# US 2015/0106310 A1 · Method and apparatus for constructing a neuroscience-inspired artificial neural network

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2015/0106310 A1
- Document ID: US20150106310A1
- Status: Published application
- Kind: A1
- Application number: 14/513,280
- Filed: 2014-10-14
- Published: 2015-04-16
- Earliest claimed: 2013-10-16
- Assignees: University of Tennessee Research Foundation
- Applicants: University of Tennessee Research Foundation
- Inventors: J. Douglas Birdwell; Catherine Schuman
- Web page: https://patentgrep.com/patent/US20150106310A1

## CPC

- G06N 3/063: using electronic means
- G06N 3/02: Neural networks
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/049: Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
- G06N 3/0495: Quantised networks; Sparse networks; Compressed networks
- G06N 3/0499: Feedforward networks
- G06N 3/065: Analogue means
- G06N 3/082: modifying the architecture, e.g. adding, deleting or silencing nodes or connections
- G06N 3/086: using evolutionary algorithms, e.g. genetic algorithms or genetic programming
- G06N 3/09: Supervised learning
- G06N 3/092: Reinforcement learning
- G06N 3/0985: Hyperparameter optimisation; Meta-learning; Learning-to-learn
- G06N 3/10: Interfaces, programming languages or software development kits, e.g. for simulating neural networks
- G06F 16/22: Indexing; Data structures therefor; Storage structures
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

A method and apparatus for constructing a neuroscience-inspired dynamic architecture (NIDA) for an artificial neural network is disclosed. The method comprises constructing, in one embodiment, an artificial neural network embodiment in a multi-dimensional space in memory such that a neuron is connected by a synapse to another neuron. The neuron and the synapse each have parameters and have features of long-term potentiation and long-term depression. Furthermore, crossover and mutation are employed to select children of parents. Through learning, an initial network may evolve into a different network when NIDA is applied to solve different problems of control, anomaly detection and classification over selected time units. The apparatus comprises in one embodiment a computational neuroscience-inspired artificial neural network with at least one affective network coupled to receive input data from an environment and to output data to the environment.

## Claims

### Claim 1

A method of detecting a change in an input stream via sensing events to an artificial network for one of a control, detection and classification application using a processor apparatus, the processor apparatus comprising a special purpose processor and memory and executing a special purpose stored program for the one of the control, the detection and the classification application, the method comprising:
constructing an artificial neural network in the memory, the artificial neural network comprising at least one input neuron and at least one output neuron, each neuron having a threshold parameter;
defining at least one synapse between a pair of neurons of the artificial neural network, one neuron comprising a source neuron and the other neuron comprising a destination neuron, the synapse having a parameter that specifies a propagation delay through the synapse, the synapse adapted to permit at least two discrete events to propagate together through the synapse from the source neuron to the destination neuron; and
executing the stored program in the special purpose processor, the execution resulting in the simulation of the artificial neural network, the simulation causing the artificial neural network to produce at least one output event at an output neuron in response to said input stream.

### Claim 2 (depends on 1)

The method of claim 1 additionally comprising:
evaluating the performance of the artificial neural network by comparing at least one of (a) the events in the input stream to detect changes in the input stream, (b) the behavior of one of a physical system and of a process in response to output events to a desired behavior and (c) classification error rates to desired error rates; and
using the artificial neural network and performance evaluation to create at least one new artificial neural network.

### Claim 3 (depends on 2)

The method of claim 2 being repeated at least once to obtain a third artificial neural network such that said third artificial neural network performs better than the artificial neural network first constructed responsive to the performance evaluation.

### Claim 4 (depends on 1)

The method of claim 1, each neuron having a spatial position in the artificial neural network.

### Claim 5 (depends on 4)

The method of claim 4 wherein the spatial positions of the neurons are positions in an N-dimensional coordinate system and N is at least two.

### Claim 6 (depends on 1)

The method of claim 1 wherein the delay parameter of the synapse is determined by a distance from the source neuron to the destination neuron.

### Claim 7 (depends on 1)

The method of claim 1 wherein the application is a detection application, the input stream is a stream of events and the method further comprises detecting a change in an arrival rate of the events.

### Claim 8 (depends on 1)

The method of claim 1 wherein the at least one synapse has a second parameter specifying a weight that is applied to a destination neuron that receives an event traversing the synapse.

### Claim 9 (depends on 1)

The method of claim 1 wherein each neuron has a parameter specifying a refractory period and wherein once a neuron fires the neuron is prevented from firing again until after a time has elapsed equal to the refractory period of the neuron.

### Claim 10 (depends on 1)

The method of claim 1 wherein at least two artificial neural networks are constructed in the memory and each of the two artificial neural networks has at least one defined synapse, and the performances of the at least two artificial neural networks are evaluated and compared, said comparison causing at least one new artificial neural network to be constructed.

### Claim 11 (depends on 10)

The method of claim 10 wherein said at least one new artificial neural network is constructed by modification of one of said at least two artificial neural networks.

### Claim 12 (depends on 10)

The method of claim 10 wherein said construction of said at least two artificial networks results in a third artificial neural network stored in the memory using information obtained from the said at least two artificial neural networks.

### Claim 13 (depends on 10)

The method of claim 10 wherein said construction of said at least two artificial networks comprises identifying a useful substructure for performing a particular sub-task of one of said at least two artificial networks, forming a new artificial neural network using the previously identified useful substructure and forming at least a portion of the new artificial neural network to perform a task comprising the particular sub-task, the new artificial neural network comprising the identified useful substructure.

### Claim 14 (depends on 1)

The method of claim 1, the method further comprising:
connecting an affective system to said artificial neural network wherein the affective system is designed to alter the behavior of the artificial neural network, and
executing the stored program in the processor apparatus responsive to the connected affective system, the execution resulting in the simulation of the artificial neural network and the affective system, the simulation causing the artificial neural network to produce at least one output event at an output neuron in response to said input stream.

### Claim 15 (depends on 14)

The method of claim 14 wherein the affective system is also an artificial neural network.

### Claim 16 (depends on 14)

The method of claim 14 wherein the affective system regulates the firing rate of the artificial neural network in order to limit the power consumed by the artificial neural network.

### Claim 17 (depends on 14)

The method of claim 14 wherein the affective system regulates the firing rate of the artificial neural network in order to achieve a desirable output of the artificial neural network in response to a change in a characteristic of the input stream.

### Claim 18 (depends on 17)

The method of claim 17 wherein the input stream is a stream of events and said characteristic is the arrival rate of events.

### Claim 19 (depends on 14)

The method of claim 14, wherein the affective system has first and second different parameter values.

### Claim 20 (depends on 14)

The method of claim 14, wherein the affective system has first, second and third parameter values.

### Claim 21 (depends on 14)

The method of claim 14 further comprising connecting a second, different affective system to the artificial neural network.

### Claim 22 (depends on 1)

The method of claim 1, the method further comprising:
during constructing the artificial neural network in the memory, the artificial neural network having a variable time period between discrete events.

### Claim 23 (depends on 22)

The method of claim 22, the method further comprising
analyzing at least one causality neural pathway from a source neuron to a destination neuron in discrete time units greater in duration than real time units.

### Claim 24 (depends on 22)

The method of claim 22, the method additionally comprising
evaluating the performance of the artificial neural network by comparing times of output events to times of changes in detected events in a detection application; and
using the artificial neural network and performance evaluation to create at least one new artificial neural network.

### Claim 25 (depends on 1)

The method of claim 1 wherein the application is a control application and the input stream is a stream of events responsive to one of a position and a change in a position of a physical object, and wherein said at least one output event causes a force to be exerted on said physical object in order to cause said physical object to move in a desired manner.

### Claim 26 (depends on 1)

The method of claim 1 wherein the application is a detection application and the input stream is a stream of events wherein at least one event at a neuron corresponds to a detection of an anomaly that is indicated by a change in an arrival rate of the stream of events.

### Claim 27 (depends on 1)

The method of claim 1 wherein the application is a classification application and the input stream is representative of the intensity values of pixels of an image and wherein said at least one output event corresponds to the classification of at least a portion of said image.

### Claim 28 (depends on 1)

The method of claim 1 wherein the at least one synapse has a second parameter specifying a weight that is applied to a destination neuron that receives an event traversing the synapse.

### Claim 29 (depends on 1)

The method of claim 1 wherein the application comprises classification and the classification application comprises the processor apparatus determining that a substructure of the artificial neural network is a useful substructure for use in classification.

### Claim 30 (depends on 29)

The method of claim 29 further comprising identifying said useful substructure by analysis of at least one causality path in said artificial neural network.

### Claim 31 (depends on 30)

The method of claim 30 further comprising identifying said useful substructure as a substructure of an artificial neural network for identifying a portion of an image.

### Claim 32 (depends on 29)

The method of claim 29 wherein the classification application comprises the processor generating an artificial neural network from a plurality of determined useful substructures.

### Claim 33 (depends on 32)

The method of claim 32 wherein the plurality of determined useful substructures comprise different useful substructures, each such substructure far performing a subtask of a classification application task.

### Claim 34 (depends on 32)

The method of claim 32 wherein the application further comprises generating an artificial neural network from a plurality of useful substructures.

### Claim 35 (depends on 1)

The method of claim 1, wherein the application comprises a classification application and the classification application comprises one of the recognition of at least a portion of a hand-written image, the recognition of at least one line in space, the diagnosis of a disease and the recognition of a biological item.

### Claim 36 (depends on 1)

The method of claim 1 wherein the application comprises a classification application and the method further comprises using a voting scheme to determine a classification.

### Claim 37 (depends on 36)

The method of claim 36 wherein the voting scheme is used to classify one of a portion of an image, a line in space, a diagnosis of a disease and the recognition of a biological item.

### Claim 38 (depends on 1)

The method of claim 1 wherein the construction of an artificial neural network comprises determining a causality path in a different artificial neural network from a source neuron to a destination neuron to detect a useful substructure and constructing the artificial neural network that contains the useful substructure.

### Claim 39 (depends on 38)

The method of claim 38 wherein the useful substructure is stored in a library of a plurality of different useful substructures, two different substructures for recognizing at least two different classes of events and the constructed artificial neural network comprising the two different substructures for recognizing the two classes of events.

### Claim 40 (depends on 1)

The method of claim 1 further comprising a learning method employing one of a tournament and a roulette method to select a child population of networks of a parent population of networks.

### Claim 41 (depends on 40)

The method of claim 40, the learning method comprising selecting first and second parent networks.

### Claim 42 (depends on 41)

The method of claim 41, the learning method comprising selecting a child population of networks responsive to one of crossover and mutation responsive to selecting of first and second parent networks.

### Claim 43 (depends on 1)

The method of claim 1, wherein the application comprises one of a classification application and a detection operation and wherein said execution operates in real-time with respect to one of a physical system and an external process.

### Claim 44

Apparatus adapted to detect a change in an input stream via sensing events to an artificial network for one of a control, detection and classification application, the apparatus comprising special purpose processor apparatus, the apparatus further comprising a memory, the special purpose processor for executing a special purpose stored program for the one of the control, the detection and the classification application, the apparatus further comprising:
the special purpose processor constructing an artificial neural network in the memory, the artificial neural network comprising at least one input neuron and at least one output neuron, each neuron having a threshold parameter;
the special purpose processor defining at least one synapse between a pair of neurons of the artificial neural network, one neuron comprising a source neuron and the other neuron comprising a destination neuron, the synapse having a parameter that specifies a propagation delay through the synapse, the synapse adapted to permit at least two discrete events to propagate together through the synapse from the source neuron to the destination neuron; and
the special purpose processor executing the special purpose stored program, the execution resulting in simulation of the artificial neural network, the simulation causing the artificial neural network to produce at least one output event at an output neuron in response to said input stream.

### Claim 45 (depends on 44)

The apparatus of claim 44 further comprising:
the special purpose processor evaluating the performance of the artificial neural network by comparing at least one of (a) the events in the input stream to detect changes in the input stream, (b) the behavior of one of a physical system and of a process in response to output events to a desired behavior and (c) classification error rates to desired error rates; and
the special purpose processor using the artificial neural network and performance evaluation to create at least one new artificial neural network.

### Claim 46 (depends on 45)

The apparatus of claim 45 wherein the special purpose processor repeats the performance evaluation at least once to obtain a third artificial neural network such that said third artificial neural network performs better than the artificial neural network first constructed responsive to the performance evaluation.

### Claim 47 (depends on 44)

The apparatus of claim 44, each neuron having a spatial position in the artificial neural network.

### Claim 48 (depends on 47)

The apparatus of claim 47 wherein the spatial positions of the neurons are positions in an N-dimensional coordinate system and N is at least two.

### Claim 49 (depends on 44)

The apparatus of claim 44 wherein the delay parameter of the synapse is determined by a distance from the source neuron to the destination neuron.

### Claim 50 (depends on 44)

The apparatus of claim 44 wherein the application is a detection application, the input stream being a stream of events and the special purpose processor detects a change in an arrival rate of the events at the input neuron.

### Claim 51 (depends on 44)

The apparatus of claim 44 wherein the at least one synapse has a second parameter specifying a weight that is applied to a destination neuron that receives an event traversing the synapse.

### Claim 52 (depends on 44)

The apparatus of claim 44 wherein each neuron has a parameter specifying a refractory period and wherein once a neuron fires the neuron is prevented from firing again until after a time has elapsed equal to the refractory period of the neuron.

### Claim 53 (depends on 44)

The apparatus of claim 44, the special purpose processor for constructing at least two artificial neural networks in the memory, each of the two artificial neural networks having at least one defined synapse, and the special purpose processor comparing and evaluating the performances of the at least two artificial neural networks, said comparison and evaluation causing the processor to construct at least one new artificial neural network.

### Claim 54 (depends on 53)

The apparatus of claim 53 wherein said at least one new artificial neural network is constructed by the special purpose processor modifying one of said at least two artificial neural networks.

### Claim 55 (depends on 53)

The apparatus of claim 53 wherein said construction of said at least two artificial networks results in a third artificial neural network stored in the memory using information obtained from the said at least two artificial neural networks.

### Claim 56 (depends on 53)

The apparatus of claim 53 wherein said construction of said at least two artificial networks comprises the special purpose processor identifying a useful substructure for performing a particular sub-task of one of said at least two artificial networks, forming a new artificial neural network using the previously identified useful substructure and forming at least a portion of the new artificial neural network to perform a task comprising the particular sub-task, the new artificial neural network comprising the identified useful substructure.

### Claim 57 (depends on 44)

The apparatus of claim 44, the special purpose processor further connecting an affective system to said artificial neural network wherein the affective system is designed to alter the behavior of the artificial neural network, and executing the stored program responsive to the connected affective system, the execution resulting in the simulation of the artificial neural network and the affective system, the simulation causing the artificial neural network to produce at least one output event at an output neuron in response to said input stream.

### Claim 58 (depends on 57)

The apparatus of claim 57 wherein the connected affective system is also an artificial neural network.

### Claim 59 (depends on 57)

The apparatus of claim 57 wherein the connected affective system regulates the firing rate of the artificial neural network in order to limit the power consumed by the artificial neural network.

### Claim 60 (depends on 57)

The apparatus of claim 57 wherein the connected affective system regulates the firing rate of the artificial neural network in order to achieve a desirable output of the artificial neural network in response to a change in a characteristic of the input stream.

### Claim 61 (depends on 60)

The apparatus of claim 60 wherein the input stream is a stream of events and said characteristic is the arrival rate of events.

### Claim 62 (depends on 57)

The apparatus of claim 57, wherein the connected affective system has first and second different parameter values.

### Claim 63 (depends on 57)

The apparatus of claim 57, wherein the connected affective system has first, second and third parameter values.

### Claim 64 (depends on 57)

The apparatus of claim 57, the processor further connecting a second, different affective system to the artificial neural network.

### Claim 65 (depends on 44)

The apparatus of claim 44, the processor, during construction of the artificial neural network in the memory, the artificial neural network having a variable time period between discrete events.

### Claim 66 (depends on 65)

The apparatus of claim 65, the processor for further analyzing at least one causality neural pathway from a source neuron to a destination neuron in discrete time units greater in duration than real time units.

### Claim 67 (depends on 65)

The apparatus of claim 65, the processor for evaluating the performance of the artificial neural network by comparing times of output events to times of changes in detected events in a detection application and using the artificial neural network and performance evaluation to create at least one new artificial neural network.

### Claim 68 (depends on 44)

The apparatus of claim 44 wherein the application is a control application and the input stream is a stream of events responsive to one of a position and a change in a position of a physical object and wherein said at least one output event causes a force to be exerted on said physical object in order to cause said physical object to move in a desired manner.

### Claim 69 (depends on 44)

The apparatus of claim 44 wherein the application is a detection application and the input stream is a stream of events wherein at least one event at a neuron corresponds to a detection of an anomaly that is indicated by a change in an arrival rate of the stream of events.

### Claim 70 (depends on 44)

The apparatus of claim 44 wherein the application is a classification application and the input stream is representative of the intensity values of pixels of an image and wherein said at least one output event corresponds to the classification of at least a portion of said image.

### Claim 71 (depends on 44)

The apparatus of claim 44 wherein the at least one synapse has a second parameter specifying a weight that is applied to a destination neuron that receives an event traversing the synapse.

### Claim 72 (depends on 44)

The apparatus of claim 44 wherein the application comprises classification and the classification application comprises the processor apparatus determining that a substructure of the artificial neural network is a useful substructure for use in classification.

### Claim 73 (depends on 72)

The apparatus of claim 72 further comprising identifying said useful substructure by analysis of at least one causality path in said artificial neural network for use in classification.

### Claim 74 (depends on 73)

The apparatus of claim 73 further comprising identifying said useful substructure as a substructure of an artificial neural network for identifying a portion of an image.

### Claim 75 (depends on 72)

The apparatus of claim 72 wherein the classification application comprises the processor generating an artificial neural network from a plurality of determined useful substructures.

### Claim 76 (depends on 75)

The apparatus of claim 75 wherein the plurality of determined useful substructures comprise different useful substructures, each such substructure for performing a subtask of a classification application.

### Claim 77 (depends on 75)

The apparatus of claim 75 wherein the application further comprises generating an artificial neural network from a plurality of useful substructures.

### Claim 78 (depends on 44)

The apparatus of claim 44, wherein the application comprises a classification application and the classification application comprises one of the recognition of at least a portion of a hand-written image, the recognition of at least one line in space, the diagnosis of a disease and the recognition of a biological item.

### Claim 79 (depends on 44)

The apparatus of claim 44 wherein the application comprises a classification application and the method further comprises using a voting scheme to determine a classification.

### Claim 80 (depends on 79)

The apparatus of claim 79 wherein the voting scheme is used to classify one of a portion of an image, a line in space, a diagnosis of a disease and the recognition of a biological item.

### Claim 81 (depends on 44)

The apparatus of claim 44 wherein the construction of an artificial neural network comprises determining a causality path in a different artificial neural network from a source neuron to a destination neuron to detect a useful substructure and constructing the artificial neural network that contains at least one said useful substructure.

### Claim 82 (depends on 81)

The apparatus of claim 81 wherein the useful substructure is stored in a library of a plurality of different substructures, two different substructures for recognizing at least two different classes of events and the constructed artificial neural network comprises the two different substructures for recognizing the at least two different classes of events.

### Claim 83 (depends on 44)

The apparatus of claim 44 further comprising a learning method employing one of a tournament and a roulette method to select a child network population.

### Claim 84 (depends on 83)

The apparatus of claim 83, the learning method comprising selecting first and second parent networks from a plurality of parent networks fitting s fitness function.

### Claim 85 (depends on 84)

The apparatus of claim 84, the learning method comprising selecting a child network population responsive to one of crossover and mutation responsive to the selecting of first and second parent networks.

### Claim 86 (depends on 44)

The apparatus of claim 44, wherein the application comprises one of a classification application and a detection operation and wherein said execution operates in real-time with respect to one of a physical system and an external process.
