Adaptive configuration of a neural network device

Patent metadata

Field	Value
Publication number	US-10331997-B2
Application number	US-201414522192-A
Country	US
Kind code	B2
Filing date	Oct 23, 2014
Priority date	May 7, 2014
Publication date	Jun 25, 2019
Grant date	Jun 25, 2019

Abstract

A first input is processed via a first configuration of a neural network to produce a first output. The first configuration defines attributes of the neural network, such as connections between neural elements of the neural network. If the neural network requires a context switch to process a second input, a second configuration is applied to the neural network to change the attributes, and the second input is processed via the second configuration of the neural network to produce a second output.

First claim

What is claimed is:

1. A method comprising:
processing a first input using a first configuration of a sparse neural network to produce a first output, the first configuration defining attributes of the sparse neural network, the attributes a plurality of connections between neural elements of the neural network, the neural elements comprising specialized neural network circuitry, each of the neural elements having a transfer function defined during a previous learning phase that converts inputs to outputs of the neural elements, each of the outputs corresponding to one of the connections and having a weighting defined during the previous learning phase that affects the corresponding connection, the attributes defining which of the inputs and the outputs are coupled by the connections, the transfer functions and the weightings the attributes further defining;
determining that the neural network requires a context switch to process a second input;
loading a second configuration from a persistent memory, the second configuration comprising changed attributes which include second connections between the neural elements, second values of the transfer functions, and second values the weightings, the changed attributes having been defined during the previous learning phase;
copying the second configuration to a configuration register of the neural network to apply the changed attributes to the neural network; and
processing the second input using the second configuration of the neural network to produce a second output.

2. The method of claim 1 , wherein the first and second configurations correspond to layers of a virtual neural network, and the second input comprises the first output.

3. The method of claim 2 , wherein the virtual neural network is too large to be represented in the neural network circuitry.

4. The method of claim 1 , wherein the first and second configurations correspond to first and second neural networks that produce parallel outputs.

5. The method of claim 4 , wherein the first and second neural networks are too large to be represented in the neural network circuitry together.

6. The method of claim 1 , wherein the persistent memory provides context configuration data out of order relative to data access requests, the method further comprising:
receiving scheduling information regarding the data access requests directed to the persistent memory; and
generate an ordering of context configuration to ensure a desired delivery order of the context configuration data.

7. The method of claim 1 , wherein the neural network comprises two or more configuration registers, and copying the second configuration to the configuration register comprises:
copying the second configuration to a second of the two or more configuration registers, wherein the first configuration is stored in a first of the two or more configuration registers; and
signaling to the neural network to use the second configuration register.

8. The method of claim 1 , further comprising pausing the processing of the first input before processing the second input.

9. An apparatus comprising:
a sparse neural network comprising specialized neural network circuitry that includes:
a plurality of neural elements each having a transfer function that converts inputs to outputs of the neural elements;
a plurality of connections that couple the inputs to the outputs of different ones of the neural elements, each of the connections and having a weighting that affects the connection; and
a configuration register that changes attributes of the sparse neural network, the attributes having been defined during a previous learning phase and including which of the inputs and the outputs are coupled by the connections, the transfer functions of the neural elements, and the weightings of the connections;

a controller coupled to the sparse neural network and operable to:
determine that the sparse neural network requires a context switch to process a second input, the sparse neural network currently processing a first input via a first configuration of the attributes of the sparse neural network;
load a second configuration from a persistent memory, the second configuration comprising changed attributes which include second connections between the neural elements, second values of the transfer functions, and second values the weightings, the changed attributes having been defined during the previous learning phase;
copy the second configuration to the configuration register of the sparse neural network to apply the changed attributes to the sparse neural network; and
process the second input via the second configuration of the sparse neural network to produce a second output.

10. The apparatus of claim 9 , wherein the first and second configurations correspond to layers of a virtual neural network.

11. The apparatus of claim 10 , wherein the virtual neural network is too large to be represented in the neural network circuitry.

12. The apparatus of claim 9 , wherein the first and second configurations correspond to first and second neural networks.

13. The apparatus of claim 12 , wherein and the first and second neural networks are too large to be represented in the neural network circuitry together.

14. The apparatus of claim 12 , wherein the persistent memory provides context configuration data out of order relative to data access requests, the controller further configured to:
receive scheduling information regarding the data access requests; and
generate an ordering of context configuration to ensure a desired delivery order.

15. The apparatus of claim 12 , wherein the neural network comprises two or more configuration registers, and copying the second configuration to the configuration register comprises:
copying the second configuration to a second of the two or more configuration registers, wherein the first configuration is stored in a first of the two or more configuration registers; and
signaling to the neural network to use the second configuration register.

16. A system comprising:
a host processor; and
at least one storage compute device comprising:
a persistent memory;
a sparse neural network comprising specialized neural network circuitry that includes:
a plurality of neural elements each having a transfer function that converts inputs to outputs of the neural elements;
a plurality of connections that couple the inputs to the outputs of different ones of the neural elements, each of the connections having a weighting that affects the connection; and
a configuration register that applies first attributes of the neural network, the first attributes having been defined during a previous learning phase and including which of the inputs and the outputs of the neural elements are coupled by the connections, the transfer functions of the neural elements, and the weightings of the connections; and

a processing unit coupled to the sparse neural network, the persistent memory, and the host processor, the processing unit configured to:
process a first input via a first configuration of the sparse neural network to produce a first output, the first configuration comprising the first attributes of the sparse neural network;
determine that the neural network requires a switch to process a second input;
load a second configuration from a persistent memory, the second configuration comprising changed attributes which include second connections between the neural elements of the sparse neural network, second values of the transfer functions, and second values the weightings, the changed attributes having bee

Assignees

• Seagate Technology Llc

Inventors

• Bohn Richard Esten
• Li Peng
• Tetzlaff David

Classifications

• G06N3/04 Primary

  Architecture, e.g. interconnection topology · CPC title

• G06N3/0495

  Quantised networks; Sparse networks; Compressed networks · CPC title