Time-division multiplexed neurosynaptic module with implicit memory addressing for implementing a neural network

• US 9,239,984 B2
• Filed: 12/21/2012
• Issued: 01/19/2016
• Est. Priority Date: 12/21/2012
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Arthur, John V., Brezzo, Bernard V., Chang, Leland, Friedman, Daniel J., Merolla, Paul A., Modha, Dharmendra S., Montoye, Robert K., Seo, Jae-sun, Tierno, Jose A.
• Primary Examiner(s): Gaffin, Jeffrey A
• Assistant Examiner(s): Traktovenko, Ilya
• Application Number:
  US13/725,476
• Publication Number:
  US 20140180987A1
• Time in Patent Office: 1,124 Days
• Field of Search: None
• US Class Current: 1/1
• 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

Abstract

Embodiments of the invention relate to a time-division multiplexed neurosynaptic module with implicit memory addressing for implementing a neural network. One embodiment comprises maintaining neuron attributes for multiple neurons and maintaining incoming firing events for different time steps. For each time step, incoming firing events for said time step are integrated in a time-division multiplexing manner. Incoming firing events are integrated based on the neuron attributes maintained. For each time step, the neuron attributes maintained are updated in parallel based on the integrated incoming firing events for said time step.

Claims

1. A method, comprising:

maintaining neuron attributes for multiple neurons;

maintaining a collection of axon activity bit maps for multiple axons, wherein each axon activity bit map corresponds to a particular time step, and the axon activity bit map maintains bits identifying incoming firing events for delivery to the multiple axons in the particular time step; and

for each time step;

integrating incoming firing events identified in an axon activity bit map corresponding to the time step in a time-division multiplexing manner, wherein the incoming firing events are integrated based on the neuron attributes;

based on the integrated incoming firing events, generating an update vector identifying outgoing firing events generated by the multiple neurons, wherein at least one of the outgoing firing events targets an axon of the multiple axons in a different time step, and an axon activity bit map corresponding to the different time step is updated to include information identifying the at least one outgoing firing event for delivery to the axon in the different time step; and

updating the neuron attributes based on the integrated incoming firing events.

2. The method of claim 1, further comprising:

maintaining synaptic connectivity information for the multiple neurons, wherein each incoming firing event is integrated based on the synaptic connectivity information.

3. The method of claim 2, further comprising:

dividing each time step into multiple time slots, wherein each time slot corresponds to an axon of the multiple axons.

4. The method of claim 3, further comprising:

for each time slot, integrating incoming firing events targeting a corresponding axon.

5. The method of claim 4, wherein maintaining neuron attributes for multiple neurons comprises:

for each neuron, maintaining a membrane potential variable for the neuron and a threshold parameter for the neuron.

6. The method of claim 5, further comprising:

maintaining routing information for the multiple neurons.

7. The method of claim 6, further comprising:

for each neuron;

determining whether a membrane potential variable for the neuron exceeds a threshold parameter for the neuron; and

generating an outgoing firing event in response to determining the membrane potential variable for the neuron exceeds the threshold parameter for the neuron.

8. The method of claim 7, further comprising:

routing each outgoing firing event based on the routing information.

9. A neurosynaptic device, comprising:

a memory device that maintains neuron attributes for multiple neurons;

a collection of axon activity bit maps for multiple axons, wherein each axon activity bit map corresponds to a particular time step, and the axon activity bit map maintains bits identifying incoming firing events for delivery to the multiple axons in the particular time step;

a scheduler that manages incoming firing events for different time steps; and

a multi-way processor configured to;

for each time step;

integrate incoming firing events identified in an axon activity bit map corresponding to the time step in a time-division multiplexing manner, wherein the incoming firing events are integrated based on the neuron attributes;

based on the integrated incoming firing events, generate an update vector identifying outgoing firing events generated by the multiple neurons, wherein at least one of the outgoing firing events targets an axon of the multiple axons in a different time step, and an axon activity bit map corresponding to the different time step is updated to include information identifying the at least one outgoing firing event for delivery to the axon in the different time step; and

update the neuron attributes based on the integrated incoming firing events.

10. The neurosynaptic device of claim 9, wherein:

the multi-way processor is further configured to maintain synaptic connectivity information for the multiple neurons, wherein each incoming firing event is integrated based on the synaptic connectivity information.

11. The neurosynaptic device of claim 10, wherein:

each time step is divided into multiple time slots; and

each time slot corresponds to an axon of the multiple axons.

12. The neurosynaptic device of claim 11, wherein:

for each neuron, the neuron attributes include a membrane potential variable for the neuron and a threshold parameter for the neuron.

13. The neurosynaptic device of claim 12, wherein:

routing information for the multiple neurons is maintained in a routing lookup table.

14. The neurosynaptic device of claim 13, wherein:

the multi-way processor is further configured to;

for each neuron;

determine whether a membrane potential variable for the neuron exceeds a threshold parameter for the neuron; and

generate an outgoing firing event in response to determining the membrane potential variable for the neuron exceeds the threshold parameter for the neuron.

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

maintaining neuron attributes for multiple neurons;

maintaining a collection of axon activity bit maps for multiple axons, wherein each axon activity bit map corresponds to a particular time step, and the axon activity bit map maintains bits identifying incoming firing events for delivery to sthe multiple axons in sthe particular time step; and

for each time step;

integrating incoming firing events identified in an axon activity bit map corresponding to the time step in a time-division multiplexing manner, wherein the incoming firing events are integrated based on the neuron attributes;

based on the integrated incoming firing events, generating an update vector identifying outgoing firing events generated by the multiple neurons, wherein at least one of the outgoing firing events targets an axon of the multiple axons in a different time step, and an axon activity bit map corresponding to the different time step is updated to include information identifying the at least one outgoing firing event for delivery to the axon in the different time step; and

updating the neuron attributes neurons based on the integrated incoming firing events.

16. The computer program product of claim 15, the method further comprising:

maintaining synaptic connectivity information for the multiple neurons, wherein each incoming firing event is integrated based on the synaptic connectivity information.

17. The computer program product of claim 16, the method further comprising:

dividing each time step into multiple time slots, wherein each time slot corresponds to an axon of the multiple axons.

18. The computer program product of claim 17, the method further comprising:

for each time slot, integrating incoming firing events targeting a corresponding axon.

19. The computer program product of claim 18, the method further comprising:

for each neuron, maintaining a membrane potential variable for the neuron and a threshold parameter for the neuron; and

maintaining routing information for the multiple neurons.

20. The computer program product of claim 19, the method further comprising:

for each neuron;

determining whether a membrane potential variable for the neuron exceeds a threshold parameter for the neuron; and

generating an outgoing firing event in response to determining the membrane potential variable for the neuron exceeds the threshold parameter for the neuron.