Routing spike messages in spiking neural networks
AssigneeMicron Technology, Inc.
InventorDmitri Yudanov
Life of this patent
Estimating the term…
Abstract
The present disclosure is directed to routing of data in a spiking neural network (SNN) that performs in-memory operations. To model a computer-implemented SNN after a biological neural network, the architecture in the present disclosure involves different memory sections for storing inbound spike messages, synaptic connection data, and synaptic connection parameters. Embodiments are directed to routing spike messages through various router-based topologies. For example, spike messages may be multicasted to target routers using address tables.
Claims17
A computing system comprising:
- a first router that implements a Spiking Neural Network (SNN), the first router configured to store a plurality of neurons and synaptic connections;
- a first port of the first router coupled to a port translator;
- a second port of the first router coupled to a plurality of other routers, the plurality of other routers implementing the SNN;
- and
- memory configured to store an address table, the address table configured to route a first spike message generated by the first router to a subset of the plurality of other routers via the port translator, the address table is configured to be updated in response to a sync message;
- wherein the first router is configured to periodically transmit a sync message to the plurality of other routers;
- wherein the address table is configured to be updated in response to a sync message received from a second router among the plurality of other routers.
A system comprising:
- a first router that implements a Spiking Neural Network (SNN), the first router configured to store a plurality of neurons and synaptic connections, the plurality of neurons configured to generate spike messages targeting a plurality of other routers within the SNN;
- memory configured to store an address table, the address table configured to route a first spike message generated by a first neuron among the plurality of neurons to a subset of routers among the plurality of other routers via a port translator, the address table is configured to be updated in response to a sync message;
- and
- wherein the first router is configured to receive a second spike message from the port translator and forward the second spike message to a second router among the plurality of other routers;
- wherein the first router is configured to periodically transmit a sync message to the plurality of other routers.
A method for routing spike messages in a computer-implemented Spiking Neuron Network (SNN), the method comprising:
- generating, by a first neuron of a first router, a first spike message, the SNN comprising a set of neurons stored among a plurality of other routers that are syntactically connected to the first neuron,
- routing the first spike message to a subset of the plurality of other routers by accessing an address table, the address table being configured to map an identifier of the first neuron to at least one router identifier that identifies the subset of the plurality of other routers, the address table is configured to be updated in response to a sync message;
- and
- updating the address table in response to receiving a sync message from a second router among the plurality of other routers.
Dependencies are inferred from the claim text.