Multi-modal neural network for universal, online learning

• US 10,282,661 B2
• Filed: 03/10/2017
• Issued: 05/07/2019
• Est. Priority Date: 12/14/2011
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Modha, Dharmendra S.
• Primary Examiner(s): Starks, Wilbert L
• Application Number:
  US15/456,429
• Publication Number:
  US 20170185896A1
• Time in Patent Office: 788 Days
• Field of Search: None
• CPC Class Codes: G06N 3/02 Neural networks
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06N 3/08 Learning methods

Abstract

In one embodiment, the present invention provides a neural network comprising multiple modalities. Each modality comprises multiple neurons. The neural network further comprises an interconnection lattice for cross-associating signaling between the neurons in different modalities. The interconnection lattice includes a plurality of perception neuron populations along a number of bottom-up signaling pathways, and a plurality of action neuron populations along a number of top-down signaling pathways. Each perception neuron along a bottom-up signaling pathway has a corresponding action neuron along a reciprocal top-down signaling pathway. An input neuron population configured to receive sensory input drives perception neurons along a number of bottom-up signaling pathways. A first set of perception neurons along bottom-up signaling pathways drive a first set of action neurons along top-down signaling pathways. Action neurons along a number of top-down signaling pathways drive an output neuron population configured to generate motor output.

Claims

1. A method comprising:

receiving motor output from one or more neurons along a top-down signaling pathway of a neuromorphic and synaptronic hardware architecture comprising electronic circuitry in response to the one or more neurons firing, wherein the one or more neurons correspond to one or more other neurons along a bottom-up signaling pathway of the neuromorphic and synaptronic hardware architecture;

determining whether the motor output is a first type of motor output or a second type of motor output based on firing events of a first neuron of the one or more neurons and firing events of a second neuron of the one or more other neurons, wherein the second neuron corresponds to the first neuron; and

training the one or more other neurons to approximate the one or more neurons by;

in response to determining the motor output is the first type of motor output, training the one or more other neurons to learn the motor output by routing the motor output and a first input to a third neuron and a fourth neuron of the neuromorphic and synaptronic hardware architecture, respectively; and

in response to determining the motor output is the second type of motor output, training the one or more other neurons to unlearn the motor output by routing the motor output and the first input to the fourth neuron and the third neuron, respectively;

wherein the neuromorphic and synaptronic hardware architecture resulting from the training is a single universal substrate for unsupervised learning, supervised learning, and reinforcement learning.

2. The method of claim 1, wherein:

the first input represents a zero input for preventing deadlock in active propagation;

the motor output is the first type of motor output in response to determining the first neuron does not fire but the second neuron fires; and

the motor output is the second type of motor output in response to determining the first neuron fires but the second neuron does not fire.

3. The method of claim 2, wherein, in response to determining the motor output is the first type of motor output, the motor output is routed to the third neuron to train the third neuron to learn the motor output.

4. The method of claim 3, wherein the first type of motor output represents a false negative.

5. The method of claim 1, wherein, in response to determining the motor output is the second type of motor output, the motor output is routed to the fourth neuron to train the fourth neuron to unlearn the motor output.

6. The method of claim 5, wherein the second type of motor output represents a false positive.

7. The method of claim 1, wherein the neuromorphic and synaptronic hardware architecture comprises a plurality of neurons interconnected via a first set of signaling pathways and a second set of signaling pathways, and each neuron along a signaling pathway of the first set corresponds to another neuron along a signaling pathway of the second set.

8. The method of claim 7, further comprising:

training each neuron along a signaling pathway of the first set using a first learning rule; and

training each neuron along a signaling pathway of the second set using a second learning rule different from the first learning rule.

9. A system comprising a computer processor, a computer-readable hardware storage device, and program code embodied with the computer-readable hardware storage device for execution by the computer processor to implement a method comprising:

receiving motor output from one or more neurons along a top-down signaling pathway of a neuromorphic and synaptronic hardware architecture comprising electronic circuitry in response to the one or more neurons firing, wherein the one or more neurons correspond to one or more other neurons along a bottom-up signaling pathway of the neuromorphic and synaptronic hardware architecture;

determining whether the motor output is a first type of motor output or a second type of motor output based on firing events of a first neuron of the one or more neurons and firing events of a second neuron of the one or more other neurons, wherein the second neuron corresponds to the first neuron; and

training the one or more other neurons to approximate the one or more neurons by;

in response to determining the motor output is the first type of motor output, training the one or more other neurons to learn the motor output by routing the motor output and a first input to a third neuron and a fourth neuron of the neuromorphic and synaptronic hardware architecture, respectively; and

in response to determining the motor output is the second type of motor output, training the one or more other neurons to unlearn the motor output by routing the motor output and the first input to the fourth neuron and the third neuron, respectively;

wherein the neuromorphic and synaptronic hardware architecture resulting from the training is a single universal substrate for unsupervised learning, supervised learning, and reinforcement learning.

10. The system of claim 9, wherein:

the first input represents a zero input for preventing deadlock in active propagation;

the motor output is the first type of motor output in response to determining the first neuron does not fire but the second neuron fires; and

the motor output is the second type of motor output in response to determining the first neuron fires but the second neuron does not fire.

11. The system of claim 10, in response to determining the motor output is the first type of motor output, the motor output is routed to the third neuron to train the third neuron to learn the motor output.

12. The system of claim 11, wherein the first type of motor output represents a false negative.

13. The system of claim 9, wherein, in response to determining the motor output is the second type of motor output, the motor output is routed to the fourth neuron to train the fourth neuron to unlearn the motor output.

14. The system of claim 13, wherein the second type of motor output represents a false positive.

15. The system of claim 9, wherein the neuromorphic and synaptronic hardware architecture comprises a plurality of neurons interconnected via a first set of signaling pathways and a second set of signaling pathways, and each neuron along a signaling pathway of the first set corresponds to another neuron along a signaling pathway of the second set.

16. The system of claim 15, the method further comprising:

training each neuron along a signaling pathway of the first set using a first learning rule; and

training each neuron along a signaling pathway of the second set using a second learning rule different from the first learning rule.

17. A computer program product comprising a computer-readable hardware storage device having program code embodied therewith, the program code being executable by a computer to implement a method comprising:

receiving motor output from one or more neurons along a top-down signaling pathway of a neuromorphic and synaptronic hardware architecture comprising electronic circuitry in response to the one or more neurons firing, wherein the one or more neurons correspond to one or more other neurons along a bottom-up signaling pathway of the neuromorphic and synaptronic hardware architecture;

determining whether the motor output is a first type of motor output or a second type of motor output based on firing events of a first neuron of the one or more neurons and firing events of a second neuron of the one or more other neurons, wherein the second neuron corresponds to the first neuron; and

training the one or more other neurons to approximate the one or more neurons by;

in response to determining the motor output is the first type of motor output, training the one or more other neurons to learn the motor output by routing the motor output and a first input to a third neuron and a fourth neuron of the neuromorphic and synaptronic hardware architecture, respectively; and

in response to determining the motor output is the second type of motor output, training the one or more other neurons to unlearn the motor output by routing the motor output and the first input to the fourth neuron and the third neuron, respectively;

wherein the neuromorphic and synaptronic hardware architecture resulting from the training is a single universal substrate for unsupervised learning, supervised learning, and reinforcement learning.

18. The computer program product of claim 17, wherein:

the motor output is the first type of motor output in response to determining the first neuron does not fire but the second neuron fires;

in response to determining the motor output is the first type of motor output, the motor output is routed to the third neuron to train the third neuron to learn the motor output; and

the first type of motor output represents a false negative.

19. The computer program product of claim 17, wherein:

the motor output is the second type of motor output in response to determining the first neuron fires but the second neuron does not fire;

in response to determining the motor output is the second type of motor output, the motor output is routed to the fourth neuron to train the fourth neuron to unlearn the motor output; and

the second type of motor output represents a false positive.

20. The computer program product of claim 17, wherein the neuromorphic and synaptronic hardware architecture comprises a plurality of neurons interconnected via a first set of signaling pathways and a second set of signaling pathways, each neuron along a signaling pathway of the first set corresponds to another neuron along a signaling pathway of the second set, each neuron along a signaling pathway of the first set is trained using a first learning rule, and each neuron along a signaling pathway of the second set is trained using a second learning rule different from the first learning rule.