Method and apparatus for constructing, using and reusing components and structures of an artifical neural network

Patent metadata

Field	Value
Publication number	US-10019470-B2
Application number	US-201414513388-A
Country	US
Kind code	B2
Filing date	Oct 14, 2014
Priority date	Oct 16, 2013
Publication date	Jul 10, 2018
Grant date	Jul 10, 2018

Abstract

A method and apparatus for constructing a neuroscience-inspired artificial neural network (NIDA) or a dynamic adaptive neural network array (DANNA) or combinations of substructures thereof comprises one of constructing a substructure of an artificial neural network for performing a subtask of the task of the artificial neural network or extracting a useful substructure based on one of activity, causality path, behavior and inputs and outputs. The method includes identifying useful substructures in artificial neural networks that may be either successful at performing a subtask or unsuccessful at performing a subtask. Successful substructures may be implanted in an artificial neural network and unsuccessful substructures may be extracted from the artificial neural network for performing the task. The method and apparatus supports constructing, using and reusing components and structures of a neuroscience-inspired artificial neural network dynamic architecture in software and a dynamic adaptive neural network array.

First claim

What we claim is:

1. A method of constructing a neuromorphic network for use in a process control application for an external automated process or external device upon prediction of a component failure by a signal processor,
the neuromorphic network comprising a reconfigurable structure of components,
the structure of components comprising a connection of at least two different components comprising an addressably reconfigurable two-dimensional array of neuron and synapse circuit elements forming the neuromorphic network,
the structure of components comprising at least one addressably configured input neuron circuit element, an addressably configured output neuron circuit element, the addressably configured input and output neuron circuit elements located at an edge of the two-dimensional addressably reconfigurable array of circuit elements and the addressably configured synapse circuit element adapted to connect one addressably configured neuron circuit element to another addressably configured neuron circuit element to form an initial addressably configured array of an initial neuromorphic network,
the addressably configured synapse circuit element having a programmable delay or distance determined by a stored selected value of a delay/distance parameter,
a signal processor for processing input signals comprising one of a signal generated or sampled in the external automated process or external device and an error signal indicating a deviation of an external automated process signal from an expected value predicted by a model;
the structure of components and the value of the delay/distance parameter determined by an evolutionary optimization process, the structure of components configured in response to the external automated process or external device requiring control, the input neuron and output neuron circuit elements providing inputs or outputs to components external to the array comprising one of a display, a camera, a radio and a scanner of the external automated process or external device;
the addressably configured neuron circuit element generating at least two discrete events,
the addressably configured synapse circuit element, responsive to the addressably configured input neuron generating the at least two discrete events, permitting the at least two generated discrete events to propagate together through the addressably configured synapse circuit element,
the neuromorphic network performing the process control application for the external automated process or external device requiring control receiving a first output signal representing a measured output of the external automated process or external device and a second output signal predicted by a model of a desired behavior of the external automated process or external device, said neuromorphic network generating a process control signal via an interface and control structure connected to the input and output neuron circuit elements and a configuration control and optimizing device in response to said received first and second output signals, the process control signal being applied to the external automated process or external device to control the external automated process or external device upon the prediction of a component failure; and
the process control signal to further optimize a reconfiguration of the addressably configurable circuit elements of the initial neuromorphic network responsive to data received from the external automated process or external device via an input neuron circuit element.

2. The method of constructing a neuromorphic network of claim 1 ,
each of a plurality of connected addressably configurable neurons and synapses forming a two-dimensional dynamic addressably configurable array of circuit neuron and synapse circuit elements,
the addressably configurable synapse further having a programmable weight or strength parameter determining a value of charge arriving at a destination neuron circuit element that is addressably configurable.

3. The method of constructing a neuromorphic network of claim 1 ,
each of a plurality of connected addressably configurable neurons and synapses forming a three-dimensional array of circuit elements,
the addressably configurable synapse circuit element further having a programmable weight or strength parameter determining a value of charge arriving at a destination neuron that is an addressably configurable circuit element.

4. The method of constructing the neuromorphic network of claim 2 wherein each addressably configurable neuron circuit element comprises a programmable threshold of firing and a programmable refractory period,
the addressably configurable neuron circuit element firing if it is not in its refractory period, the firing of a neuron that is an addressably configurable circuit element generating one discrete event.

5. The method of constructing a neuromorphic network of claim 1 wherein the neuromorphic network comprises at least one substructure of addressably configurable neuron and synapse circuit elements for performing a computational sub-task of the process control application for the automated process, and
at least one substructure being an affective system of addressably configurable neuron and synapse circuit elements for performing an affective sub-task of the process control application for the automated process.

6. The method of constructing a neuromorphic network of claim 5 further comprising:
controlling the neuromorphic network construction using a programmed computer processor of the configuration control and optimizing device and an associated database of useful substructures of addressably configurable neuron and synapse circuit elements,
the useful substructures comprising at least one neuron and at least one synapse circuit element,
the useful substructures being measured by controlling the external automated process and addressably reconfiguring the neuromorphic network using the configuration control and optimizing device.

7. The method of constructing a neuromorphic network of claim 6 wherein said computer processor is adapted to extract a useful substructure for a computational sub-task of the process control application for the external automated process from the neuromorphic network and store the useful substructure for the computational sub-task in the associated database.

8. The method of constructing the neuromorphic network of claim 6 wherein said computer processor is adapted to implant a useful substructure of addressably reconfigurable neuron and synapse circuit elements for a computational sub-task of the process control application in the neuromorphic network,
the useful substructure retrieved from a plurality of useful substructures stored in the associated database.

9. The method of constructing the neuromorphic network of claim 6 ,
the programmed computer processor adapted to identify a useful substructure of addressably configurable neuron and synapse circuit elements based on decreasing error value of the computational sub-task of the useful substructure,
the programmed computer processor for measuring error of performing a task of the process control application of the external automated process and extract the substructure from the neuromorphic network as a useful substructure for storage in the associated database responsive to the decreasing error value.

10. A method of constructing a central pattern generator of a neuromorphic network for use in a process control application of an external automated process, the external automated system comprising one of an external automated system or an external device, the method comprising:
controlling the external automated system or external device upon prediction of a

Assignees

• Univ Tennessee Res Found

Inventors

• Birdwell J Douglas
• Dean Mark E
• Schuman Catherine

Classifications

• G06N3/065

  Analogue means · CPC title

• G06N3/086

  using evolutionary algorithms, e.g. genetic algorithms or genetic programming · CPC title

• G06N3/02

  Neural networks · CPC title

• G06N3/049

  Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs · CPC title

• G06N3/10 Primary

  Interfaces, programming languages or software development kits, e.g. for simulating neural networks · CPC title