APPARATUS AND METHODS FOR EFFICIENT UPDATES IN SPIKING NEURON NETWORK

• US 20140032458A1
• Filed: 07/27/2012
• Published: 01/30/2014
• Est. Priority Date: 07/27/2012
• Status: Active Grant

• Current Assignee: Qualcomm, Inc.
• Original Assignee: Qualcomm Technologies Incorporated (Qualcomm, Inc.)
• Inventors: Sinyavskiy, Oleg, Polonichko, Vadim, Izhikevich, Eugene
• Granted Patent: US 9,256,823 B2
• Time in Patent Office: Days
• US Class Current: 706/16
• CPC Class Codes: G06N 3/049 Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs

Abstract

Efficient updates of connections in artificial neuron networks may be implemented. A framework may be used to describe the connections using a linear synaptic dynamic process, characterized by stable equilibrium. The state of neurons and synapses within the network may be updated, based on inputs and outputs to/from neurons. In some implementations, the updates may be implemented at regular time intervals. In one or more implementations, the updates may be implemented on-demand, based on the network activity (e.g., neuron output and/or input) so as to further reduce computational load associated with the synaptic updates. The connection updates may be decomposed into multiple event-dependent connection change components that may be used to describe connection plasticity change due to neuron input. Using event-dependent connection change components, connection updates may be executed on per neuron basis, as opposed to per-connection basis.

Claims

1. A computer-readable storage medium having instructions embodied thereon, the instructions being executable by a processor to perform a method for implementing an update of learning parameters of a plurality of connections of a spiking neuron, the method comprising:

operating the neuron in accordance with a process configured to be updated at time intervals;

storing a time history of one or more inputs provided to the neuron via the plurality of connections, the time history of a given input being descriptive of a time at which the given input occurs, the storing of the time history being performed within a time window comprising a plurality of the time intervals;

receiving an indication conveying whether the update is to be performed;

responsive to the indication, determining a plurality of input-dependent connection change components (IDCC), individual ones of the plurality of the IDCC components being associated with individual ones of the plurality of the time intervals, the plurality of IDCC components being based on the time of the indication and time of individual ones of the one or more inputs corresponding to individual ones of the plurality of the time intervals; and

effectuating the update by adjusting the learning parameters, the adjustment determined based on the plurality of the IDCC components.

2. The computer-readable storage medium of claim 1, wherein:

the update further comprises a determination of a plurality of eligibility traces, individual ones of the plurality of eligibility traces being associated with a given one of the plurality of connections and comprising a temporary record of occurrence of at least one of the one or more inputs on the given one of the plurality of connections; and

the determination of the plurality of eligibility traces being effectuated based on an integration of respective ones of the plurality of IDCC components.

3. A computer-implemented method of operating a plurality of data interfaces in a computerized network, the method being performed by one or more processors configured to execute computer program modules, the method comprising:

storing a time record of one or more data items capable of being provided via the plurality of data interfaces, the time record of a given data item being descriptive of a time at which the given data item occurred;

based on an indication conveying whether an update is to be performed, updating a plurality of parameters, the plurality of parameters being associated with the plurality of data interfaces, the updating comprising;

reading previous values of the plurality of parameters;

determining updated values of the plurality of parameters; and

storing the updated values of the plurality of parameters;

wherein;

the determining updated values is based on at least a portion of the time record, the portion covering a time interval prior to the indication; and

the determining updated values comprises a number of operations that is proportional to number of the one or more data items, the number of operations being independent from a number of the plurality of data interfaces.

4. The method of claim 3, further comprising:

operating individual ones of the plurality of data interfaces in accordance with an interface dynamic process, the dynamic process being characterized by a dynamic state capable of being updated at time periods, the dynamic process being characterized by a decay interval;

wherein;

the time interval is configured to cover a period of at least the decay time window; and

determine

the determining updated values further comprises obtaining, based on the indication and the time record, a plurality of interface change components, individual ones of the plurality interface change components being configured to characterize a change of individual ones of the plurality of parameters being effectuated at a time instance associated with the indication, due to the one or more data items occurring within the time interval window.

5. The method of claim 4, wherein:

the computerized network comprises spiking neuron network (SNN);

individual ones of the plurality of data interfaces comprise synaptic connections of the SNN;

individual ones of the plurality of parameters comprise a synaptic weight of respective synaptic connection, the weight;

is based at least on an occurrence of a data item of the one or more data items within the time interval; and

the decay time window comprises an exponential decay window characterized by a decrease of the weight by a factor of about 2.7 within a duration of the exponential decay window.

6. The method of claim 5, wherein:

the SNN is capable of implementing a learning process characterized by an external signal, the learning process configured to produce an outcome;

the update is capable of aiding the outcome being produced; and

the indication comprises at least a portion of the external signal.

7. The method of claim 4, wherein:

individual ones of the plurality of parameters comprise a connection delay associated with a respective one of the plurality of the data interfaces; and

the plurality interface change components is capable of providing update of the plurality of data interfaces.

8. The method of claim 4, wherein:

the time period comprises a regular period of equal duration;

individual ones of the plurality interface change components are associated with a respective time period occurring an integer number of the durations prior to the indication time instance; and

one or more interface change components are determined based on a difference between (i) the respective time period, and (ii) the indication time instance.

9. The method of claim 8, wherein individual ones of the one or more interface change components are associated with at least one data item of the one or more data items occurring within the respective time period.

10. The method of claim 4, wherein the indication is based on one or both of (i) an occurrence of a data item of one or more data items or (ii) a timer event indicative of an expiration of current update time period.

11. A neural network system configured to reduce a computational load for operating a plurality of communication interfaces of a network node, the system comprising:

one or more processors configured to execute one or more computer program modules to perform one or more operations comprising;

operate the node in accordance with a node dynamic process being capable of being updated at time intervals; and

based on an indication conveying whether an update is to be performed;

reduce the computational load by effectuating an update of a plurality of parameters associated with the plurality of communication interfaces, the update being based on one or more data items being communicated via at least one of the plurality of communication interfaces prior to the indication;

wherein the update of the plurality of parameters comprises a number of operations, the number of operations being independent from a number of the plurality of communication interfaces.

12. The system of claim 11, wherein the indication is based on an output being generated by the spiking neuron.

13. The system of claim 11, wherein individual ones of the plurality of data interfaces are capable of being operated in accordance with a linear interface dynamic process characterized by a stable equilibrium having a decay time scale associated therewith.

14. The system of claim 13, wherein:

individual ones of the plurality of parameters comprise a weight, the weight being capable of modifying a state of the node based on an occurrence of a data item of the one or more data items within the decay time scale; and

the update is based on one or more of;

(1) a decomposition of an interface state associated with the linear interface dynamic process into a plurality of stereotypical components;

(2) a determination of individual ones of the plurality of stereotypical components based on a time difference between occurrence of the indication and occurrence of the one or more data items;

or

(3) a linear combination of the plurality of stereotypical components, the linear combination being determined based on a value of individual ones of the plurality of stereotypical components corresponding to a time of the occurrence of the one or more data items, the value being discounted in accordance with the decay time scale and the time difference.

15. The system of claim 11, wherein execution of individual ones of the one or more computer program modules causes the one or more processors to store a time history conveying when the one or more data items were provided within a time window via the plurality of communication interfaces, the time window being equal or longer the decay time scale.

16. The system of claim 15, wherein:

the number of the components is selected between 2 and 40, inclusive; and

the number of the plurality of communication interfaces is selected between 41 and 100000, inclusive.

17. The system of claim 15, wherein the update of the plurality of parameters comprises determining a plurality of interface change components based on the indication and the time-history, individual ones of the plurality interface change components being configured to characterize a change of the state being effectuated at an indication time instance associated with the indication responsive to the one or more data items occurring within the decay window prior to the indication time instance.

18. The system of claim 17, wherein:

the linear interface dynamic process is capable of being periodically updated at a time interval;

the decay window comprising two or more of the time intervals;

the determining of the plurality of interface change components is effectuated via a look-up table, the look-up table comprising two or more entries associated with the two or more of the time intervals; and

individual ones of the plurality of interface change components are configured independent of the state of the node to enable determination of the look-up table prior to a first occurrence of the time interval to alleviate determination of the look-up table for individual ones of the updates.

19. The system of claim 17, wherein:

the linear interface dynamic process is capable of being periodically updated at a time interval;

the decay window comprising two or more the time intervals;

the determining of the plurality of interface change components is effectuated via a look-up table comprising two or more entries associated with the two or more of the time intervals;

individual ones of the a plurality of interface change components are configured based on the state of the node; and

the look-up table is determined for two or more of the updates.

20. The system of claim 11, wherein:

the update comprises an on-demand update; and

the indication comprises a buffer event associated with the storing of the history.

21. The system of claim 11, wherein:

the node state update characterized by a cyclic update effectuated at a regular time interval; and

the indication comprises a timer event associated with the regular time interval.

22. The system of claim 11, wherein:

the node state update comprises a cyclic update effectuated at regular time interval; and

the indication is based on the data item of the one or more data items communicated via at least one of the plurality of communication interfaces.

23. The system of claim 22, wherein the update is delayed until a next regular time interval occurring subsequent to occurrence of the data item of the one or more data items.

24. The system of claim 11, wherein:

the node dynamic process comprises a reinforcement learning process configured to produce an outcome;

the update of the plurality of parameters is capable of aiding the outcome being produced; and

the indication comprises a reinforcement spike generated based on an evaluation of node output versus the outcome.