Image analysis – Learning systems – Trainable classifiers or pattern recognizers
Reexamination Certificate
2000-08-17
2004-02-17
Mariam, Daniel (Department: 2623)
Image analysis
Learning systems
Trainable classifiers or pattern recognizers
C382S197000, C382S224000
Reexamination Certificate
active
06694049
ABSTRACT:
BACKGROUND OF THE INVENTION
(1) Field of the Invention
The present invention relates generally to the field of electronic neural networks, and more particularly to a new architecture for neural networks having a plurality of hidden layers, or multi-layer neural networks, and further to a new neural network processor for classifying patterns in optical image data, or other arrays of input data having one or more input dimensions.
(2) Description of the Prior Art
Electronic neural networks have been developed to rapidly identify patterns in certain types of input data, or to accurately classify the input patterns into one of a plurality of predetermined classifications. For example, neural networks have been developed which can recognize and identify patterns, such as the identification of hand-written alphanumeric characters, in response to input data constituting the pattern of on/off picture elements, or “pixels,” representing the images of the characters to be identified. In such a neural network, the pixel pattern is represented by, for example, electrical signals coupled to a plurality of input terminals, which, in turn, are connected to a number of processing nodes, or neurons, each of which is associated with one of the alphanumeric characters which the neural network can identify. The input signals from the input terminals are coupled to the processing nodes through certain weighting functions, and each processing node generates an output signal which represents a value that is a non-linear function of the pattern of weighted input signals applied thereto. Based on the values of the weighted pattern of input signals from the input terminals, if the input signals represent a character, which can be identified by the neural network, one of the processing nodes that is associated with that character will generate a positive output signal, and the others will not. On the other hand, if the input signals do not represent a character, which can be identified by the neural network, none of the processing nodes will generate a positive output signal. Neural networks have been developed which can perform similar pattern recognition in a number of diverse areas.
The particular patterns that the neural network can identify depend on the weighting functions and the particular connections of the input terminals to the processing nodes, or elements. As an example, the weighting functions in the above-described character recognition neural network essentially will represent the pixel patterns that define each particular character. Typically, each processing node will perform a summation operation in connection with the weight values, also referred to as connection values or weighting values, representing the weighted input signals provided thereto, to generate a sum that represents the likelihood that the character to be identified is the character associated with that processing node. The processing node then applies the non-linear function to that sum to generate a positive output signal if the sum is, for example, above a predetermined threshold value. The non-linear functions, which the processing nodes may use in connection with the sum of weighted input signals, are generally conventional functions, such as step functions, threshold functions, or sigmoids. In all cases the output signal from the processing node will approach the same positive output signal asymptotically.
Before a neural network can be useful, the weighting functions for a set of the respective input signals must be established. In special cases, the weighting functions can be established a priori. Normally, however, a neural network goes through a training phase, in which input signals representing a number of training patterns for the types of items to be classified (e.g., the pixel patterns of the various hand-written characters in the character-recognition example) are applied to the input terminals, and the output signals from the processing nodes are tested. Based on the pattern of output signals from the processing nodes for each training example, the weighting functions are adjusted over a number of trials. Once trained, a neural network can generally accurately recognize patterns during an operational phase. The degree of success is based in part on the number of training patterns applied to the neural network during the training stage and the degree of dissimilarity between patterns to be identified. Such a neural network can also typically identify patterns that are similar to the training patterns.
One of the problems with conventional neural network architectures as described above is that the training methodology, generally known as the “back-propagation” method, is often extremely slow in a number of important applications. Also, under the back-propagation method, the neural network may provide erroneous results, which may require restarting the training. In addition, even after a neural network has been through a training phase, confidence that the best training has been accomplished may sometimes be poor. If a new classification is to be added to a trained neural network, the complete neural network must be retrained. Further, the weighting functions generated during the training phase often cannot be interpreted in ways that readily provide understanding of what they particularly represent.
In my related patent application entitled “NEURAL DIRECTORS” (Ser. No. 09/436,957, which is now U.S. Pat. No. 6,618,713), incorporated herein in its entirety by reference, a new neural network architecture, or neural director, was described in which the weighting functions may be determined a priori, i.e., the new neural network architecture is constructed rather then trained. The neural director has an input processing node layer, which receives the input vector X and an output processing node layer, which generates the output vector Y. In a type
1
neural director containing linear neurons, the connections between the input and output processing node layers are a unique weighting set w(i,j) that contains an internal representation of a uniform spatial distribution of “J” unit vectors throughout a unit sphere of “I” dimensions. Thus the cosine value between any two adjacent unit vectors is a constant everywhere in the unit sphere. A type
1
neural director is thus described as linear in both its neural circuit, i.e., classically linear, and in its space, i.e., spatially linear. A type
2
neural director, is generally classically linear but spatially nonlinear, though it will be understood that either classic or spatial nonlinearity will result in a neural director type
2
. A spatial nonlinearity causes an input vector pair to diverge in direction in the output space and is analogous to a system nonlinearity in chaos theory where two similar initial condition points diverge over time. In the case of spatial nonlinearity, the system divergence occurs as the input data flows through repetitious stages of nonlinearity versus a chaotic system recursion over time. One of the many important characteristics of a constructed neural network is that a classification of an input pattern is greatly defined by a vector's direction in a multidimensional space. Reduced to its most basic concept, a constructed neural network senses features from a specific input pattern to provide a deterministic direction through a connecting circuit as a feature vector. This deterministic direction in a multidimensional space is the information that is used for the recognition and classification of the pattern. When compared to a neural director type
1
of the same input and output dimensions, a neural director type
2
nonlinearly shifts an input vector away from the output direction which one would anticipate using the neural director type
1
. A neural director type
2
produces a nonlinear gradient between two poles it its multidimensional output space, one pole lying in the center of a sub space that is directed by all positive elements and the other pole being the opposite polarity. The spatial nonlinearities of the type
Kasischke James M.
Mariam Daniel
Nasser Jean-Paul A.
Oglo Michael F.
The United States of America as represented by the Secretary of
LandOfFree
Multimode invariant processor does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Multimode invariant processor, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Multimode invariant processor will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3298215