Pre-synaptic learning using delayed causal updates

Patent metadata

Field	Value
Publication number	US-10748060-B2
Application number	US-201615294666-A
Country	US
Kind code	B2
Filing date	Oct 14, 2016
Priority date	Oct 14, 2016
Publication date	Aug 18, 2020
Grant date	Aug 18, 2020

Abstract

A processor or integrated circuit includes a memory to store weight values for a plurality neuromorphic states and a circuitry coupled to the memory. The circuitry is to detect an incoming data signal for a pre-synaptic neuromorphic state and initiate a time window for the pre-synaptic neuromorphic state in response to detecting the incoming data signal. The circuitry is further to, responsive to detecting an end of the time window: retrieve, from the memory, a weight value for a post-synaptic neuromorphic state for which an outgoing data signal is generated during the time window, the post-synaptic neuromorphic state being a fan-out connection of the pre-synaptic neuromorphic state; perform a causal update to the weight value, according to a learning function, to generate an updated weight value; and store the updated weight value back to the memory.

First claim

What is claimed is:

1. A processor comprising:
a memory to store weight values for a plurality neuromorphic states; and
a circuitry coupled to the memory, the circuitry to:
detect an incoming data signal for a pre-synaptic neuromorphic state;
initiate a time window for the pre-synaptic neuromorphic state in response to detection of the incoming data signal, the time window comprising a plurality of time steps, each time step defined by a number of clock cycles, wherein each time step of the plurality of time steps comprises an acausal update period, an outgoing data signal accumulation period following the acausal update period, and a causal update period following the outgoing data signal accumulation period;
detect generation of an outgoing data signal during the time window; and
responsive to detection of an end of the time window:
retrieve, from the memory, a weight value for a post-synaptic neuromorphic state corresponding to the outgoing data signal, the post-synaptic neuromorphic state being a fan-out connection of the pre-synaptic neuromorphic state;
perform, in the causal update period of a time step following the end of the time window, a causal update to the weight value, according to a learning function, to generate an updated weight value; and
store the updated weight value back to the memory.

2. The processor of claim 1 , wherein to perform the causal update is to increase the weight value according to the learning function.

3. The processor of claim 1 , further comprising a set of counters coupled to the circuitry, the set of counters including:
a first counter associated with the pre-synaptic neuromorphic state to track passage of the plurality of time steps through the time window; and
a second counter associated with the post-synaptic neuromorphic state, wherein the circuitry is further to detect the generation of the outgoing data signal by detection of a non-zero value of that the second counter.

4. The processor of claim 1 , wherein, responsive to detection of the end of the time window, the circuitry further to:
retrieve, from the memory, a second weight value for a second post-synaptic neuromorphic state for which a second outgoing data signal is generated during the time window, the second post-synaptic neuromorphic state also being a fan-out connection of the pre-synaptic neuromorphic state;
perform a causal update to the second weight value according to the learning function, to generate an updated second weight value; and
store the updated second weight value back to the memory.

5. The processor of claim 1 , wherein the weight values in the memory are indexed within a weight table according to pre-synaptic neuromorphic states and corresponding post-synaptic neuromorphic states.

6. The processor of claim 5 , wherein the memory is further to store a pointer table comprising a plurality of pointers, each pointer to identify a starting position of a pre-synaptic neuromorphic state within the weight table.

7. The processor of claim 1 , wherein, in response to detection of the incoming data signal for the pre-synaptic neuromorphic state, the circuitry further to:
retrieve, from the memory, a second weight value for a second post-synaptic neuromorphic state for which an outgoing data signal is generated during a preceding time window, the second post-synaptic neuromorphic state being a fan-out connection of the pre-synaptic neuromorphic state;
perform an acausal update to the second weight value according to the learning function, to generate an updated second weight value; and
store the updated second weight value back to the memory.

8. The processor of claim 1 , wherein the circuitry is further to set a refractory period to a period of at least a number of the plurality of time steps defined by the time window, the refractory period defining a minimum number of time steps between generating the outgoing data signal.

9. The processor of claim 1 , wherein, responsive to detection, during the time window, generation of a second outgoing data signal for the post-synaptic neuromorphic state, the circuitry further to ignore the generation of the second outgoing data signal.

10. An integrated circuit comprising:
a memory to store weight values for a plurality neuromorphic states; and
a circuitry coupled to the memory, the circuitry to:
detect an incoming data signal for a pre-synaptic neuromorphic state;
initiate a time window for the pre-synaptic neuromorphic state in response to detecting the incoming data signal, the time window comprising a plurality of time steps, each time step defined by a number of clock cycles, wherein each time step of the plurality of time steps comprises an acausal update period, an outgoing data signal accumulation period following the acausal update period, and a causal update period following the outgoing data signal accumulation period;
detect generation of an outgoing data signal during the time window; and
responsive to detection of an arrival of a second incoming data signal before an end of the time window:
retrieve, from the memory, a weight value for a post-synaptic neuromorphic state corresponding to the outgoing data signal, the post-synaptic neuromorphic state being a fan-out connection of the pre-synaptic neuromorphic state;
perform, during the causal update period, a causal update to the weight value according to a learning function, to generate an updated weight value;
perform, during the acausal update period, an acausal update to the updated weight value according to the learning function, to generate a twice updated weight value; and
store the twice updated weight value back to the memory.

11. The integrated circuit of claim 10 , wherein to perform the causal update is to increase the weight value according to the learning function, and to perform the acausal update is to decrease the updated weight value according to the learning function.

12. The integrated circuit of claim 10 , further comprising a set of counters coupled to the circuitry, the set of counters including:
a first counter associated with the pre-synaptic neuromorphic state to track passage of the plurality of time steps through the time window; and
a second counter associated with the post-synaptic neuromorphic state, wherein the circuitry is further to detect the generation of the outgoing data signal by detection of a non-zero value of the second counter.

13. The integrated circuit of claim 10 , responsive to detection of the arrival of the second incoming data signal before the end of the time window, the circuitry further to:
retrieve, from the memory, a second weight value for a second post-synaptic neuromorphic state for which a second outgoing data signal is generated during the time window, the second post-synaptic neuromorphic state also being a fan-out connection of the pre-synaptic neuromorphic state;
perform a causal update to the second weight value according to a learning function, to generate an updated second weight value;
perform an acausal update to the updated second weight value according to the learning function, to generate a twice updated second weight value; and
store the twice updated second weight value back to the memory.

14. The integrated circuit of claim 10 , wherein the weight values in the memory are indexed within a weight table according to pre-synaptic neuromorphic states and corresponding post-synaptic neuromorphic states.

15. The integrated circuit of claim 14 , wherein the memory is further to store a pointer table comprising a plurality of pointers, each pointer to identify a starting position of a pre-sy

Assignees

• Intel Corp

Inventors

• Paul Somnath
• Augustine Charles
• Khellah Muhammad M

Classifications

• G06N3/065 Primary

  Analogue means · CPC title

• G06N3/0495

  Quantised networks; Sparse networks; Compressed networks · CPC title

• G06N3/049

  Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs · CPC title

• G06N3/063

  using electronic means · CPC title

• G06N20/00

  Machine learning · CPC title