Note on page numbering

- the page numbering here uses PDF updated May 2011 which has Introduction starting page 1.

- between brackets, the printed edition numbering.

- note: PDF from Christoph’s website starts at p186

Bold for important errata (as opposed to typos which won’t confuse the reader)

pXXX (15) (eq 2.28) - should read argmax -E() at the end

pXXX (19) (prob 2.4) - data in -> data and

pXXX (33) (algo 2, line 14 and 19) - for (i,F) in callig E

p34 (42) (3.3.1) - parag “The mean field approach”: "it has been used /in/ parameter estimation"

p40 (algo 5, line 5) - y^t -> y^{(t)}

p61 (middle of page) - assume that g>=0

p71 (78) (4.5.2.2, fig4.13) - n_{i,jk} -> n_{j,jk} twice

p78 (86) (4.6.1, line 3) - becomes -> become

p79 (legend of fig 4.24) - should explicitly state notation for ESS rectangle set is B

p81 (91) (4.6.1, eq 4.31) - summation should be over {p,q} \in E, not {i,j}

p81 (91) (4.6.1, eq 4.32, above eq 4.33, in eq 4.33) - g(x,y) -> g(y)

p82 (92) (4.6.1, eq 4.36) - summation should be over {p,q} not {i,j}

p82 (92) (4.6.1, eq 4.37) - g(x,y) -> g(y)

p87 (97) (4.7.1, text above eq 4.44) - change doubled Y to calligraph Y in definition of \theta_{1,2}

p90 sqq (4.7.2, eq 4.62) - running erratum (in all of section 4.7.2): regarding the last term \mu^T v(y) : we should have either 1) minus this term in eq 4.62 and thm 4.3, or 2) \mu negative, not positive as stated, or 3) v positive, or 4) turn the max problem into a min problem (but conventionally the primal is a minimization problem)

p91 (thm 4.3, second parag) - v(u) -> v(y)

p92 (above eq 4.67) - second step size condition should read: \sum_{t’=0}^{t’=t} \alpha^{t’} \tendsto{t \tendsto +\infty} +\infty

pXXX (104) (thm 4.4) - necessary condition instead of sufficient (though I think sufficient is fine, SB)

p95 (105) (4.7.2, text after eq 4.75) - necessary -> advisable in "It is, however, necessary to make it explicit."

p99 (eq 4.94) - missing definition of notation conv (convex hull)

p100 (fig 4.34) - missing attribution of points to either Y_1 or Y_2 (could be marked in colour: red for Y_1, red for Y_2). Furthermore: point noted “y” in printed edition should be noted y^*. Finally, arrow at the bottom right symbolizes vector c, not scalar c^T y.

p101 (112) (4.7.3, eq 4.95) - first term should not be const but depend on the fg/bg model and y_i (reported by Tomas, but I (SB) can’t find where it says the first term is constant)

p104 (115) (4.7.4, line -1) - definition of J_{LOCAL} should start (A,B,y_B) not (A,B,x_B), and should have B strict subset of A, not subset

p109 (expl 4.13, above eq 4.108) - “five possible intensity values” : should specify explicitly that y_i belongs to {1...5} . Further: parag starting “To this end” : p(y|y_noisy) -> p(y|y’)

pXXX (117) (4.8.1) generating samples -> generate samples

pXXX (122) (4.8) - in practise. -> in practice.

p114 (def 5.1, eq 5.2) - first sum should run from n=1

p117 (above eq 5.14) - Hessian matrix is noted H, not \Delta

p117 (eq 5.14 and 5.15) - p(y|x^n) -> p(y|x^n,w) (should make dependence on w explicit)

p119 (algo 10, line 12) - p -> d

p119 (text below algo 10) - minimize -> minimized, requires -> require, compared the -> compared to, method -> methods

pXXX (132) (5.3) to be minimize -> to be minimized

pXXX (133) (5.3) better model that -> better model than

pXXX (133) (5.3) typically requires -> typically require

pXXX (133) (5.3) converge compared the -> converge compared to the

pXXX (134) (5.3) sentence starting "This estimation is possible..." is weird

p120 - d x d matrix -> D x D matrix (top and bottom of page)

p121 (135) (5.4 last line of first parag) - it given -> it is given, provides can be -> remove “provides”

p120 (algo 11 line 11) - remove ^-1

pXXX (135) (5.4) as a form a logistic -> as a form of logistic

pXXX (135) (5.4) as it given -> as given

pXXX (135) (5.4) model that are -> models that are

p122 (136) (5.5) - extremal -> extreme, yield -> yields, require -> requires, has let -> has lead, such that -> so that

pXXX (138) (5.6) which justified -> which justifies

p124 (138) - very noise -> very noisy

pXXX (140) (5.8) makes us of -> makes use of

pXXX (142) (5.8) conditional random with -> conditional random field with

pXXX (143) (5.9) it is common stop -> it is common to stop

pXXX (143) (5.9) only approximately minimizes -> only approximately maximizes

pXXX (144) (Eq. 5.23) superfluous trailing 0

pXXX (145) (5.9) independent outputs nodes -> independent output nodes

pXXX (145) (Eq. 5.27) inconsistent usage of w_F and w in the same sense

                    Z_F(x^n_f, y^n_F, w_F) -> Z_F(x^n_F, w_F)

pXXX (146) (Eq. 5.28) Z_F(x) -> Z_F(x_F, w_F)

pXXX (146) (Eq. 5.30) p(y_F|x) -> p(y_F|x_F,w_F)

                    Z_F(x, w_F) -> Z_F(x_F, w_F)

p134 (first parag) - from we -> from which we

pXXX (165) (6.4) terms have have -> terms have

pXXX (165) (6.4) can be thought of a -> can be thought of as

pXXX (166) (6.5) ordinary, loss-free, prediction -> ordinary, loss-free prediction

p150 (parag below Def 6.5) - such that -> so that

p150 (line -2) - has shown -> has proven

p151 (algo 16) - ConvexConcave -> ConcaveConvex

p151 - each images -> each image, idea -> ideas

pXXX (168) (6.5) all weight vector -> all weight vectors

pXXX (168) (6.5) the predicted segmenetation -> the predicted segmentation

pXXX (170) (6.6) loss-augment prediction -> loss-augmented prediction

p155 (170) (6.6) - real values -> real-valued

pXXX (171) (7) in terms of what -> in terms of which

pXXX (174) - weight vector -> Weight vector

This errata list was compiled by Sébastien Bratières from errata collected with the IST Austria reading group 2011/2012, organized by Prof Christoph Lampert.