Synaptic time multiplexing neuromorphic network that forms subsets of connections during different time slots

• US 8,977,578 B1
• Filed: 06/27/2012
• Issued: 03/10/2015
• Est. Priority Date: 06/27/2012
• Status: Expired due to Fees

• Current Assignee: HRL Laboratories LLC (The Boeing Co.)
• Original Assignee: HRL Laboratories LLC (The Boeing Co.)
• Inventors: Cruz-Albrecht, Jose M., Srinivasa, Narayan, Petre, Peter, Cho, Youngkwan, Nogin, Aleksey
• Primary Examiner(s): VINCENT, DAVID ROBERT
• Application Number:
  US13/535,114
• Time in Patent Office: 986 Days
• Field of Search: 706/12, 706/45, 706/62
• US Class Current: 706/12
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology

Abstract

A synaptic time-multiplexed (STM) neuromorphic network includes a neural fabric that includes nodes and switches to define inter-nodal connections between selected nodes of the neural fabric. The STM neuromorphic network further includes a neuromorphic controller to form subsets of a set of the inter-nodal connections representing a fully connected neural network. Each subset is formed during a different time slot of a plurality of time slots of a time multiplexing cycle of the STM neuromorphic network. In combination, the inter-nodal connection subsets implement the fully connected neural network. A method of synaptic time multiplexing a neuromorphic network includes providing the neural fabric and forming the subsets of the set of inter-nodal connections.

Claims

1. A synaptic time-multiplexed (STM) neuromorphic network comprising:

a neural fabric having nodes and switches to define inter-nodal connections between selected nodes of the neural fabric; and

a neuromorphic controller to form subsets of a set of the inter-nodal connections representing a fully connected neural network, each subset being formed during a different time slot of a plurality of time slots of a time multiplexing cycle of the STM neuromorphic network,

wherein in combination, the inter-nodal connection subsets implement the fully connected neural network.

2. The STM neuromorphic network of claim 1, wherein a node of the neural fabric comprises a neuron and a synapse, an output of the neuron being connected to an output of the node, the synapse being connected between an input of the node and an input of the neuron to provide an input signal to the neuron.

3. The STM neuromorphic network of claim 2, wherein the node comprises a plurality of synapses, each synapse of the plurality of synapses being connected between the neuron and a different one of a plurality of inputs of the node.

4. The STM neuromorphic network of claim 3, wherein the plurality of synapses comprises four synapses.

5. The STM neuromorphic network of claim 2, wherein the synapse is configured to implement spike timing dependent plasticity, the synapse having an input connected to the input of the node and an input connected to an output of the neuron.

6. The STM neuromorphic network of claim 1, wherein the nodes of the neural fabric are arranged in a two dimensional (2-D) array of nodes, the neural fabric further having a plurality of signal routing channels, the nodes being bounded on four sides by the signal routing channels, and wherein some of the switches of the neural fabric connecting between the signal routing channels and the nodes to facilitate communicating signals into and out of the nodes.

7. The STM neuromorphic network of claim 6, wherein the neural fabric associated with a node is divided into blocks comprising:

a input/output (I/O) block having an input switch to selectively connect a signal routing channel to an input of the node and an output switch to selectively connect an output of the node to the signal routing channel; and

a diagonal block having a switch to selectively connect a first signal routing channel to a substantially orthogonal second signal routing channel.

8. The STM neuromorphic network of claim 1, further comprising:

a switch state memory to store a switch state of the switches; and

a synapse conductance state memory to store a synapse conductance state of a synapse of a node of the neural fabric,

wherein the stored switch state and the stored synapse conductance state correspond to a time slot of the plurality of time slots of the time multiplexing cycle of the STM neuromorphic network.

9. The STM neuromorphic network of claim 1, further comprising a neuromorphic compiler, the neuromorphic compiler to convert a description of the fully connected neural network into an STM compatible routing schedule, the STM compatible routing schedule to assign the nodes to different ones of the time slots of the plurality and to provide connection routing between the assigned nodes during the different time slots.

10. The STM neuromorphic network of claim 9, wherein the neuromorphic compiler is further configured to provide connection routing verification by comparing a binary image of the connection routing in the different time slots to a binary image created from the description of the fully connected neural network.

11. A method of synaptic time multiplexing a neuromorphic network, the method comprising:

receiving a neural fabric of the neuromorphic network, the neural fabric comprising a plurality of neurons and associated synapses interconnected by a plurality of switches, selective switch activation defining a connection between an output of a first neuron and a synapse connected to an input of a second neuron of the neural fabric;

forming subsets of a set of connections representing a fully connected neural network within the neural fabric using the selective switch activation, each subset being formed sequentially during a different time slot of a plurality of time slots of a synaptic time multiplexing cycle of the neuromorphic network,

wherein the sequentially formed connection subsets in combination implement the fully connected neural network.

12. The method of synaptic time multiplexing of claim 11, wherein the neurons and associated synapses of the neural fabric are arranged in a two dimensional (2-D) array of nodes, each node comprising a neuron and a plurality of associated synapses, and wherein the neural fabric further comprises a plurality of signal routing channels, each node being bounded on a side by a signal routing channel of the plurality.

13. The method of synaptic time multiplexing of claim 11, wherein the neuromorphic network further comprises a switch state memory to store a switch state of the plurality of switches and a synapse conductance state memory to store a synapse conductance state of the associated synapses, the method further comprising:

retrieving a stored switch state from the switch state memory and retrieving a stored synapse conductance state from the synapse conductance state memory during a time slot of the synaptic time multiplexing cycle, the stored switch state and the stored synapse conductance state corresponding to the time slot;

using the retrieved switch state to set a switch of the plurality of switches during the time slot; and

using the retrieved synapse conductance state to set a synapse of the associated synapses during the time slot.

14. The method of synaptic time multiplexing of claim 13, wherein the synapse provides spike timing dependent plasticity, and wherein using the retrieved synapse conductance state comprises:

updating the synapse conductance state of the synapse according to a spike timing dependent plasticity rule during the time slot; and

storing the updated synapse conductance state in the synapse conductance state memory,

wherein updating and storing are performed during the time slot.

15. The method of synaptic time multiplexing of claim 11, further comprising converting a description of the fully connected neural network into a synaptic time multiplexing compatible routing schedule, the synaptic time multiplexing compatible routing schedule both to assign the neurons and the associated synapses to different ones of the time slots of the synaptic time multiplexing cycle of the neuromorphic network and to provide connection routing between the assigned neurons and associated synapse during the time slots.

16. The method of synaptic time multiplexing of claim 11, further comprising providing connection routing verification by comparing a binary image representing connection routing in the time slots to a binary image created from a connection matrix of the fully connected neural network.