9:30, NEURAL-L2.1
OPTIMIZATION IN COMPANION SEARCH SPACES: THE CASE OF CROSS-ENTROPY AND THE LEVENBERG-MARQUARDT ALGORITHM
C. FANCOURT, J. PRINCIPE
We present a new learning algorithm for the supervised training of multilayer perceptrons for classification that is significantly faster than any previously known method. Like existing methods, the algorithm assumes a multilayer perceptron with a normalized exponential (softmax) output trained under a cross-entropy criterion. However, this output-criteria pairing turns out to have poor properties for existing optimization methods (backpropagation and its second order extensions) because second-order expansion of the network weights about the optimal solution is not a good approximation. The proposed algorithm overcomes this limitation by defining a new search space for which a second-order expansion is valid and such that the optimal solution in the new space coincides with the original criterion. This allows the application of the Levenberg-Marquardt search procedure to the cross-entropy criterion, which was previously thought applicable only to a mean square error criteria.

9:50, NEURAL-L2.2
ON LEARNING AND COMPUTATIONAL COMPLEXITY OF FIR RADIAL BASIS FUNCTION NETWORKS, PART I: LEARNING OF FIR RBFN'S
K. NAJARIAN
Recently, the complexity control of dynamic neural networks has gained significant attention from signal processing community. The performance of such a process depends highly on the applied definition of "medel complexity", i.e. complexity models that give simpler networks with better model accuracy and reliability are preferred. The learning theory creates a framework to assess the learning properties of models. In this paper (and the paper that accomapnies it), we apply the learning properties of FIR Radial Basis Function Networks (RBFN's) to introduce new complexity measures that reflect the learning properties of such neural models. Then, based on these complexity terms, we define cost functions which provide a balance between training and testing performances of the model, and give desirable levels of accuracy and confidence.

10:10, NEURAL-L2.3
ON LEARNING AND COMPUTATIONAL COMPLEXITY OF FIR RADIAL BASIS FUNCTION NETWORKS, PART II: COMPLEXITY MEASURES
K. NAJARIAN
Recently, the complexity control of dynamic neural networks has gained significant attention from signal processing community. The performance of such a process depends highly on the applied definition of "medel complexity", i.e. complexity models that give simpler networks with better model accuracy and reliability are preferred. The learning theory creates a framework to assess the learning properties of models. In this paper (and the paper that accomapnies it), we apply the learning properties of FIR Radial Basis Function Networks (RBFN's) to introduce new complexity measures that reflect the learning properties of such neural models. Then, based on these complexity terms, we define cost functions which provide a balance between training and testing performances of the model, and give desirable levels of accuracy and confidence.

10:30, NEURAL-L2.4
LEARNING TOPOGRAPHIC REPRESENTATION FOR MULTI-VIEW IMAGE PATTERNS
S. LI, X. LU, H. ZHANG, Q. FU, Y. CHENG
In 3D object detection and recognition from images, the object of interest is subject to view-point as well as illumination changes. In this paper, we consider the problem of learning view based features from a set of un-labeled images containing the appearances of the object viewed from various poses. Topographic Independent Component Analysis (TICA) is applied to produce a topographic map of basis components. Topographic ordering emerges as the iteration continues. The basis components are orderly spaced on the 2D map such that variations in viewing angle and in illumination can be clearly identified in the two axis directions. The components in each column, or columns of similar angle together, form a subspace so that a sample of that angle can be well represented as a linear combination of the basis components. This provides a view subspace representation for appearance based multi-view object detection and recognition.

10:50, NEURAL-L2.5
CLASSIFICATION BY PROBABILISTIC CLUSTERING
T. BREUEL
This paper describes an approach to classification based on a probabilistic clustering method. Most current classifiers perform classification by modeling class conditional densities directly or by modeling class-dependent discriminant functions. The approach described in this paper uses class-independent multi-layer perceptrons (MLPs) to estimate the probability that two given feature vectors are in the same class. These probability estimates are used to partition the input into separate classes in a probabilistic clustering method related to Markov Random Fields (MRFs). Classification by probabilistic clustering potentially offers greater robustness to different compositions of training and test sets than existing classification methods. Experimental results demonstrating the effectiveness of the method are given for an optical character recognition (OCR) problem. The relationship of the current approach to mixture density estimation, mixture discriminant analysis, and other OCR and handwriting recognition techniques is discussed.