1 / 11

Lecture 6. Learning (III): Associative Learning and Principal Component Analysis

Lecture 6. Learning (III): Associative Learning and Principal Component Analysis. Outline. Associative Learning: Hebbian Learning Use Associative Learning to compute Principal Component. Associative (Hebbian) Learning.

hunter
Download Presentation

Lecture 6. Learning (III): Associative Learning and Principal Component Analysis

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Lecture 6.Learning (III):Associative Learning and Principal Component Analysis

  2. Outline • Associative Learning: Hebbian Learning • Use Associative Learning to compute Principal Component (C) 2001 by Yu Hen Hu

  3. Associative (Hebbian) Learning • A Learning rule postulated by Hebb (1949) on the associative learning at the cellular level. D wkj(n) = wkj(n+1)  wkj(n) =  dk(n)xi(n) or D w(n) =  dk(n)x(n) The change of synaptic weights is proportional to the correlation of the (desired) activation dk(n) and the input x(n). • The weight vector w(n) converges to the cross-correlation of dk(n) and x(n): (C) 2001 by Yu Hen Hu

  4. Potential Applications of Associative Learning • Pattern Retrieval • Each pattern (input: x) is associated with a label (output: d). • The association information is stored in weights of a single layer network W. • When a test pattern is presented to the input, the label of a stored pattern that is closest to the test pattern should appear as the output. • This is the mechanism of an “associative memory” (C) 2001 by Yu Hen Hu

  5. M pairs of training vectors {(bm, am); m = 1, 2, ..., M}. Weight matrix: Retrieval: y=Wx. Let x = bk, Linear Associative Memory (C) 2001 by Yu Hen Hu

  6. Example • Consider 3 input patterns: x(1) = [1 2 –1]T, x(2) = [–1 1 1]T, x(3) = [1 0 1]T, and their corresponding 2D labels: y(1) = [–1 2]T, y(2) = [2 4]T, y(3) = [1 –3]T. • Compute the cross correlation matrix using the Hebbian learning,with learning rate  = 1: • Recall: Present x(2). Then y = W x(2) = [6 12]T = 3[2 4]T which is proportional to y(2)! (C) 2001 by Yu Hen Hu

  7. Hebbian Learning Modifications • Problem with Hebbian learning: when both dk(n) and xj(n) are positive, the weight wkj(n) will grow unbound. • Kohonen (1988) proposed to add a forgetting factor: Dwkj(n) = dk(n) xj(n) – a dk(n) wkj(n) or Dwkj(n) = adk(n)[c xj(n) – wkj(n)]; where c = h/a • Analysis: Suppose that xj(n) > 0 and dk(n) > 0. When wkj(n) increases until c xj(n) < wkj(n), Dwkj(n) < 0. This, in turn, will reduces wkj(n) for future n. • It can be shown that when |adk(n)| < 1, |wkj(n)| will be bounded if dk(n) and xj(n) are bounded. (C) 2001 by Yu Hen Hu

  8. Principal Component Analysis (PCA) • Pricipal component: part of signal (feature vectors) which consists of most energy of the signal! • Also known as Karhunen-Loeve Transformation • A linear transform of n-dim vector x, Tx such that ||x–Tx||2 is minimum among all n by n matrices T whose rank = p < n. • Applications: Feature dimension reduction Data Compression (C) 2001 by Yu Hen Hu

  9. PCA Method Given {x(k); 1  k  K}. Let dimension of x(k) = n. 1. Form a n by n Covariance matrix estimate 2. Eigenvalue decomposition 3. Principal components y(k) = VTpx(k) = [v1v2 … vp]Tx(k), p  n. Tx(k) = Vpy(k) = VpVTpx(k) (C) 2001 by Yu Hen Hu

  10. Optimality of Principal Components • Among all n by n matrices T' whose rank = p  n, || x – Tx ||2= li || x – T'x ||2 This is the orthogonality principle: • y(k) is the p-dimensional representation of x in the subspace spanned by columns of Vp. || y(k)|| = || Tx(k)|| is the length of the shadow of x(k). (C) 2001 by Yu Hen Hu

  11. y1 x1  y2 x2        yN xM Principal Component Network • Linear neuron model: Choose w to maximize E{||y||2} = wTE{xxT}w = wTCw subject to ||w||2 = 1. • Generalized Hebbian learning rule: ghademo.m, ghafun.m (C) 2001 by Yu Hen Hu

More Related