressing random walks. This can be done using `overrelaxed' versions of univariate slice sampling procedures, or by using `reflective' multivariate slice sampling methods, which bounce off the edges of the slice.

physics.data-an↗

arXiv1998

Annealed Importance Sampling

Simulated annealing - moving from a tractable distribution to a distribution of interest via a sequence of intermediate distributions - has traditionally been used as an inexact method of handling isolated modes in Markov chain samplers. Here, it is shown how one can use the Markov chain transitions for such an annealing sequence to define an importance sampler. The Markov chain aspect allows this method to perform acceptably even for high-dimensional problems, where finding good importance sampling distributions would otherwise be very difficult, while the use of importance weights ensures that the estimates found converge to the correct values as the number of annealing runs increases. This annealed importance sampling procedure resembles the second half of the previously-studied tempered transitions, and can be seen as a generalization of a recently-proposed variant of sequential importance sampling. It is also related to thermodynamic integration methods for estimating ratios of normalizing constants. Annealed importance sampling is most attractive when isolated modes are present, or when estimates of normalizing constants are required, but it may also be more generally useful, since its independent sampling allows one to bypass some of the problems of assessing convergence and autocorrelation in Markov chain samplers.

physics.comp-ph↗

arXiv1997

Monte Carlo Implementation of Gaussian Process Models for Bayesian Regression and Classification

Gaussian processes are a natural way of defining prior distributions over functions of one or more input variables. In a simple nonparametric regression problem, where such a function gives the mean of a Gaussian distribution for an observed response, a Gaussian process model can easily be implemented using matrix computations that are feasible for datasets of up to about a thousand cases. Hyperparameters that define the covariance function of the Gaussian process can be sampled using Markov chain methods. Regression models where the noise has a t distribution and logistic or probit models for classification applications can be implemented by sampling as well for latent values underlying the observations. Software is now available that implements these methods using covariance functions with hierarchical parameterizations. Models defined in this way can discover high-level properties of the data, such as which inputs are relevant to predicting the response.

physics.data-an↗