Unsupervised, supervised and reinforced learning via spiking computation

• US 10,019,669 B2
• Filed: 11/18/2015
• Issued: 07/10/2018
• Est. Priority Date: 09/16/2011
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Modha, Dharmendra S.
• Primary Examiner(s): CHANG, LI WU
• Application Number:
  US14/945,316
• Publication Number:
  US 20170017878A1
• Time in Patent Office: 965 Days
• Field of Search: None
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06N 3/049 Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
• CPC Class Codes: G06N 3/063 using electronic means
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06N 3/088 Non-supervised learning, e.g. competitive learning

Abstract

The present invention relates to unsupervised, supervised and reinforced learning via spiking computation. The neural network comprises a plurality of neural modules. Each neural module comprises multiple digital neurons such that each neuron in a neural module has a corresponding neuron in another neural module. An interconnection network comprising a plurality of edges interconnects the plurality of neural modules. Each edge interconnects a first neural module to a second neural module, and each edge comprises a weighted synaptic connection between every neuron in the first neural module and a corresponding neuron in the second neural module.

Claims

1. A method for feature extraction, comprising:

at a neural network comprising a plurality of neural modules interconnected via a plurality of weighted synaptic connections;

receiving a first input;

extracting one or more input features from the first input as the first input propagates through the neural network via at least one of the weighted synaptic connections;

training the neural network to associate the one or more input features with the first input by applying a learning rule to at least one of the weighted synaptic connections to strengthen conductance of the at least one weighted synaptic connection; and

after the training, in response to receiving a second input comprising first input with noise, reproducing the first input without noise via the trained neural network;

wherein the training comprises unlearning network-created information including false positives by copying firing events generated by a neural module positioned along a learning, top-down pathway of the neural network, and providing the copied firing events as input to another neural module positioned along an unlearning, bottom-up pathway of the neural network; and

wherein each neural module comprises an electronic circuit including multiple neurons.

2. The method of claim 1, wherein:

each neuron in each neural module is connected to a corresponding neuron in another neural module via a weighted synaptic connection of the weighted synaptic connections.

3. The method of claim 2, further comprising:

at the neural network;

adaptively updating weights of the weighted synaptic connections as a function of firing events of interconnected neurons.

4. The method of claim 1, wherein the training the neural network comprises applying at least one learning rule to the at least one weighted synaptic connection to adjust conductance of the at least one weighted synaptic connection as the first input propagates through the neural network via the at least one weighted synaptic connection.

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

at a neural network comprising a plurality of neural modules interconnected via a plurality of weighted synaptic connections;

receiving a first input;

extracting one or more input features from the first input as the first input propagates through the neural network via at least one of the weighted synaptic connections;

training the neural network to associate the one or more input features with the first input by applying a learning rule to at least one of the weighted synaptic connections to strengthen conductance of the at least one weighted synaptic connection; and

after the training, in response to receiving a second input comprising the first input with noise, reproducing the first input without noise via the trained neural network;

wherein the training comprises unlearning network-created information including false positives by copying firing events generated by a neural module positioned along a learning, top-down pathway of the neural network, and providing the copied firing events as input to another neural module positioned along an unlearning, bottom-up pathway of the neural network; and

wherein each neural module comprises an electronic circuit including multiple neurons.

6. The system of claim 5, wherein:

each neuron in each neural module is connected to a corresponding neuron in another neural module via a weighted synaptic connection.

7. The system of claim 6, the method further comprising:

at the neural network;

adaptively updating weights of the weighted synaptic connections as a function of firing events of interconnected neurons.

8. The system of claim 5, wherein the training the neural network comprises applying at least one learning rule to the at least one weighted synaptic connection to adjust conductance of the at least one weighted synaptic connection as the first input propagates through the neural network via the at least one weighted synaptic connection.

9. 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:

at a neural network comprising a plurality of neural modules interconnected via a plurality of weighted synaptic connections;

receiving a first input;

extracting one or more input features from the first input as the first input propagates through the neural network via at least one of the weighted synaptic connections;

training the neural network to associate the one or more input features with the first input by applying a learning rule to at least one of the weighted synaptic connections to strengthen conductance of the at least one weighted synaptic connection; and

after the training, in response to receiving a second input comprising the first input with noise, reproducing the first input without noise via the trained neural network;

wherein the training comprises unlearning network-created information including false positives by copying firing events generated by a neural module positioned along a learning, top-down pathway of the neural network, and providing the copied firing events as input to another neural module positioned along an unlearning, bottom-up pathway of the neural network; and

wherein each neural module comprises an electronic circuit including multiple neurons.

10. The computer program product of claim 9, wherein:

each neuron in each neural module is connected to a corresponding neuron in another neural module via a weighted synaptic connection of the weighted synaptic connections.

11. The computer program product of claim 10, the method further comprising:

at the neural network;

adaptively updating weights of the weighted synaptic connections as a function of firing events of interconnected neurons.

12. The computer program product of claim 9, wherein the training the neural network comprises applying at least one learning rule to the at least one weighted synaptic connection to adjust conductance of the at least one weighted synaptic connection as the first input propagates through the neural network via the at least one weighted synaptic connection.