We introduce an object recognition system, based on the well known Elastic Graph Matching (EGM), but includes significant improvements compared to earlier versions. Our basic features are banana wavelets, which are generalized Gabor wavelets. In addition to the qualities frequency and orientation, banana wavelets have the attributes curvature and size. Banana wavelets can be metrically organized. A sparse and efficient representation of object classes is learned utilizing this metric organization. Learning is guided by a sensible amount of a priori knowledge in form of basic principles. The learned representation is used for a fast matching. Significant speed up can be achieved by hierarchical processing of features. Furthermore manual construction of ground truth is replaced by an automatic generation of suitable training examples using motor controlled feedback. We motivate the biological plausibility of our approach by utilizing concepts like hierarchical processing or metrical orga...

Random input patterns induce a partition of the coupling space of feed-forward neural networks into different cells according to the generated output sequence. For the perceptron this partition forms a random multifractal for which the spectrum f(α) can be calculated analytically using the replica trick. Phase transition in the multifractal spectrum correspond to the crossover from percolating to nonpercolating cell sizes. Instabilities of negative moments are related to the VC-dimension. PACS numbers: 02.50.-r, 64.60.Ak, 87.10.+e Multifractal concepts were originally introduced in the context of developed turbulence [1] and chaotic dynamical systems [2] and have become since then a standard tool to analyze physical systems with richer structure than that induced by dilation symmetry alone (for reviews see [3]). Contrary to simple scale invariant situations as provided, e.g., by systems at second order

This paper describes a method to select a suitable feature for speech recognition using information theoretic measure. Conventional speech recognition systems heuristically choose a portion of frequency components, cepstrum, mel-cepstrum, energy, and their time differences of speech waveforms as their speech features. However, these systems never have good performance if the selected features are not suitable for speech recognition. Since the recognition rate is the only performance measure of speech recognition system, it is hard to judge how suitable the selected feature is. To solve this problem, it is essential to analyze the feature itself, and measure how good the feature itself is. Good speech features should contain all of the classrelated information and as small amount of the class-irrelevant variation as possible. In this paper, we suggest a method to measure the class-related information and the amount of the class-irrelevant variation based on the Shannon's information the...

To study local structures in proteins, we previously developed an autoassociative artificial neural network (autoANN) and clustering tool to discover intrinsic features of macromolecular structures. The hidden unit activations computed by the trained autoANN are a convenient low-dimensional encoding of the local protein backbone structure. Clustering these activation vectors results in a unique classification of protein local structural features called Structural Building Blocks (SBBs). Here we describe application of this method to a larger database of proteins, verification of the applicability of this method to structure classification, and subsequent analysis of amino acid frequencies and several commonly occurring patterns of SBBs. The SBB classification method has several interesting properties: 1) it identifies the regular secondary structures, a helix and b strand; 2) it consistently identifies other local structure features (e.g., helix caps and strand caps); 3) strong amino acid preferences are revealed at some positions in some SBBs; and 4) distinct patterns of SBBs occur in the "random coil" regions of proteins. Analysis of these patterns identifies interesting structural motifs in the protein backbone structure, indicating that SBBs can be used as "building blocks" in the analysis of protein structure. This type of pattern analysis should increase our understanding of the relationship between protein sequence and local structure, especially in the prediction of protein structures. Proteins 27:249--271 r 1997 Wiley-Liss, Inc. Key words: protein structure; secondary structure; protein conformation; protein backbone structure; protein structure classification; helix capping; strand capping; neural networks; structural building blocks