DOCSLIB.ORG

Arxiv:1811.04251V4 [Cs.IT] 20 May 2020

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Arxiv:1811.04251V4 [Cs.IT] 20 May 2020 Formal Limitations on the Measurement of Mutual Information David McAllester Karl Stratos Toyota Technological Institute at Chicago Rutgers University Abstract larger than O(ln N) where N is the number of samples. Thus a meaningful high-confidence lower bound is infeasi- Measuring mutual information from finite data is ble when underlying mutual information is large (e.g., hun- difficult. Recent work has considered variational dreds of bits). methods maximizing a lower bound. In this pa- Our result corrects, generalizes, and unifies past work on per, we prove that serious statistical limitations measuring mutual information. Unlike previous intractabil- are inherent to any method of measuring mutual ity results that are tied to specific estimators (Gao et al., information. More specifically, we show that any 2015; Oord et al., 2018), our result is universal to all es- distribution-free high-confidence lower bound on timators. Hence we show that without making strong as- mutual information estimated from N samples sumptions on the population distribution, such as the small cannot be larger than O(ln N). support assumption in Valiant and Valiant (2011) and min- imax bounds (Jiao et al., 2015), it is generally impossible to guarantee an accurate estimate of mutual information. 1 INTRODUCTION Our results contradict a theorem in Belghazi et al. (2018) claiming a polynomial sample size convergence rate for a Mutual information has important applications to unsuper- mutual information estimator and we point out an error in vised learning. It underlies classical representation learn- their proof. ing methods such as Brown clustering (Brown et al., 1992), While it is infeasible to give meaningful high-confidence INFOMAX (Bell and Sejnowski, 1995), and the informa- lower bound guarantees for large mutual information, es- tion bottleneck method (Tishby et al., 1999). Maximiz- timators lacking formal guarantees might still be useful ing mutual information is also central to recent works on in practice. To this end, we propose expressing mutual unsupervised representation learning with neural networks information as a difference of entropies and estimating (McAllester, 2018; Belghazi et al., 2018; Oord et al., 2018; the entropy terms by cross-entropy upper bounds. This Stratos, 2019; Hjelm et al., 2019). difference-of-entropies (DoE) estimator gives neither an Unfortunately, measuring mutual information from finite upper bound nor a lower bound guarantee on mutual in- data is a notoriously difficult estimation problem. This dif- formation. Nevertheless, we give theoretical and empirical ficulty has motivated researchers to consider more compu- evidence that DoE can meaningfully estimate large mutual tationally amenable measurement methods that maximize information from feasible samples. a parameterized lower bound on mutual information. The arXiv:1811.04251v4 [cs.IT] 20 May 2020 hope is that the optimized value of the bound is close to 1.1 Overview the true value of mutual information and can serve as a good enough approximation. For instance, this is the ap- We state our main result below. We write (X; Y ) to denote proach advocated by Mutual Information Neural Estimator a pair of random variables with ranges (X ; Y). Unless oth- (MINE) (Belghazi et al., 2018) and contrastive predictive erwise specified, they can be discrete or continuous. For coding (CPC) (Oord et al., 2018). simplicity, all distributions are assumed to have full sup- port. Here we prove that serious statistical limitations are in- herent to all methods of measuring mutual information. Theorem 1.1. Let B be any mapping from N samples of More specifically, we show that any distribution-free high- (X; Y ) to a real number with the following property: for confidence lower bound on mutual information cannot be any distribution pXY over (X; Y ), given N iid samples (x1; y1) ::: (xN ; yN ) ∼ pXY , with probability at least 0.99 ≥ Proceedings of the 23rdInternational Conference on Artificial In- I(X; Y ; pXY ) B((x1; y1) ::: (xN ; yN )) telligence and Statistics (AISTATS) 2020, Palermo, Italy. PMLR: where I(X; Y ; pXY ) is the mutual information between Volume 108. Copyright 2020 by the author(s). (X; Y ) under pXY . Now pick any distribution qXY Formal Limitations on the Measurement of Mutual Information over (X; Y ). Then given N ≥ 50 iid samples 0 0 0 0 (x1; y1) ::: (xN ; yN ) ∼ qXY , with probability at least 0.96 N N ! N 1 1 0 X X f(xi) DVdf (pX jjqX ) := f(xi) − ln e (3) 0 0 0 0 N N B((x1; y1) ::: (xN ; yN )) ≤ 2 ln N + 5 i=1 i=1 An application of the DV bound is measuring KL diver- In the rest of the paper, we build toward this result by prov- gence. Specifically, we estimate KL divergence by esti- ing statistical limitations on measuring lower bounds on mating the associated DV bound from samples: Kullback-Leibler (KL) divergence and entropy. We derive jj ≥ jj several intermediate results, specifically: DKL(pX qX ) DVf (pX qX ) N & DVdf (pX jjqX ) • We first consider the Donsker-Varadhan lower bound on KL divergence which underlies the approach in MINE. The first inequality holds for all choices of f by Theo- We give an intuitive explanation on why estimating the rem 2.1. We require that the second inequality holds with bound from samples is problematic. We show that the high probability (with respect to random sampling). We polynomial sample complexity result given by Belghazi now show that this approach to measuring KL divergence et al. (2018) is incorrect. is problematic. • We formally prove that any lower bound on KL diver- gence cannot be larger than O(ln N). This subsumes the 2.1 Statistical Limitations on Measuring the DV Donsker-Varadhan bound as a special case. Bound • We formally prove that any lower bound on entropy can- A crucial observation is that the second term in the DV not be larger than O(ln N). Theorem 1.1 is a special bound (2) involves an expectation of the exponential case of this result. h f(x)i • We motivate measuring mutual information as a differ- E e ence of entropies, each of which measured by minimiz- x∼qX ing cross entropy. We empirically show that it outper- This expression has the same form as the moment gener- forms existing variational lower bounds in synthetic ex- ating function used in analyzing large deviation probabil- periments and produce realistic estimates of mutual in- ities. The utility of expectations of exponentials in large formation in real-world datasets. deviation theory is that such expressions can be dominated by extremely rare events (large deviations). The rare events 2 ISSUES WITH THE dominating the expectation will never be observed by sam- DONSKER-VARADHAN LOWER pling from qX . BOUND To quantitatively analyze the risk of unseen outlier events, we will make use of the following simple lemma where we The Donsker-Varadhan (DV) lower bound on KL diver- write pX (Φ[x]) for the probability over drawing x from pX gence is stated below. For completeness, a simple proof that the statement Φ[x] holds. is given in the supplementary material. Lemma 2.2 (Outlier risk lemma). Given N ≥ 2 samples from and a property such that ≤ , Theorem 2.1 (Donsker and Varadhan, 1983). Let pX and pX Φ[x] pX (Φ[x]) 1=N the probability that no sample satisfies is at least qX be distributions over X with finite KL divergence. Then x Φ[x] 1/4. for all bounded functions f : X! R DKL(pX jjqX ) ≥ DVf (pX jjqX ) (1) Proof. The probability that Φ[x] is unseen in the sample is at least (1 − 1=N)N which is at least 1=4 for N ≥ 2 and where N where limN!1(1 − 1=N) = 1=e. h f(x)i DVf (pX jjqX ) := E [f(x)] − ln E e (2) x∼pX x∼qX We can use the outlier risk lemma to perform a quantitative Moreover, (1) holds with equality for some f with range risk analysis of the DV bound. Assume without loss of X [0;F ]. generality that [0;Fmax] is the range of f taken on . Let max us consider the best case scenario in which the empirical The DV bound (2) computes the difference between the estimate (3) is the largest. It is easy to see that the largest expected value of f(x) under pX and the log of the ex- value is Fmax and attained when pected value of the exponential of f(x) under q . It can be X f(x ) = F 8i = 1 :::N ∼ i max easily estimated by sampling: given x1 : : : xN pX and 0 0 0 f(xi) = 0 8i = 1 :::N x1 : : : xN ∼ qX , we can compute the empirical estimate David McAllester, Karl Stratos 0 0 where x1 : : : xN are samples from pX and x1 : : : xN are lower bound on KL divergence (e.g., DKL(pX jjqX ) = samples from qX . But by the outlier risk lemma, there is supf>0 E [ln f(x)] − E [f(x)] + 1 in Nguyen et al. x∼pX x∼qX still at least a 1/4 probability that (2010) which does not involve an expectation of the expo- nential) that overcomes the limitation. h f(x)i 1 F E e ≥ e max x∼qX N We now strengthen the result by formally proving that no Since we require that (3) is a high-confidence lower bound lower bound on KL divergence estimated from samples can on the DV bound, it must account for the unseen outlier be large. We consider a more challenging setting in which risk.1 In particular, we must have we have perfect knowledge of pX (i.e., we can compute probabilities under the distribution) and only sample from F N e max qX to estimate a lower bound on DKL(pX jjqX ).
Load more

Recommended publications
  • On Measures of Entropy and Information
    On Measures of Entropy and Information Tech. Note 009 v0.7 http://threeplusone.com/info Gavin E. Crooks 2018-09-22 Contents 5 Csiszar´ f-divergences 12 Csiszar´ f-divergence ................ 12 0 Notes on notation and nomenclature 2 Dual f-divergence .................. 12 Symmetric f-divergences .............. 12 1 Entropy 3 K-divergence ..................... 12 Entropy ........................ 3 Fidelity ........................ 12 Joint entropy ..................... 3 Marginal entropy .................. 3 Hellinger discrimination .............. 12 Conditional entropy ................. 3 Pearson divergence ................. 14 Neyman divergence ................. 14 2 Mutual information 3 LeCam discrimination ............... 14 Mutual information ................. 3 Skewed K-divergence ................ 14 Multivariate mutual information ......... 4 Alpha-Jensen-Shannon-entropy .......... 14 Interaction information ............... 5 Conditional mutual information ......... 5 6 Chernoff divergence 14 Binding information ................ 6 Chernoff divergence ................. 14 Residual entropy .................. 6 Chernoff coefficient ................. 14 Total correlation ................... 6 Renyi´ divergence .................. 15 Lautum information ................ 6 Alpha-divergence .................. 15 Uncertainty coefficient ............... 7 Cressie-Read divergence .............. 15 Tsallis divergence .................. 15 3 Relative entropy 7 Sharma-Mittal divergence ............. 15 Relative entropy ................... 7 Cross entropy
    [Show full text]
  • Information Theory 1 Entropy 2 Mutual Information
    CS769 Spring 2010 Advanced Natural Language Processing Information Theory Lecturer: Xiaojin Zhu [email protected] In this lecture we will learn about entropy, mutual information, KL-divergence, etc., which are useful concepts for information processing systems. 1 Entropy Entropy of a discrete distribution p(x) over the event space X is X H(p) = − p(x) log p(x). (1) x∈X When the log has base 2, entropy has unit bits. Properties: H(p) ≥ 0, with equality only if p is deterministic (use the fact 0 log 0 = 0). Entropy is the average number of 0/1 questions needed to describe an outcome from p(x) (the Twenty Questions game). Entropy is a concave function of p. 1 1 1 1 7 For example, let X = {x1, x2, x3, x4} and p(x1) = 2 , p(x2) = 4 , p(x3) = 8 , p(x4) = 8 . H(p) = 4 bits. This definition naturally extends to joint distributions. Assuming (x, y) ∼ p(x, y), X X H(p) = − p(x, y) log p(x, y). (2) x∈X y∈Y We sometimes write H(X) instead of H(p) with the understanding that p is the underlying distribution. The conditional entropy H(Y |X) is the amount of information needed to determine Y , if the other party knows X. X X X H(Y |X) = p(x)H(Y |X = x) = − p(x, y) log p(y|x). (3) x∈X x∈X y∈Y From above, we can derive the chain rule for entropy: H(X1:n) = H(X1) + H(X2|X1) + ..
    [Show full text]
  • Information Theory and Maximum Entropy 8.1 Fundamentals of Information Theory
    NEU 560: Statistical Modeling and Analysis of Neural Data Spring 2018 Lecture 8: Information Theory and Maximum Entropy Lecturer: Mike Morais Scribes: 8.1 Fundamentals of Information theory Information theory started with Claude Shannon's A mathematical theory of communication. The first building block was entropy, which he sought as a functional H(·) of probability densities with two desired properties: 1. Decreasing in P (X), such that if P (X1) < P (X2), then h(P (X1)) > h(P (X2)). 2. Independent variables add, such that if X and Y are independent, then H(P (X; Y )) = H(P (X)) + H(P (Y )). These are only satisfied for − log(·). Think of it as a \surprise" function. Definition 8.1 (Entropy) The entropy of a random variable is the amount of information needed to fully describe it; alternate interpretations: average number of yes/no questions needed to identify X, how uncertain you are about X? X H(X) = − P (X) log P (X) = −EX [log P (X)] (8.1) X Average information, surprise, or uncertainty are all somewhat parsimonious plain English analogies for entropy. There are a few ways to measure entropy for multiple variables; we'll use two, X and Y . Definition 8.2 (Conditional entropy) The conditional entropy of a random variable is the entropy of one random variable conditioned on knowledge of another random variable, on average. Alternative interpretations: the average number of yes/no questions needed to identify X given knowledge of Y , on average; or How uncertain you are about X if you know Y , on average? X X h X i H(X j Y ) = P (Y )[H(P (X j Y ))] = P (Y ) − P (X j Y ) log P (X j Y ) Y Y X X = = − P (X; Y ) log P (X j Y ) X;Y = −EX;Y [log P (X j Y )] (8.2) Definition 8.3 (Joint entropy) X H(X; Y ) = − P (X; Y ) log P (X; Y ) = −EX;Y [log P (X; Y )] (8.3) X;Y 8-1 8-2 Lecture 8: Information Theory and Maximum Entropy • Bayes' rule for entropy H(X1 j X2) = H(X2 j X1) + H(X1) − H(X2) (8.4) • Chain rule of entropies n X H(Xn;Xn−1; :::X1) = H(Xn j Xn−1; :::X1) (8.5) i=1 It can be useful to think about these interrelated concepts with a so-called information diagram.
    [Show full text]
  • A Characterization of Guesswork on Swiftly Tilting Curves Ahmad Beirami, Robert Calderbank, Mark Christiansen, Ken Duffy, and Muriel Medard´
    View metadata, citation and similar papers at core.ac.uk brought to you by CORE provided by MURAL - Maynooth University Research Archive Library 1 A Characterization of Guesswork on Swiftly Tilting Curves Ahmad Beirami, Robert Calderbank, Mark Christiansen, Ken Duffy, and Muriel Medard´ Abstract Given a collection of strings, each with an associated probability of occurrence, the guesswork of each of them is their position in a list ordered from most likely to least likely, breaking ties arbitrarily. Guesswork is central to several applications in information theory: Average guesswork provides a lower bound on the expected computational cost of a sequential decoder to decode successfully the transmitted message; the complementary cumulative distribution function of guesswork gives the error probability in list decoding; the logarithm of guesswork is the number of bits needed in optimal lossless one-to-one source coding; and guesswork is the number of trials required of an adversary to breach a password protected system in a brute-force attack. In this paper, we consider memoryless string-sources that generate strings consisting of i.i.d. characters drawn from a finite alphabet, and characterize their corresponding guesswork. Our main tool is the tilt operation on a memoryless string-source. We show that the tilt operation on a memoryless string-source parametrizes an exponential family of memoryless string-sources, which we refer to as the tilted family of the string-source. We provide an operational meaning to the tilted families by proving that two memoryless string-sources result in the same guesswork on all strings of all lengths if and only if their respective categorical distributions belong to the same tilted family.
    [Show full text]
  • Language Models
    Recap: Language models Foundations of Natural Language Processing Language models tell us P (~w) = P (w . w ): How likely to occur is this • 1 n Lecture 4 sequence of words? Language Models: Evaluation and Smoothing Roughly: Is this sequence of words a “good” one in my language? LMs are used as a component in applications such as speech recognition, Alex Lascarides • machine translation, and predictive text completion. (Slides based on those from Alex Lascarides, Sharon Goldwater and Philipp Koehn) 24 January 2020 To reduce sparse data, N-gram LMs assume words depend only on a fixed- • length history, even though we know this isn’t true. Alex Lascarides FNLP lecture 4 24 January 2020 Alex Lascarides FNLP lecture 4 1 Evaluating a language model Two types of evaluation in NLP Intuitively, a trigram model captures more context than a bigram model, so Extrinsic: measure performance on a downstream application. • should be a “better” model. • – For LM, plug it into a machine translation/ASR/etc system. – The most reliable evaluation, but can be time-consuming. That is, it should more accurately predict the probabilities of sentences. • – And of course, we still need an evaluation measure for the downstream system! But how can we measure this? • Intrinsic: design a measure that is inherent to the current task. • – Can be much quicker/easier during development cycle. – But not always easy to figure out what the right measure is: ideally, one that correlates well with extrinsic measures. Let’s consider how to define an intrinsic measure for LMs.
    [Show full text]
  • Neural Networks and Backpropagation
    10-601 Introduction to Machine Learning Machine Learning Department School of Computer Science Carnegie Mellon University Neural Networks and Backpropagation Neural Net Readings: Murphy -- Matt Gormley Bishop 5 Lecture 20 HTF 11 Mitchell 4 April 3, 2017 1 Reminders • Homework 6: Unsupervised Learning – Release: Wed, Mar. 22 – Due: Mon, Apr. 03 at 11:59pm • Homework 5 (Part II): Peer Review – Expectation: You Release: Wed, Mar. 29 should spend at most 1 – Due: Wed, Apr. 05 at 11:59pm hour on your reviews • Peer Tutoring 2 Neural Networks Outline • Logistic Regression (Recap) – Data, Model, Learning, Prediction • Neural Networks – A Recipe for Machine Learning Last Lecture – Visual Notation for Neural Networks – Example: Logistic Regression Output Surface – 2-Layer Neural Network – 3-Layer Neural Network • Neural Net Architectures – Objective Functions – Activation Functions • Backpropagation – Basic Chain Rule (of calculus) This Lecture – Chain Rule for Arbitrary Computation Graph – Backpropagation Algorithm – Module-based Automatic Differentiation (Autodiff) 3 DECISION BOUNDARY EXAMPLES 4 Example #1: Diagonal Band 5 Example #2: One Pocket 6 Example #3: Four Gaussians 7 Example #4: Two Pockets 8 Example #1: Diagonal Band 9 Example #1: Diagonal Band 10 Example #1: Diagonal Band Error in slides: “layers” should read “number of hidden units” All the neural networks in this section used 1 hidden layer. 11 Example #1: Diagonal Band 12 Example #1: Diagonal Band 13 Example #1: Diagonal Band 14 Example #1: Diagonal Band 15 Example #2: One Pocket
    [Show full text]
  • Entropy and Decisions
    Entropy and decisions CSC401/2511 – Natural Language Computing – Winter 2021 Lecture 5, Serena Jeblee, Frank Rudzicz and Sean Robertson CSC401/2511 – Spring 2021 University of Toronto This lecture • Information theory and entropy. • Decisions. • Classification. • Significance and hypothesis testing. Can we quantify the statistical structure in a model of communication? Can we quantify the meaningful difference between statistical models? CSC401/2511 – Spring 2021 2 Information • Imagine Darth Vader is about to say either “yes” or “no” with equal probability. • You don’t know what he’ll say. • You have a certain amount of uncertainty – a lack of information. Darth Vader is © Disney And the prequels and Rey/Finn Star Wars suck CSC401/2511 – Spring 2021 3 Star Trek is better than Star Wars Information • Imagine you then observe Darth Vader saying “no” • Your uncertainty is gone; you’ve received information. • How much information do you receive about event ! when you observe it? 1 ! 1 = log ! )(1) For the units For the inverse of measurement 1 1 ! "# = log = log = 1 bit ! )("#) ! 1 ,2 CSC401/2511 – Spring 2021 4 Information • Imagine Darth Vader is about to roll a fair die. • You have more uncertainty about an event because there are more possibilities. • You receive more information when you observe it. 1 ! 5 = log ! )(5) " = log! ! ≈ 2.59 bits ⁄" CSC401/2511 – Spring 2021 5 Information is additive • From k independent, equally likely events ", 1 " 1 1 $ 8 = log = log $ % binary decisions = log! = % bits ! 9(8") ! 9 8 " 1 " 52 • For a unigram model, with each of 50K words # equally likely, 1 $ < = log ≈ 15.61 bits ! 1 550000 and for a sequence of 1K words in that model, " 1 $ < = log! ≈ 15,610???bits 1 #$$$ 550000 CSC401/2511 – Spring 2021 6 Information with unequal events • An information source S emits symbols without memory from a vocabulary #!, #", … , ## .
    [Show full text]
  • A Gentle Tutorial on Information Theory and Learning Roni Rosenfeld Carnegie Mellon University Outline • First Part Based Very
    Outline Definition of Information • First part based very loosely on [Abramson 63]. (After [Abramson 63]) • Information theory usually formulated in terms of information Let E be some event which occurs with probability channels and coding — will not discuss those here. P (E). If we are told that E has occurred, then we say that we have received 1. Information 1 I(E) = log2 2. Entropy P (E) bits of information. 3. Mutual Information 4. Cross Entropy and Learning • Base of log is unimportant — will only change the units We’ll stick with bits, and always assume base 2 • Can also think of information as amount of ”surprise” in E (e.g. P (E) = 1, P (E) = 0) • Example: result of a fair coin flip (log2 2= 1 bit) • Example: result of a fair die roll (log2 6 ≈ 2.585 bits) Carnegie Carnegie Mellon 2 ITtutorial,RoniRosenfeld,1999 Mellon 4 ITtutorial,RoniRosenfeld,1999 A Gentle Tutorial on Information Information Theory and Learning • information 6= knowledge Concerned with abstract possibilities, not their meaning Roni Rosenfeld • information: reduction in uncertainty Carnegie Carnegie Mellon University Mellon Imagine: #1 you’re about to observe the outcome of a coin flip #2 you’re about to observe the outcome of a die roll There is more uncertainty in #2 Next: 1. You observed outcome of #1 → uncertainty reduced to zero. 2. You observed outcome of #2 → uncertainty reduced to zero. =⇒ more information was provided by the outcome in #2 Carnegie Mellon 3 ITtutorial,RoniRosenfeld,1999 Entropy Entropy as a Function of a Probability Distribution A Zero-memory information source S is a source that emits sym- Since the source S is fully characterized byP = {p1,...pk} (we bols from an alphabet {s1, s2,...,sk} with probabilities {p1, p2,...,pk}, don’t care what the symbols si actually are, or what they stand respectively, where the symbols emitted are statistically indepen- for), entropy can also be thought of as a property of a probability dent.
    [Show full text]
  • Entropy, Relative Entropy, Cross Entropy Entropy
    Entropy, Relative Entropy, Cross Entropy Entropy Entropy, H(x) is a measure of the uncertainty of a discrete random variable. Properties: ● H(x) >= 0 ● Entropy Entropy ● Lesser the probability for an event, larger the entropy. Entropy of a six-headed fair dice is log26. Entropy : Properties Primer on Probability Fundamentals ● Random Variable ● Probability ● Expectation ● Linearity of Expectation Entropy : Properties Primer on Probability Fundamentals ● Jensen’s Inequality Ex:- Subject to the constraint that, f is a convex function. Entropy : Properties ● H(U) >= 0, Where, U = {u , u , …, u } 1 2 M ● H(U) <= log(M) Entropy between pair of R.Vs ● Joint Entropy ● Conditional Entropy Relative Entropy aka Kullback Leibler Distance D(p||q) is a measure of the inefficiency of assuming that the distribution is q, when the true distribution is p. ● H(p) : avg description length when true distribution. ● H(p) + D(p||q) : avg description length when approximated distribution. If X is a random variable and p(x), q(x) are probability mass functions, Relative Entropy/ K-L Divergence : Properties D(p||q) is a measure of the inefficiency of assuming that the distribution is q, when the true distribution is p. Properties: ● Non-negative. ● D(p||q) = 0 if p=q. ● Non-symmetric and does not satisfy triangular inequality - it is rather divergence than distance. Relative Entropy/ K-L Divergence : Properties Asymmetricity: Let, X = {0, 1} be a random variable. Consider two distributions p, q on X. Assume, p(0) = 1-r, p(1) = r ; q(0) = 1-s, q(1) = s; If, r=s, then D(p||q) = D(q||p) = 0, else for r!=s, D(p||q) != D(q||p) Relative Entropy/ K-L Divergence : Properties Non-negativity: Relative Entropy/ K-L Divergence : Properties Relative Entropy of joint distributions as Mutual Information Mutual Information, which is a measure of the amount of information that one random variable contains about another random variable.
    [Show full text]
  • Reduced Perplexity: a Simplified Perspective on Assessing Probabilistic Forecasts Kenric P
    Reduced perplexity: A simplified perspective on assessing probabilistic forecasts Kenric P. Nelson1,2 A simple, intuitive approach to the assessment of probabilistic inferences is introduced. The Shannon information metrics are translated to the probability domain. The translation shows that the negative logarithmic score and the geometric mean are equivalent measures of the accuracy of a probabilistic inference. Thus there is both a quantitative reduction in perplexity, which is the inverse of the geometric mean of the probabilities, as good inference algorithms reduce the uncertainty and a qualitative reduction due to the increased clarity between the original set of probabilistic forecasts and their central tendency, the geometric mean. Further insight is provided by showing that the Rényi and Tsallis entropy functions translated to the probability domain are both the weighted generalized mean of the distribution. The generalized mean of probabilistic forecasts forms a spectrum of performance metrics referred to as a Risk Profile. The arithmetic mean is used to measure the decisiveness, while the -2/3 mean is used to measure the robustness. 1 Introduction The objective of this chapter is to introduce a clear and simple approach to assessing the performance of probabilistic forecasts. This is important because machine learning and other techniques for decision making are often only evaluated in terms of percentage of correct decisions. Management of uncertainty in these systems requires that accurate probabilities be assigned to decisions. Unfortunately the existing assessment methods based on “scoring rules” [1], [2], which are defined later in the chapter, are poorly understood and often misapplied and/or misinterpreted. The approach here will be to ground the assessment of probability forecasts using information theory [3]–[5], while framing the results from the perspective of the central tendency and fluctuation of the forecasted probabilities.
    [Show full text]
  • K-Nearest Neighbor Based Consistent Entropy Estimation for Hyperspherical Distributions
    Entropy 2011, 13, 650-667; doi:10.3390/e13030650 OPEN ACCESS entropy ISSN 1099-4300 www.mdpi.com/journal/entropy Article k-Nearest Neighbor Based Consistent Entropy Estimation for Hyperspherical Distributions Shengqiao Li 1,, Robert M. Mnatsakanov 1,2, and Michael E. Andrew 1 1 Health Effects Laboratory Division, National Institute for Occupational Safety and Health, Morgantown, WV 26505, USA; E-Mail: [email protected] 2 Department of Statistics, West Virginia University, Morgantown, WV 26506, USA Authors to whom correspondence should be addressed; E-Mails: [email protected] (S.L.); [email protected] (R.M.). Received: 22 December 2010; in revised form: 27 January 2011 / Accepted: 28 February 2011 / Published: 8 March 2011 Abstract: A consistent entropy estimator for hyperspherical data is proposed based on the k-nearest neighbor (knn) approach. The asymptotic unbiasedness and consistency of the estimator are proved. Moreover, cross entropy and Kullback-Leibler (KL) divergence estimators are also discussed. Simulation studies are conducted to assess the performance of the estimators for models including uniform and von Mises-Fisher distributions. The proposed knn entropy estimator is compared with the moment based counterpart via simulations. The results show that these two methods are comparable. Keywords: hyperspherical distribution; directional data; differential entropy; cross entropy; Kullback-Leibler divergence; k-nearest neighbor 1. Introduction The Shannon (or differential) entropy of a continuously distributed random variable (r.v.) X with probability density function (pdf ) f is widely used in probability theory and information theory as a measure of uncertainty. It is defined as the negative mean of the logarithm of the density function, i.e., H(f)=−Ef [ln f(X)] (1) Entropy 2011, 13 651 k-Nearest neighbor (knn) density estimators were proposed by Mack and Rosenblatt [1].
    [Show full text]
  • Entropy Methods for Joint Distributions in Decision Analysis Ali E
    146 IEEE TRANSACTIONS ON ENGINEERING MANAGEMENT, VOL. 53, NO. 1, FEBRUARY 2006 Entropy Methods for Joint Distributions in Decision Analysis Ali E. Abbas, Member, IEEE Abstract—A fundamental step in decision analysis is the elicita- When used to construct joint probability distributions in de- tion of the decision maker’s information about the uncertainties of cision analysis, the maximum entropy approach has several ad- the decision situation in the form of a joint probability distribution. vantages: 1) it incorporates as much (or as little) information as This paper presents a method based on the maximum entropy prin- ciple to obtain a joint probability distribution using lower order there is available at the time of making the decision; 2) it makes joint probability assessments. The approach reduces the number no assumptions about a particular form of a joint distribution; of assessments significantly and also reduces the number of con- 3) it applies to both numeric and nonnumeric variables (such as ditioning variables in these assessments. We discuss the order of the categorical events Sunny/Rainy for weather); and 4) it does the approximation provided by the maximum entropy distribution not limit itself to the use of only moments and correlation co- with each lower order assessment using a Monte Carlo simulation and discuss the implications of using the maximum entropy distri- efficients, which may be difficult to elicit in decision analysis bution in Bayesian inference. We present an application to a prac- practice. tical decision situation faced by a semiconductor testing company We do observe, however, that there is a tradeoff between the in the Silicon Valley.
    [Show full text]