Image analysis – Learning systems – Neural networks
Reexamination Certificate
1999-03-18
2003-02-25
Boudreau, Leo (Department: 2621)
Image analysis
Learning systems
Neural networks
C706S020000, C706S031000
Reexamination Certificate
active
06526168
ABSTRACT:
REFERENCE TO A MICROFICHE APPENDIX
Not Applicable
INCORPORATION BY REFERENCE
The following publications which are referenced herein using numbers in square brackets (e.g., [1]) are incorporated herein by reference:
[1] R. A. Jacobs and M. I. Jordan, “Adaptive Mixtures of Local Experts,”
Neural Computation
3, pp. 79-87, 1991.
[2] D. DeMers and G. Cotrell, “Non-Linear Dimensionality Reduction,”
Advances in Neural Information Processing Systems,
pp. 580-587, 1993.
[3] G. Cybenko, “Approximation by Superpositions of a Sigmodal Function,”
Mathematics of Control, Signals, and Systems,
Vol. 2, pp. 303-314, 1989.
[4] J. Mao and A. K. Jain, “Artificial Neural Networks for Feature Extraction and Multivariate Data Projection,”
IEEE Transactions on Neural Networks,
Vol. 6, No. 2, March 1995.
[5] J. Sklansky and L. Michelotti, “Locally Trained Piecewise Linear Classifiers,”
IEEE Transactions on Pattern Analysis and Machine
Intelligence, Vol. 6, No. 2, pp. 195-222, 1989.
[6] R. Caruana, “Learning Many Related Tasks at the Same Time,”
Advances in Neural Information Processing Systems
7, pp. 657-664, 1995.
[7] L. K. Hansen and P. Salamon, “Neural Network Ensembles,”
IEEE Transactions on Pattern Analysis and Machine Intelligence,
Vol. 12, No. 10, pp. 993-1001, October 1990.
[8] Y. Park, “A Comparison of Neural Net Classifiers and Linear Tree Classifiers: Their Similarities and Differences,”
Pattern Recognition,
Vol. 27, No. 11, pp. 1494-1503, 1994.
[9] T. Kohonen,
Self
-
Organization and Associative Memory,
Second Edition, Springer-Verlag, Berlin, 1988.
[10] E. Y. Tao and J. Sklansky, “Analysis of Mammograms Aided by Database of Images of Calcifications and Textures,” Proc. of 1996 SPIE Conf. on Medical Imaging—Computer-Aided Diagnosis, February 1996.
[11] B. Lofy, O. Pätz, M. Vriesenga, J. Bernarding, K. Haarbeck, J. Sklansky, “Landmark Enhancement for Spoke-Directed Anisotropic Diffusion,” Proc. of the IAPR Workshop on Methods for Extracting and Mapping Buildings, Roads, and other Man-Made Structures from Images, Technical University, Graz, Austria, September 1996.
[12] H. C. Zuckerman, “The Role of Mammography in the Diagnosis of Breast Cancer,” in
Breast Cancer, Diagnosis, and Treatment,
eds. I. M. Ariel and J. B. Cleary, Chap. 12, McGraw-Hill, N.Y. pp. 152-172, 1987.
[13] A. P. M. Forrest and R. J. Aitken, “Mammography Screening for Breast Cancer,”
Ann. Rev. Medicine
41, pp. 117-132, 1990.
[14] M. Vriesenga and J. Sklansky, “Genetic Selection and Neural Modeling of Piecewise-Linear Classifiers,”
International Journal of Pattern Recognition and Artificial Intelligence,
Vol. 10, No. 5, pp. 587-612, 1996.
BACKGROUND OF THE INVENTION
1. Field of the Invention
This invention pertains generally to classifiers constructed in the form of neural networks, and more particularly to neural classifiers that can map design data and decision curves on the same two-dimensional display.
2. Description of the Background Art
The applications to which neural networks can be applied continues to expand. Examples include medical analysis, character recognition, speech recognition, remote sensing, and geophysical prospecting among others.
An example of the use of neural networks for medical analysis can be found in U.S. Pat. No. 5,872,861 issued to Makram-Elbeid on Feb. 16, 1999, which is incorporated by reference herein. That patent describes a method for processing digital angiographic images for automatic detection of stenoses in blood vessels using a neural network. The digital image is processed in order to determine the central points and the edge points of the objects represented, provided that these objects constitute sufficiently uniform, contrasting masses on a sufficiently uniform background. The neural network with a hidden layer and two outputs is used to determine the probability that a potential stenosis is real or concerns a false alarm. The input of the neural network receives a vector whose components are characteristic traits of a candidate stenosis detected by means of the above method. The vector may be formed, for example by the intensities of the pixels of the icon of the candidate stenosis. The two outputs of the neural network encode the class of the non-stenoses (output
1
) and that of the stenoses (output
2
), respectively. Once reduced to the interval (
0
,
1
) by a mathematical transformation, the two activations of the output of the network can be interpreted as probabilities of association with either the class
1
or the class
2
, given the vector of characteristic traits (probabilities a posterior). The two probabilities are stored for each of the candidate stenoses. This enables the operator himself to define the degree of reliability, on a scale of probability, so as to retain or reject a candidate stenosis. The storage of the probabilities enables the user to try out several reliability levels without having to repeat the entire procedure for the detection and recognition of stenoses as described above. A graphic display method visualizes the stenosis retained in each of the individual images.
An example of neural networks applied to character recognition can be found in U.S. Pat. No. 5,859,925 issued to Yaeger et al. on Jan. 12, 1999, which is also incorporated herein by reference. As explained by Yaeger et al., various classification algorithms are available based on different theories and methodologies used in the particular area. In applying a classifier to a specific problem, varying degrees of success with any one of the classifiers have been obtained and, to improve the accuracy and success of the classification results, different techniques for combining classifiers have been studied. Nevertheless, problems of obtaining a high classification accuracy within a reasonable amount of time exist for the present classifying combination techniques and an optimal integration of different types of information is therefore desired to achieve high success and efficiency. Accordingly, combinations of multiple classifiers have been employed. However, none of the conventional approaches achieve the desired accuracy and efficiency in obtaining the combined classification result. The solution provided by Yaeger et al. is a classifying system having a single neural network in which multiple representations of a character are provided as input data. The classifying system analyzes the input representations through appropriate combination of their corresponding sets of data in the neural network architecture.
Another way to enhance classification performance is to use multi-expert neural classifiers [1]. This can result in computational complexity, so attempts have been made to use networks with two-neuron hidden layers. The prevailing view, however, is that networks with only two-neuron hidden layers do not have the capacity to perform large scale classification tasks and can only be used for exploratory data analysis [4] or data compression [2]. Therefore, there is a need for a classifier for networks with two-neuron hidden layers that combines information provided by several classification tasks into a visually meaningful and explanatory display, and that can display a large database of cases or objects.
BRIEF SUMMARY OF THE INVENTION
The present invention satisfies the foregoing needs by providing a neural classifier that combines information provided by several classification tasks into a visually meaningful and explanatory display. We refer to this as a “visual neural classifier.” Using the invention a designer can identify difficult-to-classify input patterns that may then be applied to an additional classification stage.
A visual neural classifier according to the invention comprises two major elements: (a) a set of experts and (b) a visualization network. Visualization is accomplished by a funnel-shaped multilayer dimensionality reduction network [2]. The dimensio
Ornes Chester
Sklansky Jack
O'Banion John P.
The Regents of the University of California
Werner Brian P.
LandOfFree
Visual neural classifier does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Visual neural classifier, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Visual neural classifier will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3115671