Chapter 10 Information-Theoretic Learning Models 《神经网络与 及机器学习》 .pptVIP

Chapter 10 Information-Theoretic Learning Models;Figure 10.1 Relationships embodied in the three lines of Eq. (10.32), involving the mutual information I(X; Y).;Figure 10.2 Four basic scenarios that lend themselves to the application of information maximization and its three variants.;Figure 10.3 Signal-flow graph of a noisy neuron.;Figure 10.4 Another noisy model of the neuron.;Figure 10.7 Block diagram of a neural processor, the goal of which is to suppress background clutter using a pair of polarimetric, noncoherent radar inputs; clutter suppression is attained by minimizing the mutual information between the outputs of the two modules.;Figure 10.8 Application of the Imin principle to radar polarimetry. (a) Raw B-scan radar images (azimuth plotted versus range) for horizontal– horizontal polarization (top) and horizontal–vertical (bottom) polarization. (b) Composite image computed by minimizing the mutual information between the two polarized radar images of part (a).;Figure 10.9 Block diagram of the processor for solving the blind source separation problem. The vectors s, x, and y are values of the respective random vectors S, X, and Y.;Figure 10.10 Two Gaussian distributed processes.;Figure 10.11 Gaussian- and uniformly-distributed processes.;Figure 10.12 The result of applying the sparse-coding algorithm to a natural image. (The figure is reproduced with the permission of Dr. Bruno Olshausen.);Figure 10.13 The result of applying the Infomax algorithm for ICA to another natural image. (The figure is reproduced with the permission of Dr. Anthony Bell.);Figure 10.14 Signal-flow graph of the blind source separation learning algorithm described in Eqs. (10.85) and (10.104): The block labeled z–1I represents a bank of uni-time delays. The graph embodies a multiplicity of feedback loops.;Figure 10.15 Illustration of the relationship between the maximum-likelihood method and independent-components analysis for blind source separation. The ma