Data processing: artificial intelligence – Neural network – Learning method
Reexamination Certificate
1999-09-01
2003-04-22
Starks, Jr., Wilbert L. (Department: 2122)
Data processing: artificial intelligence
Neural network
Learning method
C706S015000, C706S016000, C706S027000
Reexamination Certificate
active
06553357
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Field of the Invention
This invention relates to the field of information processing, and in particular to machine learning, neural networks, and evolutionary algorithms.
2. Description of Related Art
Neural networks are commonly employed as learning systems. Neural networks can be structured in a variety of forms; for ease of understanding, a feed-forward neural network architecture is used herein as a paradigm for neural networks, although the application of the principles presented herein will be recognized by one of ordinary skill in the art to be applicable to a variety of other neural network architectures. A typical feed-forward neural network comprises one or more input nodes, one or more output nodes, and a plurality of intermediate, or hidden, nodes that are arranged in a series of layers between the input and output nodes. In a common neural net architecture, each input node is connected to one or mode hidden nodes in a first layer of nodes, each hidden node in the first layer of nodes is connected to one or more hidden nodes in a second layer of nodes, and so on until each node of the last layer of hidden nodes is connected to each output node. The output of each node is typically a nonlinear function of a weighted combination of each input to the node. In a feedforward neural net, when a set of input values is applied to the input nodes, the weighted values are propagated through each layer of the network until a resultant set of output values is produced. Other configurations of nodes, interconnections, and effect propagation are also common. For example, in some architectures, a node may be connected to one or more other nodes beyond its immediately adjacent layer.
In a learning mode, the resultant set of output values is compared to the set of output values that a properly trained network should have produced, to provide an error factor associated with each output node. In the case of pattern matching, for example, each output node may represent a likelihood that the input pattern corresponds to a particular class. Each input pattern is pre-categorized to provide an “ideal” set of likelihood factors, and the error factor is a measure of the difference between this “ideal” set and the set of output node values that the neural network produced. The error factor is propagated back through the network to modify the weights of each input to each node so as to minimize a composite of the error factors. The composite is typically the sum of the square of the error factor at each output node. Conceptually, the node weights that contributed to the outputs of the incorrect class are reduced, while those that contributed to the output of the correct class are increased.
Although the error factor can be propagated back based on each comparison of the ideal output and the result of processing each input set, preferably, a plurality, or batch, of input sets of values is applied to the network, and an accumulated error factor is back-propagated to readjust the weights. Depending upon the training technique employed, this process may be repeated for additional sets or batches of input values. The entire process is repeated for a fixed number of iterations or until subsequent iterations demonstrate a convergence to the “ideal”, or until some other termination criterion is achieved. Once the set of weights is determined, the resultant network can be used to process other items, items that were not part of the training set, by providing the corresponding set of input values from each of the other items, to produce a resultant output corresponding to each of the other items.
The performance of the neural network for a given problem set depends upon a variety of factors, including the number of network layers, the number of hidden nodes in each layer, and so on. Given a particular set of network factors, or network architecture, different problem sets will perform differently. U.S. Pat. No. 5,140,530 “GENETIC ALGORITHM SYNTHESIS OF NEURAL NETWORKS”, issued Aug. 18, 1992 to Guha et al, and incorporated by reference herein, presents the use of a genetic algorithm to construct an optimized custom neural network architecture. U.S. Pat. No. 5,249,259 “GENETIC ALGORITHM TECHNIQUE FOR DESIGNING NEURAL NETWORKS”, issued Sep. 28, 1993 to Robert L. Harvey, and incorporated by reference herein, presents the use of a genetic algorithm to select an optimum set of weights associated with a neural network.
Genetic algorithms are a specific class of evolutionary algorithms and the term evolutionary algorithm is used hereinafter. Evolutionary algorithms are commonly used to provide a directed trial and error search for an optimum solution wherein the samples selected for each trial are based on the performance of samples in prior trials. In a typical evolutionary algorithm, certain attributes, or genes, are assumed to be related to an ability to perform a given task, different combinations of genes resulting indifferent levels of effectiveness for performing that task. The evolutionary algorithm is particularly effective for problems wherein the relation between the combination of attributes and the effectiveness for performing the task does not have a closed form solution.
In an evolutionary algorithm, the offspring production process is used to determine a particular combination of genes that is most effective for performing a given task. A combination of genes, or attributes, is termed a chromosome. In the genetic algorithm class of evolutionary algorithms, a reproduction-recombination cycle is used to propagate generations of offspring. Members of a population having different chromosomes mate and generate offspring. These offspring have attributes passed down from the parent members, typically as some random combination of genes from each parent. In a classic genetic algorithm, the individuals that are more effective than others in performing the given task are provided a greater opportunity to mate and generate offspring. That is, the individuals having preferred chromosomes are given a higher opportunity to generate offspring, in the hope that the offspring will inherit whichever genes allowed the parents to perform the given task effectively. The next generation of parents is selected based on a preference for those exhibiting effectiveness for performing the given task. In this manner, the number of offspring having attributes that are effective for performing the given task will tend to increase with each generation. Paradigms of other methods for generating offspring, such as asexual reproduction, mutation, and the like, are also used to produce offspring having an increasing likelihood of improved abilities to perform the given task.
As applied to neural networks, the chromosome of the referenced '530 (Guha) patent represents the architecture of a neural network. Alternative neural networks, those having different architectures, each have a corresponding different chromosome. After a plurality of neural networks have been trained, each of the networks is provided evaluation input sets, and the performance of each trained neural network on the evaluation input sets is determined, based on a comparison with an “ideal” performance corresponding to each evaluation input set. The chromosomes of the better performing trained neural networks are saved and used to generate the next set of sample neural networks to be trained and evaluated. By determining each next generation of samples based on the prior successful samples, the characteristics that contribute to successful performance are likely to be passed down from generation to generation, such that each generation tends to contain successively better performers.
The speed at which a particular neural network converges to an optimal set of weights is highly dependent upon the initial value of the weights in the neural network. Similarly, the likelihood of a particular neural network converging on a “global” optimum, rather than a “local” optimum, is highly dependent upon the initial value of the weights
Eshelman Larry J.
Mathias Keith E.
Schaffer J. David
LandOfFree
Method for improving neural network architectures using... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method for improving neural network architectures using..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method for improving neural network architectures using... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3088281