DOCSLIB.ORG

4 Mutual Information and Channel Capacity

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
4 Mutual Information and Channel Capacity 4 Mutual Information and Channel Capacity In Section 2, we have seen the use of a quantity called entropy to measure the amount of randomness in a random variable. In this section, we in- troduce several more information-theoretic quantities. These quantities are important in the study of Shannon's results. 4.1 Information-Theoretic Quantities Definition 4.1. Recall that, the entropy of a discrete random variable X is defined in Definition 2.41 to be X H (X) = − pX (x) log2 pX (x) = −E [log2 pX (X)] : (16) x2SX Similarly, the entropy of a discrete random variable Y is given by X H (Y ) = − pY (y) log2 pY (y) = −E [log2 pY (Y )] : (17) y2SY In our context, the X and Y are input and output of a discrete memory- less channel, respectively. In such situation, we have introduced some new notations in Section 3.1: SX ≡ X pX(x) ≡ p(x) p pY jX(yjx) ≡ Q(yjx) Q matrix SY ≡ Y pY (y) ≡ q(y) q pX;Y (x; y) ≡ p(x; y) P matrix Under such notations, (16) and (17) become X H (X) = − p (x) log2 p (x) = −E [log2 p (X)] (18) x2X and X H (Y ) = − q (y) log2 q (y) = −E [log2 q (Y )] : (19) y2Y Definition 4.2. The joint entropy for two random variables X and Y is given by X X H (X; Y ) = − p (x; y) log2p (x; y) = −E [log2 p (X; Y )] : x2X y2Y 48 Example 4.3. Random variables X and Y have the following joint pmf matrix P: 2 1 1 1 1 3 8 16 16 4 1 1 1 6 16 8 16 0 7 6 1 1 1 7 4 32 32 16 0 5 1 1 1 32 32 16 0 Find H(X), H(Y ) and H(X; Y ). Definition 4.4. The (conditional) entropy of Y when we know X = x is denoted by H (Y jX = x) or simply H(Y jx). It can be calculated from X H (Y jx) = − Q (y jx) log2 Q (y jx) y2Y • Note that the above formula is what we should expect it to be. When we want to find the entropy of Y , we use (19): X H (Y ) = − q (y) log2 q (y): y2Y When we have an extra piece of information that X = x, we should update the probability about Y from the unconditional probability q(y) to the conditional probability Q(yjx). • Note that when we consider Q(yjx) with the value of x fixed and the value of y varied, we simply get the whole x-row from Q matrix. So, to 49 find H (Y jx), we simply find the \usual" entropy from the probability values in the row corresponding to x in the Q matrix. Example 4.5. Consider the following DMC (actually BAC) 1/2 / 0 0 1/2 X Y / 1 1 1 8 1=4; y = 0; < Originally P [Y = y] = q(y) = 3=4; y = 1; : 0; otherwise. (a) Suppose we know that X = 0. 1=2; y = 0; 1; The \x = 0" row in the Q matrix gives Q (yj0) = 0; otherwise; that is, given x = 0, the RV Y will be uniform. (b) Suppose we know that X = 1. The \x = 1" row in the Q matrix 1; y = 1; gives Q (yj1) = that is, given x = 1, the RV Y is 0; otherwise; degenerated (deterministic). Definition 4.6. Conditional entropy: The (average) conditional entropy of Y when we know X is denoted by H(Y jX). It can be calculated from X H (Y jX ) = p (x) H (Y jx): x2X Example 4.7. In Example 4.5, 50 4.8. An alternative way to calculate H(Y jX) can be derived by first rewrit- ing it as X X X H (Y jX ) = p (x) H (Y jx) = − p (x) Q (y jx) log2 Q (y jx) x2X x2X y2Y X X = − p (x; y) log2Q (y jx) = −E [log2 Q (Y jX )] x2X y2Y p(x;y) Note that Q(yjx) = p(x) . Therefore, p (X; Y ) H (Y jX) = − [log Q (Y jX )] = − log E 2 E 2 p (X) = (−E [log2p (X; Y )]) − (−E [log2p (X)]) = H (X; Y ) − H (X) Example 4.9. In Example 4.5, Example 4.10. Continue from Example 4.3. Random variables X and Y have the following joint pmf matrix P: 2 1 1 1 1 3 8 16 16 4 1 1 1 6 16 8 16 0 7 6 1 1 1 7 4 32 32 16 0 5 1 1 1 32 32 16 0 Find H(Y jX) and H(XjY ). 51 ECS 452: In-Class Exercise # _ Instructions Date: _ _ / _ _ / 2017 1. Separate into groups of no more than three persons. Name ID (last 3 digits) 2. The group cannot be the same as your former group. 3. Only one submission is needed for each group. 4. Write down all the steps that you have done to obtain your answers. You may not get full credit even when your answer is correct without showing how you get your answer. 5. Do not panic. 1. Consider two random variables X and Y whose joint pmf matrix is given by y 3 5 x 1 1 5 18 18 P . 42 2 99 Calculate the following quantities. Except for part (e), all of your answers should be of the form X.XXXX. a. 퐻(푋, 푌) b. 퐻(푋) c. 퐻(푌) d. 퐻(푌|푋) e. Q matrix Definition 4.11. The mutual information13 I(X; Y ) between two ran- dom variables X and Y is defined as I (X; Y ) = H (X) − H (X jY ) (20) = H (Y ) − H (Y jX ) (21) = H (X) + H (Y ) − H (X; Y ) (22) p (X; Y ) X X p (x; y) = log = p (x; y) log (23) E 2 p (X) q (Y ) p (x) q (y) x2X y2Y P (X jY ) Q (Y jX ) = log XjY = log : (24) E 2 p (X) E 2 q (Y ) • Mutual information quantifies the reduction in the uncertainty of one random variable due to the knowledge of another. • Mutual information is a measure of the amount of information one random variable contains about another [5, p 13]. • It is also natural to think of I(X; Y ) as a measure of how far X and Y are from being independent. ◦ Technically, it is the (Kullback-Leibler) divergence between the joint and product-of-marginal distributions. 4.12. Some important properties (a) H(X; Y ) = H(Y; X) and I(X; Y ) = I(Y ; X). However, in general, H(XjY ) 6= H(Y jX). (b) I and H are always ≥ 0. (c) There is a one-to-one correspondence between Shannon's information measures and set theory. We may use an information diagram, which 13The name mutual information and the notation I(X; Y ) was introduced by [Fano, 1961, Ch 2]. 52 is a variation of a Venn diagram, to represent relationship between Shannon's information measures. This is similar to the use of the Venn diagram to represent relationship between probability measures. These diagrams are shown in Figure 7. Venn Diagram Information Diagram Probability Diagram 퐵 푌 퐵 퐻 푋 푃 퐴 퐴\B 퐴 ∩ 퐵 퐵\A 퐻 푋|푌 퐼 푋; 푌 퐻 푌|푋 푃 퐴\B 푃 퐴 ∩ 퐵 푃 퐵\A 퐴 푋 퐻 푌 퐴 푃 퐵 퐴 ∪ 퐵 퐻 푋, 푌 푃 퐴 ∪ 퐵 Figure 7: Venn diagram and its use to represent relationship between information measures and relationship between probabilities • Chain rule for information measures: H (X; Y ) = H (X) + H (Y jX ) = H (Y ) + H (X jY ) : (d) I(X; Y ) ≥ 0 with equality if and only if X and Y are independent. • When this property is applied to the information diagram (or def- initions (20), (21), and (22) for I(X; Y )), we have (i) H(XjY ) ≤ H(X), (ii) H(Y jX) ≤ H(Y ), (iii) H(X; Y ) ≤ H(X) + H(Y ) Moreover, each of the inequalities above becomes equality if and = only X j Y . (e) We have seen in Section 2.4 that 0 ≤ H (X) ≤ log2 jX j : (25) deterministic (degenerated) uniform Similarly, 0 ≤ H (Y ) ≤ log2 jYj : (26) deterministic (degenerated) uniform 53 For conditional entropy, we have 0 ≤ H (Y jX ) ≤ H (Y ) (27) 9g Y =g(X) X = Y j and 0 ≤ H (X jY ) ≤ H (X) : (28) 9g X=g(Y ) X = Y j For mutual information, we have 0 ≤ I (X; Y ) ≤ H (X) (29) X = Y 9g X=g(Y ) j and 0 ≤ I (X; Y ) ≤ H (Y ) : (30) X = Y 9g Y =g(X) j Combining 25, 26, 29, and 30, we have 0 ≤ I (X; Y ) ≤ min fH (X) ;H (Y )g ≤ min flog2 jX j ; log2 jYjg (31) (f) H (X jX ) = 0 and I(X; X) = H(X). Example 4.13. Find the mutual information I(X; Y ) between the two ran- h 1 1 i 2 4 dom variables X and Y whose joint pmf matrix is given by P = 1 : 4 0 Example 4.14. Find the mutual information I(X; Y ) between the two ran- h 1 3 i 1 3 4 4 dom variables X and Y whose p = 4; 4 and Q = 3 1 : 4 4 54.
Load more

Recommended publications
  • Information Theory Techniques for Multimedia Data Classification and Retrieval
    INFORMATION THEORY TECHNIQUES FOR MULTIMEDIA DATA CLASSIFICATION AND RETRIEVAL Marius Vila Duran Dipòsit legal: Gi. 1379-2015 http://hdl.handle.net/10803/302664 http://creativecommons.org/licenses/by-nc-sa/4.0/deed.ca Aquesta obra està subjecta a una llicència Creative Commons Reconeixement- NoComercial-CompartirIgual Esta obra está bajo una licencia Creative Commons Reconocimiento-NoComercial- CompartirIgual This work is licensed under a Creative Commons Attribution-NonCommercial- ShareAlike licence DOCTORAL THESIS Information theory techniques for multimedia data classification and retrieval Marius VILA DURAN 2015 DOCTORAL THESIS Information theory techniques for multimedia data classification and retrieval Author: Marius VILA DURAN 2015 Doctoral Programme in Technology Advisors: Dr. Miquel FEIXAS FEIXAS Dr. Mateu SBERT CASASAYAS This manuscript has been presented to opt for the doctoral degree from the University of Girona List of publications Publications that support the contents of this thesis: "Tsallis Mutual Information for Document Classification", Marius Vila, Anton • Bardera, Miquel Feixas, Mateu Sbert. Entropy, vol. 13, no. 9, pages 1694-1707, 2011. "Tsallis entropy-based information measure for shot boundary detection and • keyframe selection", Marius Vila, Anton Bardera, Qing Xu, Miquel Feixas, Mateu Sbert. Signal, Image and Video Processing, vol. 7, no. 3, pages 507-520, 2013. "Analysis of image informativeness measures", Marius Vila, Anton Bardera, • Miquel Feixas, Philippe Bekaert, Mateu Sbert. IEEE International Conference on Image Processing pages 1086-1090, October 2014. "Image-based Similarity Measures for Invoice Classification", Marius Vila, Anton • Bardera, Miquel Feixas, Mateu Sbert. Submitted. List of figures 2.1 Plot of binary entropy..............................7 2.2 Venn diagram of Shannon’s information measures............ 10 3.1 Computation of the normalized compression distance using an image compressor....................................
    [Show full text]
  • 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 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]
  • Mutual Information Analysis: a Comprehensive Study∗
    J. Cryptol. (2011) 24: 269–291 DOI: 10.1007/s00145-010-9084-8 Mutual Information Analysis: a Comprehensive Study∗ Lejla Batina ESAT/SCD-COSIC and IBBT, K.U.Leuven, Kasteelpark Arenberg 10, 3001 Leuven-Heverlee, Belgium [email protected] and CS Dept./Digital Security group, Radboud University Nijmegen, Heyendaalseweg 135, 6525 AJ, Nijmegen, The Netherlands Benedikt Gierlichs ESAT/SCD-COSIC and IBBT, K.U.Leuven, Kasteelpark Arenberg 10, 3001 Leuven-Heverlee, Belgium [email protected] Emmanuel Prouff Oberthur Technologies, 71-73 rue des Hautes Pâtures, 92726 Nanterre Cedex, France [email protected] Matthieu Rivain CryptoExperts, Paris, France [email protected] François-Xavier Standaert† and Nicolas Veyrat-Charvillon UCL Crypto Group, Université catholique de Louvain, 1348 Louvain-la-Neuve, Belgium [email protected]; [email protected] Received 1 September 2009 Online publication 21 October 2010 Abstract. Mutual Information Analysis is a generic side-channel distinguisher that has been introduced at CHES 2008. It aims to allow successful attacks requiring min- imum assumptions and knowledge of the target device by the adversary. In this paper, we compile recent contributions and applications of MIA in a comprehensive study. From a theoretical point of view, we carefully discuss its statistical properties and re- lationship with probability density estimation tools. From a practical point of view, we apply MIA in two of the most investigated contexts for side-channel attacks. Namely, we consider first-order attacks against an unprotected implementation of the DES in a full custom IC and second-order attacks against a masked implementation of the DES in an 8-bit microcontroller.
    [Show full text]
  • Nonparametric Maximum Entropy Estimation on Information Diagrams
    Nonparametric Maximum Entropy Estimation on Information Diagrams Elliot A. Martin,1 Jaroslav Hlinka,2, 3 Alexander Meinke,1 Filip Dˇechtˇerenko,4, 2 and J¨ornDavidsen1 1Complexity Science Group, Department of Physics and Astronomy, University of Calgary, Calgary, Alberta, Canada, T2N 1N4 2Institute of Computer Science, The Czech Academy of Sciences, Pod vodarenskou vezi 2, 18207 Prague, Czech Republic 3National Institute of Mental Health, Topolov´a748, 250 67 Klecany, Czech Republic 4Institute of Psychology, The Czech Academy of Sciences, Prague, Czech Republic (Dated: September 26, 2018) Maximum entropy estimation is of broad interest for inferring properties of systems across many different disciplines. In this work, we significantly extend a technique we previously introduced for estimating the maximum entropy of a set of random discrete variables when conditioning on bivariate mutual informations and univariate entropies. Specifically, we show how to apply the concept to continuous random variables and vastly expand the types of information-theoretic quantities one can condition on. This allows us to establish a number of significant advantages of our approach over existing ones. Not only does our method perform favorably in the undersampled regime, where existing methods fail, but it also can be dramatically less computationally expensive as the cardinality of the variables increases. In addition, we propose a nonparametric formulation of connected informations and give an illustrative example showing how this agrees with the existing parametric formulation in cases of interest. We further demonstrate the applicability and advantages of our method to real world systems for the case of resting-state human brain networks. Finally, we show how our method can be used to estimate the structural network connectivity between interacting units from observed activity and establish the advantages over other approaches for the case of phase oscillator networks as a generic example.
    [Show full text]
  • Mutual Information Analysis: a Comprehensive Study *
    Mutual Information Analysis: a Comprehensive Study ? Lejla Batina1;2, Benedikt Gierlichs1, Emmanuel Prouff3, Matthieu Rivain4, Fran¸cois-Xavier Standaert5?? and Nicolas Veyrat-Charvillon5 1 K.U.Leuven, ESAT/SCD-COSIC and IBBT, Kasteelpark Arenberg 10, B-3001 Leuven-Heverlee, Belgium. flejla.batina, [email protected] 2 Radboud University Nijmegen, CS Dept./Digital Security group, Heyendaalseweg 135, 6525 AJ, Nijmegen, The Netherlands. 3 Oberthur Technologies, 71-73 rue des Hautes P^atures,92726 Nanterre Cedex, France. [email protected] 4 CryptoExperts, Paris, France. [email protected] 5 Universit´ecatholique de Louvain, UCL Crypto Group, B-1348 Louvain-la-Neuve, Belgium. ffstandae, [email protected] Abstract. Mutual Information Analysis is a generic side-channel dis- tinguisher that has been introduced at CHES 2008. It aims to allow successful attacks requiring minimum assumptions and knowledge of the target device by the adversary. In this paper, we compile recent contri- butions and applications of MIA in a comprehensive study. From a the- oretical point of view, we carefully discuss its statistical properties and relationship with probability density estimation tools. From a practical point of view, we apply MIA in two of the most investigated contexts for side-channel attacks. Namely, we consider first order attacks against an unprotected implementation of the DES in a full custom IC and second order attacks against a masked implementation of the DES in an 8-bit microcontroller. These experiments allow to put forward the strengths and weaknesses of this new distinguisher and to compare it with standard power analysis attacks using the correlation coefficient.
    [Show full text]
  • Anatomy of a Bit: Information in a Time Series Observation
    Santa Fe Institute Working Paper 11-05-XXX arxiv.org:1105.XXXX [physics.gen-ph] Anatomy of a Bit: Information in a Time Series Observation 1, 2, 1, 2, 1, 2, 3, Ryan G. James, ∗ Christopher J. Ellison, y and James P. Crutchfield z 1Complexity Sciences Center 2Physics Department University of California at Davis, One Shields Avenue, Davis, CA 95616 3Santa Fe Institute 1399 Hyde Park Road, Santa Fe, NM 87501 (Dated: October 24, 2018) Appealing to several multivariate information measures|some familiar, some new here|we ana- lyze the information embedded in discrete-valued stochastic time series. We dissect the uncertainty of a single observation to demonstrate how the measures' asymptotic behavior sheds structural and semantic light on the generating process's internal information dynamics. The measures scale with the length of time window, which captures both intensive (rates of growth) and subextensive com- ponents. We provide interpretations for the components, developing explicit relationships between them. We also identify the informational component shared between the past and the future that is not contained in a single observation. The existence of this component directly motivates the notion of a process's effective (internal) states and indicates why one must build models. Keywords: entropy, total correlation, multivariate mutual information, binding information, en- tropy rate, predictive information rate PACS numbers: 02.50.-r 89.70.+c 05.45.Tp 02.50.Ey 02.50.Ga compressed. In fact, a single observation tells us the os- A single measurement, when considered in the cillation's phase. And, with this single bit of information, context of the past and the future, contains a we have learned everything|the full bit that the time se- wealth of information, including distinct kinds of ries contains.
    [Show full text]
  • Information Measures & Information Diagrams
    Information Measures & Information Diagrams Reading for this lecture: CMR articles Yeung & Anatomy Lecture 19: Natural Computation & Self-Organization, Physics 256B (Winter 2014); Jim Crutchfield Monday, March 10, 14 1 Information Measures ... One random variable: X Pr(x) ⇠ Use paradigm: One shot sampling. Recall information theory quantity: Entropy: H[X] Lecture 19: Natural Computation & Self-Organization, Physics 256B (Winter 2014); Jim Crutchfield Monday, March 10, 14 2 Information Measures ... Two random variables: X Pr(x) ⇠ Y Pr(y) ⇠ (X, Y ) Pr(x, y) ⇠ Recall information theory quantities: Joint entropy: H[X, Y ] Conditional Entropies: H[X Y ] H[Y X] Mutual Information: I[X|; Y ] | Information Metric: d(X, Y ) Lecture 19: Natural Computation & Self-Organization, Physics 256B (Winter 2014); Jim Crutchfield Monday, March 10, 14 3 Information Measures ... Event Space Relationships of Information Quantifiers: Lecture 19: Natural Computation & Self-Organization, Physics 256B (Winter 2014); Jim Crutchfield Monday, March 10, 14 4 Information Measures ... Event Space Relationships of Information Quantifiers: Lecture 19: Natural Computation & Self-Organization, Physics 256B (Winter 2014); Jim Crutchfield Monday, March 10, 14 4 Information Measures ... Event Space Relationships of Information Quantifiers: H(X) Lecture 19: Natural Computation & Self-Organization, Physics 256B (Winter 2014); Jim Crutchfield Monday, March 10, 14 4 Information Measures ... Event Space Relationships of Information Quantifiers: H(X) Lecture 19: Natural Computation & Self-Organization, Physics 256B (Winter 2014); Jim Crutchfield Monday, March 10, 14 4 Information Measures ... Event Space Relationships of Information Quantifiers: H(Y ) H(X) Lecture 19: Natural Computation & Self-Organization, Physics 256B (Winter 2014); Jim Crutchfield Monday, March 10, 14 4 Information Measures ..
    [Show full text]
  • 5 Mutual Information and Channel Capacity
    5 Mutual Information and Channel Capacity In Section 2, we have seen the use of a quantity called entropy to measure the amount of randomness in a random variable. In this section, we in- troduce several more information-theoretic quantities. These quantities are important in the study of Shannon's results. 5.1 Information-Theoretic Quantities Definition 5.1. Recall that, the entropy of a discrete random variable X is defined in Definition 2.41 to be X H (X) = − pX (x) log2 pX (x) = −E [log2 pX (X)] : (16) x2SX Similarly, the entropy of a discrete random variable Y is given by X H (Y ) = − pY (y) log2 pY (y) = −E [log2 pY (Y )] : (17) y2SY In our context, the X and Y are input and output of a discrete memory- less channel, respectively. In such situation, we have introduced some new notations in Section 3.1: Under such notations, (16) and (17) become X H (X) = − p (x) log2 p (x) = −E [log2 p (X)] (18) x2X and X H (Y ) = − q (y) log2 q (y) = −E [log2 q (Y )] : (19) y2Y Definition 5.2. The joint entropy for two random variables X and Y is given by X X H (X; Y ) = − p (x; y) log2p (x; y) = −E [log2 p (X; Y )] : x2X y2Y 45 Example 5.3. Random variables X and Y have the following joint pmf matrix P: 2 1 1 1 1 3 8 16 16 4 1 1 1 6 16 8 16 0 7 6 1 1 1 7 4 32 32 16 0 5 1 1 1 32 32 16 0 Find H(X), H(Y ) and H(X; Y ).
    [Show full text]
  • Two Information-Theoretic Tools to Assess the Performance of Multi-Class Classifiers
    View metadata, citation and similar papers at core.ac.uk brought to you by CORE provided by Universidad Carlos III de Madrid e-Archivo Two Information-Theoretic Tools to Assess the Performance of Multi-class Classifiers Francisco J. Valverde-Albacete∗, Carmen Pelaez-Moreno´ Departamento de Teor´ıade la Se˜naly de las Comunicaciones, Universidad Carlos III de Madrid Avda de la Universidad, 30. 28911 Legan´es,Spain Abstract We develop two tools to analyze the behavior of multiple-class, or multi-class, classifiers by means of entropic mea- sures on their confusion matrix or contingency table. First we obtain a balance equation on the entropies that captures interesting properties of the classifier. Second, by normalizing this balance equation we first obtain a 2-simplex in a three-dimensional entropy space and then the de Finetti entropy diagram or entropy triangle. We also give examples of the assessment of classifiers with these tools. Key words: Multiclass classifier, confusion matrix, contingency table, performance measure, evaluation criterion, de Finetti diagram, entropy triangle 1 1. Introduction 22 non-uniform prior distributions of input patterns (Ben- 23 David, 2007; Sindhwani et al., 2004). For instance, with n p 2 Let VX = {xi}i=1 and VY = {y j} j=1 be sets of input and 24 continuous speech corpora, the silence class may ac- 3 output class identifiers, respectively, in a multiple-class 25 count for 40–60% percent of input patterns making a 4 classification task. The basic classification event con- 26 majority classifier that always decides Y = silence, the 5 sists in “presenting a pattern of input class xi to the clas- 27 most prevalent class, quite accurate but useless.
    [Show full text]
  • The Conditional Entropy Bottleneck
    entropy Article The Conditional Entropy Bottleneck Ian Fischer Google Research, Mountain View, CA 94043, USA; [email protected] Received: 30 July 2020; Accepted: 28 August 2020; Published: 8 September 2020 Abstract: Much of the field of Machine Learning exhibits a prominent set of failure modes, including vulnerability to adversarial examples, poor out-of-distribution (OoD) detection, miscalibration, and willingness to memorize random labelings of datasets. We characterize these as failures of robust generalization, which extends the traditional measure of generalization as accuracy or related metrics on a held-out set. We hypothesize that these failures to robustly generalize are due to the learning systems retaining too much information about the training data. To test this hypothesis, we propose the Minimum Necessary Information (MNI) criterion for evaluating the quality of a model. In order to train models that perform well with respect to the MNI criterion, we present a new objective function, the Conditional Entropy Bottleneck (CEB), which is closely related to the Information Bottleneck (IB). We experimentally test our hypothesis by comparing the performance of CEB models with deterministic models and Variational Information Bottleneck (VIB) models on a variety of different datasets and robustness challenges. We find strong empirical evidence supporting our hypothesis that MNI models improve on these problems of robust generalization. Keywords: information theory; information bottleneck; machine learning 1. Introduction Despite excellent progress in classical generalization (e.g., accuracy on a held-out set), the field of Machine Learning continues to struggle with the following issues: • Vulnerability to adversarial examples. Most machine-learned systems are vulnerable to adversarial examples.
    [Show full text]
  • The Mutual Information Diagram for Uncertainty Visualization
    International Journal for Uncertainty Quantification, 1(1):xxx–xxx, 2012 THE MUTUAL INFORMATION DIAGRAM FOR UNCERTAINTY VISUALIZATION Carlos D. Correa∗ & Peter Lindstrom Center for Applied Scientific Computing (CASC), Lawrence Livermore National Laboratory, Liv- ermore, CA,USA Original Manuscript Submitted: 09/05/2012; Final Draft Received: 09/05/2012 We present a variant of the Taylor diagram, a type of 2D plot that succinctly shows the relationship between two or more random variables based on their variance and correlation. The Taylor diagram has been adopted by the climate and geophysics communities to produce insightful visualizations, e.g., for intercomparison studies. Our variant, which we call the Mutual Information Diagram, represents the relationship between random variables in terms of their entropy and mutual information, and naturally maps well-known statistical quantities to their information-theoretic counter- parts. Our new diagram is able to describe non-linear relationships where linear correlation may fail; it allows for categorical and multi-variate data to be compared; and it incorporates the notion of uncertainty, key in the study of large ensembles of data. KEY WORDS: mutual information, entropy, variation of information, uncertainty visualization, Taylor diagrams 1. INTRODUCTION Making sense of statistical data from observation and simulation is challenging, especially when data is surrounded by uncertainty. Dealing with uncertainty is a multi-faceted process that includes the quantification, modeling, propagation and visualization of the uncertainties and errors intrinsic in the data and arising from data transformations. In this paper, we focus on the visualization of uncertainty. One common task in uncertainty studies is the comparison of different models to determine whether they are effective in explaining the observed data.
    [Show full text]