Method and apparatus for pattern mapping system with self-reliability check

• US 5,335,291 A
• Filed: 12/09/1993
• Issued: 08/02/1994
• Est. Priority Date: 09/20/1991
• Status: Expired due to Fees

• Current Assignee: Massachusetts Institute of Technology
• Original Assignee: Massachusetts Institute of Technology
• Inventors: Leonard, James A., Ungar, Lyle H., Kramer, Mark A.
• Primary Examiner(s): Moore, David K.
• Assistant Examiner(s): Johns, Andrew W.
• Application Number:
  US08/165,067
• Time in Patent Office: 236 Days
• Field of Search: 382/14, 382/15, 382/57, 395/21, 395/22, 395/23, 395/50, 395/51, 395/55
• US Class Current: 382/158
• CPC Class Codes: G06N 3/047 Probabilistic or stochastic networks
• CPC Class Codes: G06V 30/194 References adjustable by an adaptive method, e.g. learning

Abstract

A pattern mapping system provides indication of reliability of pattern or category selection indicated by output nodes. The indication of reliability includes a quantitative indication of error in the output and an indication when training is insufficient for the system to respond to the input pattern accurately. The error indication is based on error in the outputs produced by the system as trained in response to given inputs, and accuracy of fit of training data to the output. The insufficiency of the mapping system as trained is determined according to training data density local to the subject input pattern.

Claims

1. A pattern mapping system comprising:

an input layer for presenting an input pattern;

a working layer of working nodes connected to the input layer for receiving the input pattern and generating activation signals in accordance with predefined training parameters, and (b) statistical signals associated with each activation signal;

an output layer of output nodes coupled to the working layer for providing from the activation signals an output pattern indicative of the input pattern; and

a node assembly coupled to the working layer for receiving the statistical signals and in response to the statistical signals providing an indication of reliability of the output pattern as a function of the predefined training parameters.

2. A system as claimed in claim 1 wherein the node assembly includes:

at least a first node for providing an indication that the predefined training parameters are insufficient for the system to adequately respond to the input pattern; and

at least a second node for providing a quantitative indication of error in the output pattern as a function of error in outputs of the working layer nodes established by the predefined training parameters.

3. A system as claimed in claim 2 wherein the second node provides a range of uncertainty of the output pattern as output from the output node.

4. A system as claimed in claim 2 wherein the node assembly further comprises plural summing nodes coupled between the working layer and the first and second nodes, one summing node being common to the first and second nodes.

5. A system as claimed in claim 1 wherein the output pattern identifies category of the input pattern such that the system is a pattern classification system.

6. A system as claimed in claim 1 wherein the output pattern is a functional mapping of the input pattern such that the system is a functional approximation system.

7. In a pattern mapping system, a method of identifying patterns comprising the steps of:

presenting input signals in a presentation field, the input signals defining an input pattern;

providing a plurality of working nodes, each for identifying patterns with respect to a plurality of predefined training parameters and generating at least one statistical signal relating the pattern identification with the predefined training parameters, the working nodes being coupled to the presentation field for receiving the input signals;

providing a plurality of output nodes coupled to the pattern identifications from the working nodes such that the output nodes provide output patterns indicative of the input pattern; and

providing an output indication of reliability of the output patterns provided by the output nodes, the output indication of reliability being based on the statistical signals of the working nodes.

8. A method as claimed in claim 7 wherein the step of providing an output indication of reliability includes:

providing an indication of sufficiency of the predefined training parameters for the mapping system to respond to the input pattern.

9. A method as claimed in claim 7 wherein the step of providing an output indication of reliability includes:

providing a quantitative indication of accuracy of the output pattern with respect to accuracy of the working nodes established by the predefined training parameters.

10. A method as claimed in claim 7 wherein the output nodes provide indications of categories which include the input pattern;

and

the step of providing an output indication of reliability includes providing an indication of sufficiency of the predefined training parameters for the mapping system to respond to the input pattern.

11. A method as claimed in claim 7 wherein the step of providing an output indication of reliability includes:

providing a quantitative indication of error in the output pattern; and

providing an indication when the predefined training parameters are insufficient for the mapping system to respond to the input pattern accurately.

12. A pattern mapping system comprising:

a plurality of input nodes for providing signals defining an input pattern;

a layer of working nodes, each working node corresponding to a different range of target patterns from training data and each working node having (a) an activation signal in response to input patterns within the corresponding range of target patterns and (b) at least one statistical signal associated with the activation signal based on the training data, the layer of working nodes coupled to the input nodes to receive the signals defining the input pattern, and in response, each working node having a corresponding range of target patterns which includes the input pattern producing a respective activation signal;

a plurality of output nodes, each coupled to receive and sum the activation signals of the working nodes, such that the plurality of output nodes provide an output pattern indicative of the input pattern; and

an output assembly coupled to receive the at least one statistical signal from the working nodes, and in response to the at least one statistical signal the output assembly providing at least one measure of reliability of the output pattern as a function of the training data.

13. A system as claimed in claim 12 wherein the output assembly includes:

at least a first node for providing a measure of amount of training data involved in the ranges of target patterns of the working nodes producing activation signals in response to the input pattern, the involved training data being local to the input pattern and forming local training data; and

a second node for providing a measure of accuracy of fit of the output pattern to the local training data.

14. A system as claimed in claim 13 wherein the first node provides an indication of insufficient training data for the system to respond to the input pattern when the measure of amount of local training data falls below a predetermined threshold.

15. A system as claimed in claim 13 wherein the first node provides an indication of training data density.

16. A system as claimed in claim 13 wherein the second node comprises a ratio node for calculating output error in the output pattern by averaging contributions of the working nodes to the output error weighted by contribution of the working node in producing the output pattern.

17. A system as claimed in claim 13 wherein the output assembly further comprises plural summing nodes coupled between the layer of working nodes and the first and second nodes, one summing node being common to the first and second nodes.

18. A system as claimed in claim 13 wherein the output assembly further comprises a maximum activation node for indicating maximum activation of the working nodes.

19. A system as claimed in claim 12 wherein the working nodes are radial basis function units having respective active signals defined according to a Gaussian density function.