Search

Most pattern recognition tasks can be abstracted to a problem of utilizing comparisons between objects to perform the given inference task. Often these comparisons are in the form of a distance measure or dissimilarity. The design of appropriate comparison functions for particular inference tasks is an area of extensive research, and often rests on expert knowledge of the problem domain. If the data of interest come from two different sensors, or consist of very different types of data, a single dissimilarity may be inappropriate; instead, one might utilize several dissimilarities, each designed for a specific sensor or data stream. In this work we consider the problem of fusing information obtained from very different sensors or sources, encoded through the use of dissimilarity functions. Given n observations from source j, we have an n x n dissimilarity measure Dj, and we wish to utilize all this information in our inference. We describe several methods of utilizing these dissimilarity matrices that are based on embedding the observations into a single space. These methods optimize either the fidelity (whether the distances in the embedded space match the original dissimilarities) or the commensurability (whether matched objects from different sensors are close in the embedded space) or both. We discuss the properties of these embeddings, apply the idea to a problem in network modeling, and point out some interesting areas of further research., Presented at the 14th International Conference on Information Fusion held in Chicago, IL on 5-8 July 2011. Sponsored in part by Office of Naval Research and U.S. Army Research Laboratory.

Many applications require analyzing vast amounts of textual data, but the size and inherent noise of such data can make processing very challenging. One approach to these issues is to mathematically reduce the data so as to represent each document using only a few dimensions. Techniques for performing such "dimensionality reduction" (DR) have been well-studied for geometric and numerical data, but more rarely applied to text. In this paper, we examine the impact of five DR techniques on the accuracy of two supervised classifiers on three textual sources. This task mirrors important real world problems, such as classifying web pages or scientific articles. In addition, the accuracy serves as a proxy measure for how well each DR technique preserves the inter-document relationships while vastly reducing the size of the data, facilitating more sophisticated analysis. We show that, for a fixed number of dimensions, DR can be very successful at improving accuracy compared to using the original words as features. Surprisingly, we also find that one of the simplest DR techniques, Multi-dimensional Scaling (MDS), is among the most effective. This suggests that textual data may often lie upon a linear manifold where the more complex non-linear DR techniques do not have an advantage., Prepared in collaboration with the Naval Surface Warfare Center, Dahlgren, VA. Presented at the IEEE International Conference on Information Reuse and Integration, IRI 2007, Las Vegas, NV on 13-15 Aug 2007. Published in proceedings of the same, p348-353; ISBN 1-4244-1500-4.

Discovery in science is the generation of novel, interesting, plausible, and intelligible knowledge about the objects of study. Literature-related discovery (LRD) is the linking of two or more literature concepts that have heretofore not been linked (i.e., disjoint), in order to produce novel interesting, plausible, and intelligible knowledge (i.e., potential discovery)., Prepared in collaboration with Rush Medical College, Rush Univ. Medical Center, Chicago, IL and DDL-OMNI Engineering, LLC, McLean, VA.

Discovery in science is the generation of novel, interesting, plausible, and intelligible knowledge about the objects of study. Literature-related discovery (LRD) is the linking of two or more literature concepts that have heretofore not been linked to produce novel interesting, plausible, and intelligible knowledge (i.e., potential discovery). Two major variants of LRD are open discovery systems (ODS), where one starts with a problem and generates a potential solution (or vice versa), and closed discovery systems (CDS), where one starts with a problem and a potential solution and generates linking mechanisms. This report reviews the state-of-the-art in ODS LRD only. It examines the major LRD concepts, evaluates each concept in detail from the perspective of discovery capability, and examines the level of potential "discovery" reported in the literature from each concept's implementation. In the evaluation of potential discovery claimed in the published literature, a vetting process is used that requires that both characteristics of ODS LRD are present for potential discovery to be affirmed: concepts are linked that have not been linked previously, and novel, interesting, plausible, and intelligible knowledge is produced. The major conclusions are that, until recently, most of the reported ODS LRD techniques had not generated discovery, and this lack of discovery had hampered the growth of ODS LRD substantially. However, ODS LRD techniques have been developed that allow significantly greater amounts of potential discovery to be generated systematically.

Image analysis is performed by defining segmentation boundaries within an image by using wavelet theory or some other suitable method. Such boundaries can be incomplete, irregular, and/or multi-valued. The segmentation boundaries are then incorporated into feature calculations related to fractal dimensions for each pixel using a diffusion related method or a Dijkstra potential related method. These features are then used in statistical techniques to distinguish among textures or classes of interest. The system performing the image analysis is trained (or supervised) on data from different classes within an image or images. This enables the system to then later identify these classes in different images. The system can be used for Computer Aided Diagnosis (CAD) of mammograms or other medical imagery., Supersedes PAT-APPL-308 112-94, AD-D017 254.

We are developing a synthetic environment to be used in combination with real environments for the purpose of mine counter measures. Two examples of our work involve applications to land-based minefields and an application to submarine defense. In the former application, imaging is done with six spectral bands. It has been shown empirically that a sequence of images taken in six spectral bands when viewed sequentially will allow one to distinguish between real mines, partially buried real mines, decoys, other metallic objects, and other manner of debris. The image can be viewed as a 2-dimensional image with an 6-dimensional vector attached to each pixel location. We use the grand-tour technique (Wegman and Shen, 1993) to find an optimal discrimant between real mines and other objects. We then use a head-mounted display (HMD) which is semitransparent so that the real-world objects can be seen through it. After processing the scene, the suspected sites of real mines are superimposed on the visual field so that the soldier wearing the HMD is alerted to the presence of mines. The submarine application is similar. We use Crystal Eyes technology and Silicon Graphics Onyx systems in our laboratory.

A self-organizing neural network and method for classifying a pattern signature having N-features is provided. The network provides a posteriori conditional class probability that the pattern signature belongs to a selected class from a plurality of classes with which the neural network was trained. In its training mode, a plurality of training vectors is processed to generate an N-feature, N-dimensional space defined by a set of non-overlapping trained clusters. Each training vector has N-feature coordinates and a class coordinate. Each trained cluster has a center and a radius defined by a vigilance parameter. The center of each trained cluster is a reference vector that represents a recursive mean of the N-feature coordinates from training vectors bounded by a corresponding trained cluster. Each reference vector defines a fractional probability associated with the selected class based upon a ratio of (1) a count of training vectors from the selected class that are bounded by the corresponding trained cluster to (2) a total count of training vectors bounded by the corresponding trained cluster. In the exercise mode, an input vector defines the pattern signature to he classified. The input vector has N-feature coordinates associated with an unknown class. One of the reference vectors is selected so as to minimize differences with the N-feature coordinates of the input vector. The fractional probability of the selected one of the reference vectors is the a posteriori conditional class probability that the input vector belongs to the selected class. (KAR) p. 1, Supersedes PAT-APPL-936 369-92, AD-D016 455.

A discrete non-linear estimator is provided to establish a probability density function useful in classification of a feature as belonging to one of two classes. A first and second known feature space associated with respective first and second classes are each converted to a plurality of voltages reflecting corresponding first and second feature space bins. A first and second one-dimensional, non-linear resistive grid has input nodes that continuously receive the voltages reflecting the respective first and second known feature space. Each resistive grid produces an output voltage at each of a corresponding plurality of first and second output nodes. Each of a plurality of voltage comparators are used to compare voltages at corresponding first and second output nodes to output a discrete probability density function that considers corresponding ones of the first and second feature space bins. A display may be used to visually or otherwise provide each probability density function from each voltage comparator. (AN), Supersedes PAT-APPL-157 553-93.

A network is provided for the detection and correction of local boundary misalignments in a two-dimensional pixel space between a reference and transformed image. An input layer has input layer sections, each of which contains a plurality of input nodes associated with a cell. The cell is centered on a pixel and divided along a straightline orientation into first and second cell sections. Each of the input nodes outputs a digital signal indicative of the presence or absence of a contrast gradient as measured by the two cell sections. A second layer has a plurality of second layer sections, each of which is associated with one of the input layer sections and contains a plurality of second layer nodes. Each second layer node is responsive to a combination of input nodes to indicate the presence or absence of a boundary misalignment between the reference and transformed images. Presence of a contrast gradient at the combination of nodes defines a local boundary misalignment., Supersedes PAT-APPL-920 700.