List of Errata

Time-stamp: "2011-01-07 10:43:31 jnw"

Haykin, S., Neural Networks and Machine Learning, Third Edition, Prentice Hall, 2008, ISBN 0-13-147139-2

Page, positionError
27, Figure 19 Since xi and xj are not normal vectors, replace xiTxj by xiTxj / ||xj|| in the diagram.
48, line 4 Delete the first sentence. The impact of McColloch-Pitts is not discussed in this edition's introductory chapter.
51, line -12 Replace wTx(n) by wT(n) x(n)
51, line -11 Replace wTx(n) by wT(n) x(n)
51, line -7 Replace w(n) - η(n)x(n) by w(n) + η(n)x(n) in this second line of equation (1.6).
Whole proof
52, line 4 to
53, line 2
The proof deals only with training samples from H1, and is thus incorrect. To fix the proof, replace occurrences of x(n) by x(n)d(n), where d is as defined on page 54, line -2. All terms in the sequences are still positive, and the conclusion is the same.
This changes the following numbered equations: 1.7, 1.8, 1.9, 1.13, 1.14.
55, line -2 Replace is drawn from subspace H1 by corresponds to an object in class C1.
56, line 2 Replace observation vector x is drawn from subspace H1 by observation vector x corresponds to an object in class C1.
56, line 5 Replace is drawn from subspace H1 by corresponds to an object in class C1.
56, lines -9, -5, and -1 Replace subspace by subset. These subsets may or not satisfy the requirements of a vector subspace.
60, figure Replace fX(x|C1) by pX(x|C1) and replace fX(x|C2) by pX(x|C2)
62, line 4 Replace 20 by 2 here or, better yet, delete this line.
The phrase size of the weight vector is ambiguous since w may or may not include the bias. The information is redundant, since m is always the same as the input vector size and must be 2 in this case.
62, Equation 1.39 Replace x on the summation symbol by x(n) and replace (-wTx) in the sum by (-wTx(n)d(n)).
Otherwise, samples from C2 will contribute negatively to the cost.
62, Equation 1.40 Replace x on the summation symbol by x(n) and replace (-x) in the sum by (-x(n)d(n))
65, Equation 1.42 Replace x on the summation symbol by x(n) and replace x in the sum by x(n)d(n)
66, line -7 Replace classic look by classic book
71, Equations 2.5 and 2.7 Unless d and x are independent, replace p D(d) by p D|X(d|x)
71, Equations 2.5 and 2.7 Unless w and x are independent, replace p W(w) by p W|X(w|x)
76, line 18 Replace Problem 2.4 by Problem 2.3
80, line 6 Replace rather the notion by rather than the notion
89, line 5 Note number 3 is not referenced in the text.
126, line -4 Replace multilayar by multilayer
126, line 10 Replace parallalization by parallelization
131, line 9 ReplaceE(n) by -∂E(n)
131, line 10 ReplaceE(n) by -∂E(n)
140, Fig. 4.7 The red circle at the intersection of the line below v1(1) and the line to the left of v3(1) should be deleted.
141, Eq. 4.47 Replace α[wji(l)(n-1)] by αwji(l)(n-1)]
153, line -2 Replace l = 0, 1, 2, ... by l = 1, 2, ...
173, line 21 Footnote reference 10 is a duplicate here. It also appears on page 183. The reference to Kearns is missing in the notes and references at the end of the chapter.
191, line 3 Footnote reference 12 is a duplicate here. It also appears on page 192. The reference to the conjugate-gradient method is missing in the notes and references at the end of the chapter.
211, Equation 4.168 Reverse the labels Approximation error and Estimation error in this equation as per Boutton 2007. He states The estimation error is determined by the number of training examples and by the capacity of the family of functions.
211, line -9 Replace approximation by estimation.
211, line -5 Replace estimation by approximation The term is used correctly on page 214 line -4.
212, Fig. 4.29 Replace the label Bound on the approximation error by Bound on the excess error
215, Eq. 4.171 Reverse the labels Approximation error and Estimation error in this equation as per Boutton 2007.
238, Eq. 5.15 Replace x by d
238, Eq. 5.16 Replace x by d
244, lines 8,11, Eq. 5.28 Replace subscript ranges j = i to K by j = 1 to K
245, Eq. 5.32 Replace x1, x2, ..., xK by μ1, μ2, ..., μK
245, Eq. 5.33 Replace both occurrences of xj by μj
247 Eq. 5.44 Replace the denominator of the fraction, namely, 1 + φ(n)R-1(n-1)φ(n), by 1 + φT(n)R-1(n-1)φ(n)
 
In addition, the numerator of the fraction, namely, R-1(n-1)φ(nT(n)R-1(n-1), should be replaced by R-1(n-1)φ(nT(n)R-1 T(n-1), else the note concerning the symmetry of R is unnecessary.

247, Eq. 5.45 In keeping with the changes to Equation 5.44, it may be preferable to express the denominator of the fraction as P(n-1)φ(nT(n)PT(n-1)
249, line -7 (unnumbered equation) Replace x1, x2, ..., xK by μ1, μ2, ..., μK
251, line -2 Replace distancd by distance
264, line 1 Replace dictomy by dichotomy
283, line -14 Replace integretion by integration
283, line -4 Replace number by form. There are, in fact, infinitely many Mercer kernels.
311, line -9 Replace Fig. P6.24 by P6.25.
311, line -6 Replace Fig. P6.24 by P6.25.
349, line 20 Replace so a to include by so as to include
350, line 13 Replace adjancy by adjacency
352, Equation 7.114 Replace fTF by adjacency fTf
354, line -6 to -5 In order for Equation 7.124 to work, all vectors must be length N. Replace
d = l-by-1 desired response vector
  = [d1, d2, ..., dl]T
by
d = N-by-1 desired response vector
  = [d1, d2, ..., dl, 0, 0, ..., 0]T
355, line 1 Replace L-by-L by N-by-N
378, Equation 8.28 Replace [ q1, q1, ..., qj] by [ q1, q2, ..., ql]
403, Equation 8.109 Replace λk by λr
454, line -3 Replace 0 ≤ yi by -∞ ≤ yi
454, Eq. 9.34 Replace log pyi by log pYi
459, line 5 Replace avereged by averaged.
472, line 8 Replace Γ(d, x) by Γ(α, x)
521, Eq. 10.93, last line Replace η[I - Φ(y)yTW-T by η[I - Φ(y)yT]W-T
522, line 6 Replace between the function G(C(n)s(n)) and the matrix product W-TA by between the product G(C(n)s(n))W-T and the matrix A
522, line -9 Replace W(n)(WT(n)W-T(n)) by W-T(n)WT(n)W(n)
575, line -17 Replace in Fig. 10.12 by in Fig. 10.9
672, line 3 Replace Laypunov by Lyapunov.
678, Eq. 13.8 Replace x by F(x)
701, Fig. 13.14 The arrowhead on one of the arcs between nodes labelled 2 and 3 should be reversed.