System for address-event-representation network simulation

• US 8,429,107 B2
• Filed: 11/04/2009
• Issued: 04/23/2013
• Est. Priority Date: 11/04/2009
• Status: Active Grant

• Current Assignee: Samsung Electronics Co. Ltd.
• Original Assignee: International Business Machines Corporation
• Inventors: Denneau, Monty M., Friedman, Daniel J., Linsker, Ralph, Ritter, Mark B.
• Primary Examiner(s): Gaffin, Jeffrey A
• Assistant Examiner(s): Traktovenko, Ilya
• Application Number:
  US12/611,968
• Publication Number:
  US 20110106741A1
• Time in Patent Office: 1,266 Days
• Field of Search: 706/25, 706/15, 706/62
• US Class Current: 706/25
• CPC Class Codes: H04L 47/521 Static queue service slot or fixed bandwidth allocation

Abstract

A system, method, and design structure for address-event-representation network simulation are provided. The system includes a hardware structure with a plurality of interconnected processing modules configured to simulate a plurality of interconnected nodes. To simulate each node, the hardware structure includes a source table configured to receive an input message and identify a weight associated with a source of the input message. The hardware structure also includes state management logic configured to update a node state as a function of the identified weight, and generate an output signal responsive to the updated node state. The hardware structure further includes a target table configured to generate an output message in response to the output signal, identify a target to receive the output message, and transmit the output message. The hardware structure may further include learning logic configured to combine information about input messages and generated output signals, and to update weights.

Claims

1. A system for address-event-representation network simulation comprising:

a hardware structure comprised of a plurality of interconnected processing modules configured to simulate a plurality of interconnected nodes, wherein each node comprises;

a source table configured to receive an input message and identify a weight associated with a source of the input message;

state management logic configured to update a node state as a function of the identified weight, and generate an output signal responsive to the updated node state; and

a target table configured to generate an output message in response to the output signal, identify a target to receive the output message, and transmit the output message.

2. The system of claim 1 wherein the address-event-representation network simulation simulates a spiking neural network, the nodes represent neurons, the input message comprises a spike from a source neuron and the output message comprises a spike emitted to a target neuron.

3. The system of claim 1 further comprising:

an output scheduler configured to receive the output signal from the state management logic, read the target table to determine a delay time, and delay the output signal for the determined delay time before sending the output signal to the target table to trigger transmission of the output message.

4. The system of claim 3 wherein the output scheduler is a shift register configured to place an indicator of the output signal from the state management logic in the shift register and shift the indicator of the output signal at time steps of the address-event-representation network simulation, and further wherein the shift register is configured to delay the output signal until the indicator of the output signal reaches a location of the shift register corresponding to the determined delay time.

5. The system of claim 1 wherein the input message comprises an entry indicator that denotes an entry of the source table to which the input message corresponds, and the output message comprises a pointer to an entry in the source table of a target node and an address identifying the target node that is to receive the output message.

6. The system of claim 1 further comprising:

learning logic configured to update the weight associated with the source of the input message responsive to a time interval between receiving a pre-output signal and a post-output signal as limited by a time limit, wherein the pre-output signal is associated with the input message and the post-output signal is associated with the output signal.

7. The system of claim 1 further comprising:

learning logic configured to monitor an order in which a pre-output signal and a post-output signal arrive within a learning time window, wherein the pre-output signal is associated with the input message and the post-output signal is associated with the output signal.

8. The system of claim 1 wherein the input message comprises one or more control bits indicating one of a simulation event and an expire event, and the output message comprises one or more control bits indicating one of a simulation event and an expire event.

9. The system of claim 8 further comprising:

a first and second shift register configured to schedule sending the simulation event and the expire event in separate output messages to the target.

10. The system of claim 1 wherein the weight associated with the source of the input message is used to generate a probability value, and the probability value determines whether to update the node state.

11. A method for address-event-representation network simulation, comprising:

receiving an input message at a source table of a node in a hardware structure comprised of a plurality of interconnected processing modules configured to simulate a plurality of interconnected nodes;

identifying a weight associated with a source of the input message, the identified weight located in the source table;

updating a node state as a function of the identified weight, the updating performed by state management logic;

generating an output signal responsive to the updated node state;

accessing a target table to identify a target to receive an output message; and

generating the output message in response to the output signal.

12. The method of claim 11 wherein the address-event-representation network simulation simulates a spiking neural network, the nodes represent neurons, the input message represents a spike from a source neuron and the output message represents a spike emitted to a target neuron.

13. The method of claim 11 further comprising:

receiving the output signal from the state management logic at an output scheduler;

reading the target table to determine a delay time; and

delaying the output signal for the determined delay time before sending the output signal to the target table to trigger transmission of the output message.

14. The method of claim 13 wherein the output scheduler is a shift register configured to place an indicator of the output signal from the state management logic in the shift register and shift the indicator of the output signal at time steps of the address-event-representation network simulation, and further wherein the shift register delays the output signal until the indicator of the output signal reaches a location of the shift register corresponding to the determined delay time.

15. The method of claim 11 wherein the input message comprises an entry indicator that denotes an entry of the source table to which the input message corresponds, and the output message comprises a pointer to an entry in the source table of a target node and an address identifying the target node that is to receive the output message.

16. The method of claim 11 further comprising:

receiving a pre-output signal associated with the input message;

receiving a post-output signal associated with the output signal; and

updating the weight associated with the source of the input message responsive to a time interval between receiving the pre-output signal and the post-output signal as limited by a time limit.

17. The method of claim 11 further comprising:

receiving a pre-output signal associated with the input message;

receiving a post-output signal associated with the output signal; and

monitoring an order in which the pre-output signal and the post-output signal arrive within a learning time window.

18. The method of claim 11 wherein the input message comprises one or more control bits indicating one of a simulation event and an expire event, and the output message comprises one or more control bits indicating one of a simulation event and an expire event.

19. The method of claim 18 further comprising:

scheduling sending the simulation event and the expire event in separate output messages to the target using a first and second shift register.

20. The method of claim 11 further comprising:

generating a probability value from the weight associated with the source of the input message; and

determining whether to update the node state in response to the probability value.

21. A design structure tangibly embodied in a non-transitory machine-readable data storage medium for designing, manufacturing, or testing an integrated circuit, the design structure comprising:

a hardware structure comprised of a plurality of interconnected processing modules configured to simulate a plurality of interconnected nodes, wherein each node comprises;

a source table configured to receive an input message and identify a weight associated with a source of the input message;

state management logic configured to update a node state as a function of the identified weight, and generate an output signal responsive to the updated node state; and

a target table configured to generate an output message in response to the output signal, identify a target to receive the output message, and transmit the output message.

22. The design structure of claim 21, wherein the design structure comprises a netlist.

23. The design structure of claim 21, wherein the design structure resides on storage medium as a data format used for the exchange of layout data of integrated circuits.

24. The design structure of claim 21, wherein the design structure resides in programmable logic.