Method and apparatus for providing real-time monitoring of an artifical neural network

Patent metadata

Field	Value
Publication number	US-10248675-B2
Application number	US-201414513447-A
Country	US
Kind code	B2
Filing date	Oct 14, 2014
Priority date	Oct 16, 2013
Publication date	Apr 2, 2019
Grant date	Apr 2, 2019

Abstract

A circuit element of a multi-dimensional dynamic adaptive neural network array (DANNA) may comprise a neuron/synapse select input functional to select the circuit element to function as one of a neuron and a synapse. In one embodiment of a DANNA array of such circuit elements, (wherein a circuit element may be digital), a destination neuron may be connected to a first neuron by a first synapse in one dimension a second destination neuron may be connected to the first neuron by a second synapse in a second dimension to form linked columns and rows of neuron/synapse circuit elements. In one embodiment, the rows and columns of circuit elements have read registers that are linked together by signal lines and clocked and controlled so as to output columnar data to an output register when a neuron/synapse data value is stored in the read register.

First claim

What we claim is:

1. Apparatus for configuring a neuromorphic network comprising an artificial neural network for implementing a solution to one of a control, detection or classification application, the artificial neural network comprising:
an artificial neural network configuration structure, an interface and control structure for connecting a two-dimensional array to an external process and the configuration structure and interface and control structure connected to a control and optimizing device for configuring the two-dimensional array, the two-dimensional array comprising addressable circuit elements and the neuromorphic network for optimizing the configuration of the two-dimensional array and for controlling connection to the external process;
the two-dimensional array of interconnected, addressable circuit elements, the two-dimensional array comprising X columns and Y rows of circuit elements where X and Y are integers greater than one, the interconnection of circuit elements of the multi-dimensional array being programmable by a special purpose computer program to implement a solution to one of the control, detection and classification application of the neuromorphic network for use with the external process,
each circuit element having the same components,
each circuit element addressably configured by the configuration structure, interface and control structure and optimizing device responsive to the external process, under special purpose program control, as selectively operating as one of a neuron and a synapse function or used as a pass-thru circuit element or excluded from the artificial neural network,
under special purpose program control, at least one circuit element of the array selectively configured as one of an input neuron and an output neuron at an edge of the two-dimensional array for internal array connection and external process connection, and,
under program control, one to multiple circuit elements of the two-dimensional array selectively operating as the neuron circuit element or the synapse circuit element being connected between the input neuron at the edge of the array or the synapse circuit element and the output neuron at the edge of the array or the synapse circuit element,
the circuit elements formed in an array of rows and columns with signal lines connecting addressably programmable circuit elements oft least two of the Y rows of programmably selected circuit elements and signal lines connecting at least two of the X columns of programmably selected circuit elements to provide monitoring of the selectively configured two-dimensional array of programmably addressed circuit elements having one of a programmable neuron and a programmable synapse function, certain of the programmably addressable circuit elements used as a pass-thru circuit element or others excluded from the artificial neural network comprising the neuromorphic network,
a given addressed circuit element of the neuromorphic network performing one of a neuron and a synapse function at a particular row and column number selectively reporting its output to an output register of a monitoring circuit at the same particular row and column number when a value is stored in a read register of the monitoring circuit for monitoring of the programmably addressed circuit element of the artificial network array of the neuromorphic network of the at least two of the X columns and Y rows of programmably addressed circuit elements interconnected to implement one of a control, detection and classification application responsive to the external process, the programmably addressed circuit elements of the neuromorphic network being configurable to implement either a neuron or a synapse function, the configuration of programmable addressed circuit elements being addressably configurable and optimized by the control and optimizing device.

2. The apparatus of claim 1 , the selected elements of the array comprising a programmable logic array, ASIC or VLSI component.

3. The apparatus of claim 2 , defined by a hardware description language, circuit diagram and/or logic equations.

4. The apparatus of claim 3 wherein said hardware description language is one of VHDL and Verilog.

5. The apparatus of claim l wherein at least one addressably configured circuit element comprises a digital circuit component, the digital circuit component comprising an accumulator for accumulating a charge value when the at least one addressably configured circuit element is addressably configured as performing a neuron function and the accumulator for storing a value of synaptic weight when the at least one circuit element is addressably configured to perform a synapse function under special purpose program control.

6. The apparatus of claim I wherein at least one circuit element contains one of an analog and a digital electronic device, the electronic device for storing a value of a long-term depression/long-term potentiation refractory period value.

7. The apparatus of claim 6 wherein a stored value of the long-term depression/long-term potentiation refractory period value is stored in one of an analog and a digital storage circuit of the circuit element addressably configured to perform one of a neuron and a synapse function under special purpose program control.

8. The apparatus of claim 1 , the special purpose program control implemented using one of a microprocessor, a microcontroller, and a processor core.

9. The apparatus of claim 1 wherein a circuit element of the array of interconnected circuit elements addressably configured under special purpose program control comprises;
an accumulator,
a comparator, responsive to the accumulator,
a refractory period input and
a threshold input,
the accumulator accumulating charge from a plurality of selectable inputs of the circuit element addressably configured to perform a neuron function for comparison at the comparator with the threshold input, and,
if the accumulated charge is greater than the threshold and, if a refractory period has lapsed, the comparator causing a tire output of the circuit element selectively configured to perform the neuron function, the fire output monitored by the monitoring circuit.

10. The apparatus of claim 1 wherein a circuit element of the array programmably configured to perform one of a neuron and a synapse function or used as an addressable configured pass-thru circuit element or excluded from the artificial neural network of the neuromorphic network, the programmably configured neuron or synapse circuit element of the neuromorphic network comprising a monitoring circuit conforming to the addressably configured circuit element of the artificial neural network array, each monitoring circuit comprising:
a read register of the monitoring circuit responsive to a global network clock,
the global network clock selectively serving to provide clock signals to the circuit element addressably configured to perform a neuron function and to the circuit element addressably configured to perform a synapse function under special purpose program control,
the read register of the monitoring circuit having an output port to the output register of the monitoring circuit and further having a load read register signal line input for selectively actuating a transmittal of a neuron/synapse event from the addressably configured circuit element of the two-dimensional artificial neural network array depending on whether the conforming circuit element addressably represents a neuron or a synapse or used as an addressably configured pass-thru element or is excluded from the two-dimensional artificial neural network comprising the neuromorphic network, the addressably configured circuit element connected to

Assignees

• Univ Tennessee Res Found

Inventors

• Birdwell J Douglas
• Dean Mark E
• Schuman Catherine

Classifications

• G06N3/065

  Analogue means · CPC title

• G06N3/02

  Neural networks · CPC title

• G06N3/049

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

• G06N3/086

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

• G06N3/10 Primary

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