Abstract not found

The ¨generic viewpointässumption states that an observer is not in a special position relative to the scene. It is commonly used to disqualify scene interpretations that assume special viewpoints, following a binary decision that the viewpoint was either generic or accidental. In this paper, we apply Bayesian statistics to quantify the probability of a view, and so derive a useful tool to estimate scene parameters. This approach may increase the scope and accuracy of scene estimates. It applies to a range of vision problems. We show shape from shading examples, where we rank shapes or reflectance functions in cases where these are otherwise unknown. The rankings agree with the perceived values.

Vision algorithms are often developed in a Bayesian framework. Two estimators are commonly used: maximum a posteriori (MAP), and minimum mean squared error (MMSE). We argue that neither is appropriate for perception problems. The MAP estimator makes insufficient use of structure in the posterior probability. The squared error penalty of the MMSE estimator does not reflect typical penalties. We describe a new

this paper I (1) examine three levels of inferential strength supported by typical social science data-gathering methods, and call for a greater degree of explicitness, when HMs and other models are applied, in identifying which level is appropriate; (2) reconsider the use of HMs in school effectiveness studies and meta-analysis from the perspective of causal inference; and (3) recommend the increased use of Gibbs sampling and other Markov-chain Monte Carlo (MCMC) methods in the application of HMs in the social sciences, so that comparisons between MCMC and better-established fitting methods---including full or restricted maximum likelihood estimation based on the EM algorithm, Fisher scoring or iterative generalized least squares---may be more fully informed by empirical practice.

Despite the widespread use of key concepts of the Neyman–Pearson (N–P) statistical paradigm—type I and II errors, significance levels, power, confidence levels—they have been the subject of philosophical controversy and debate for over 60 years. Both current and long-standing problems of N–P tests stem from unclarity and confusion, even among N–P adherents, as to how a test’s (pre-data) error probabilities are to be used for (post-data) inductive inference as opposed to inductive behavior. We argue that the relevance of error probabilities is to ensure that only statistical hypotheses that have passed severe or probative tests are inferred from the data. The severity criterion supplies a meta-statistical principle for evaluating proposed statistical inferences, avoiding classic fallacies from tests that are overly sensitive, as well as those not sensitive enough to particular errors and discrepancies.

The principle of maximum entropy is a method for assigning values to probability distributions on the basis of partial information. In usual formulations of this and related methods of inference one assumes that this partial information takes the form of a constraint on allowed probability distributions. In practical applications, however, the information consists of empirical data. A constraint rule is then employed to construct constraints on probability distributions out of these data. Usually one adopts the rule to equate the expectation values of certain functions with their empirical averages. There are, however, various other ways in which one can construct constraints from empirical data, which makes the maximum entropy principle lead to very different probability assignments. This paper shows that an argument by Jaynes to justify the usual constraint rule is unsatisfactory and investigates several alternative choices. The choice of a constraint rule is also show...

Summary. Discrete data, particularly count and contingency table data, are typically analyzed using methods that are accurate to first order, such as normal approximations for maximum likelihood estimators. By contrast continuous data can quite generally be analyzed using third order procedures, with major improvements in accuracy and with intrinsic separation of information concerning parameter components. This paper extends these higher order results to discrete data, yielding a methodology that is widely applicable and accurate to second order. The extension can be described in terms of an approximating exponential model expressed in terms of a score variable. The development is outlined and the flexibility of the approach illustrated by examples. 1.

One of the major criticisms of probabilistic risk assessment is that the requisite input distributions are often not available. Several approaches to this problem have been suggested, including creating a library of standard empirically fitted distributions, employing maximum entropy criteria to synthesize distributions from a priori constraints, and even using `default' inputs such as the triangular distribution. Since empirical information is often sparse, analysts commonly must make assumptions to select the input distributions without empirical justification. This practice diminishes the credibility of the assessment and any decisions based on it. There is no absolute necessity, however, of assuming particular shapes for input distributions in probabilistic risk assessments. It is possible to make the needed calculations using inputs specified only as bounds on probability distributions. We describe such bounds for a variety of circumstances where empirical information is extremely limited, and illustrate how these bounds can be used in computations to represent uncertainty about input distributions far more comprehensively than is possible with current approaches.

The Dempster-Shafer (DS) theory of probabilistic reasoning is presented in terms of a semantics whereby every meaningful formal assertion is associated with a triple (p, q, r) where p is the probability “for ” the assertion, q is the probability “against” the assertion, and r is the probability of “don’t know”. Arguments are presented for the necessity of “don’t know”. Elements of the calculus are sketched, including the extension of a DS model from a margin to a full state space, and DS combination of independent DS uncertainty assessments on the full space. The methodology is applied to inference and prediction from Poisson counts, including an introduction to the use of join-tree model structure to simplify and shorten computation. The relation of DS theory to statistical significance testing is elaborated, introducing along the way the new concept of “dull ” null hypothesis. Key words: Dempster-Shafer; belief functions; state space; Poisson model; join-tree computation; statistical significance; dull null hypothesis 1

Conformal prediction uses past experience to determine precise levels of confidence in new predictions. Given an error probability ε, together with a method that makes a prediction ˆy of a label y, it produces a set of labels, typically containing ˆy, that also contains y with probability 1 − ε. Conformal prediction can be applied to any method for producing ˆy: a nearest-neighbor method, a support-vector machine, ridge regression, etc. Conformal prediction is designed for an on-line setting in which labels are predicted successively, each one being revealed before the next is predicted. The most novel and valuable feature of conformal prediction is that if the successive examples are sampled independently from the same distribution, then the successive predictions will be right 1 − ε of the time, even though they are based on an accumulating data set rather than on independent data sets. In addition to the model under which successive examples are sampled independently, other on-line compression models can also use conformal prediction. The widely used Gaussian linear model is one of these. This tutorial presents a self-contained account of the theory of conformal prediction and works through several numerical examples. A more comprehensive treatment of the topic is provided in