Neural network for performing a relaxation process
First Claim
1. A neural network for implementing a probabilistic relaxation process comprisinga plurality of interconnected processing nodes, each node generating an output signal representing the time evolution of a probability that a particular object in a set of objects is a member of a particular class in a set of classes,said nodes being interconnected by weighted connection paths such that each node receives a time dependent input signal including a sum of weighted input signals of a plurality of said nodes and an externally generated input signal, the weighting of a connection path connecting two particular nodes being dependent on a compatibility that the corresponding objects of the nodes will be in the corresponding classes of the nodes.each of said processing nodes having a transfer function characteristic so that its output signal is a monotonic non-linear function of its input signal,wherein when the output signal of each node is first set to an initial estimated value of the probability that a particular object in a set of objects is a member of a particular class in a set of classes, said output signal of each node evolves over time due to the interconnection between said nodes to a constant final value indicative of whether or not a particular object is in a particular class,wherein said neural network has an energy function ##EQU18## wherein A and B are positive numbersc(i,j;
- h,k) is the compatibility that an object i is in a class j and an object h is in a class k.Vu is the output signal of a node with index ijVhk is the output signal of a node with index hkn is the number of objects.
2 Assignments
0 Petitions
Accused Products
Abstract
In accordance with the present invention, a neural network comprising an array of neurons (i.e. processing nodes) interconnected by synapses (i.e. weighted transmission links) is utilized to carry out a probabilistic relaxation process. The inventive neural network is especially suited for carrying out a variety of image processing tasks such as thresholding.
-
Citations
8 Claims
-
1. A neural network for implementing a probabilistic relaxation process comprising
a plurality of interconnected processing nodes, each node generating an output signal representing the time evolution of a probability that a particular object in a set of objects is a member of a particular class in a set of classes, said nodes being interconnected by weighted connection paths such that each node receives a time dependent input signal including a sum of weighted input signals of a plurality of said nodes and an externally generated input signal, the weighting of a connection path connecting two particular nodes being dependent on a compatibility that the corresponding objects of the nodes will be in the corresponding classes of the nodes. each of said processing nodes having a transfer function characteristic so that its output signal is a monotonic non-linear function of its input signal, wherein when the output signal of each node is first set to an initial estimated value of the probability that a particular object in a set of objects is a member of a particular class in a set of classes, said output signal of each node evolves over time due to the interconnection between said nodes to a constant final value indicative of whether or not a particular object is in a particular class, wherein said neural network has an energy function ##EQU18## wherein A and B are positive numbers c(i,j; - h,k) is the compatibility that an object i is in a class j and an object h is in a class k.
Vu is the output signal of a node with index ij Vhk is the output signal of a node with index hk n is the number of objects. - View Dependent Claims (2, 3, 4, 5)
- h,k) is the compatibility that an object i is in a class j and an object h is in a class k.
-
6. In a neural network comprising a plurality of processing nodes interconnected by weighted interconnection paths, wherein the input signal to each processing node includes a weighted sum of the output signals of a plurality of said processing nodes and an external signal, wherein the output signal of each processing node is a probability that a particular object in a set of objects is a member of a particular class in a set of classes, and is a non-linear monotonic function of the input signal, and wherein the weight of each interconnection path between two nodes is dependent on a compatibility that the corresponding objects of the two nodes are in the corresponding classes, a method for carrying out a probabilistic relaxation comprising,
setting the output signal of each node to an initial estimated value of the probability that a particular object in a set of objects is in a particular class in a set of classes, and causing the output signal of each of said nodes to evolve in time because of the transmission of signals via the interconnection paths between said nodes to a final state indicative of whether or not the particular object of each node is in the particular class, wherein wherein said neural network has an energy function ##EQU19## wherein A and B are positive numbers c(i,j; - h,k) is the compatibility that an object i is in a class j and an object h is in a class k,
Vij is the output signal of a node with index ij Vhk is the output signal of a node with index hk n is the number of objects. - View Dependent Claims (7)
- h,k) is the compatibility that an object i is in a class j and an object h is in a class k,
-
8. An electronic network comprising
electronic means for generating a plurality of time dependent output signals, each of said output signals initially representing a first estimated value of the probability that a particular object in a set of objects is a member of a particular class in a set of classes and evolving in time to a final state which indicates definitively whether the particular object is in the particular class, said electronic means comprising a plurality of processing nodes, each of which generates one of said output signals, and means for forming an input signal for each of said nodes by multiplying the output signal of each of a plurality of said nodes by a compatibility coefficient indicative of the compatibility between the object and class of the node from which the output signal is taken and the object and class of the node for which the input signal is being generated and summing the resulting products, each of said nodes having a non-linear transfer function between its input and its output, wherein said neural network has an energy function ##EQU20## wherein A and B are positive numbers c(i,j; - h,k) is the compatibility that an object i is in a class j and an object h is in a class k,
Vij is the output signal of a node with index ij Vhk is the output signal of a node with index hk n is the number of objects.
- h,k) is the compatibility that an object i is in a class j and an object h is in a class k,
Specification