Neural architecture search

Patent metadata

Field	Value
Publication number	US-10984319-B2
Application number	US-202016859781-A
Country	US
Kind code	B2
Filing date	Apr 27, 2020
Priority date	Oct 27, 2017
Publication date	Apr 20, 2021
Grant date	Apr 20, 2021

Abstract

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for determining neural network architectures. One of the methods includes generating, using a controller neural network, a batch of output sequences, each output sequence in the batch specifying a respective subset of a plurality of components of a large neural network that should be active during the processing of inputs by the large neural network; for each output sequence in the batch: determining a performance metric of the large neural network on the particular neural network task (i) in accordance with current values of the large network parameters and (ii) with only the subset of components specified by the output sequences active; and using the performance metrics for the output sequences in the batch to adjust the current values of the controller parameters of the controller neural network.

First claim

What is claimed is:

1. A method of determining an architecture for a neural network for performing a particular neural network task, the method comprising:
generating, using a controller neural network having a plurality of controller parameters and in accordance with current values of the controller parameters, a batch of output sequences, each output sequence in the batch specifying a respective subset of a plurality of components of a large neural network, wherein the large neural network has a plurality of large network parameters;
for each output sequence in the batch:
determining a performance metric of the large neural network on the particular neural network task (i) in accordance with current values of the large network parameters and (ii) with only the subset of the plurality of components, being active; and

using the performance metrics for the output sequences in the batch to adjust the current values of the controller parameters of the controller neural network.

2. The method of claim 1 , further comprising:
generating, using the controller neural network and in accordance with the adjusted values of the controller parameters, a new output sequence; and
training the large neural network with only the subset of the plurality of components specified by the new output sequence active on training data to determine adjusted values of the large network parameters.

3. The method of claim 1 , wherein using the performance metrics for the output sequences in the batch to adjust the current values of the controller parameters of the controller neural network comprises:
training the controller neural network to generate output sequences that have increased performance metrics using a reinforcement learning technique.

4. The method of claim 3 , wherein the reinforcement learning technique is a policy gradient technique.

5. The method of claim 4 , wherein the reinforcement learning technique is a REINFORCE technique.

6. The method of claim 1 , wherein the large neural network comprises a plurality of layers.

7. The method claim 1 , wherein the current values of the large network parameters are fixed while determining the performance of the large neural network.

8. The method of claim 1 , wherein the controller neural network is a recurrent neural network that is configured to, for a given output sequence and at each of a plurality of time steps:
receive as input an embedding of an output generated at a preceding time step in the given output sequence, and to process the input to update a current hidden state of the recurrent neural network and to generate a score distribution over possible outputs for the time step.

9. The method of claim 1 , wherein each output sequence comprises respective outputs at each of a plurality of time steps, wherein each time step corresponds to a respective node in a directed acyclic graph (DAG) that represents the large neural network, wherein the DAG comprises a plurality of edges connecting nodes in the DAG, and wherein the output sequence defines, for each node, an input received by the node and a computation performed by the node.

10. The method of claim 9 , wherein, for each particular node of a plurality of nodes in the DAG, the controller neural network generates, at a first time step corresponding to the node, a probability distribution over nodes that are connected to the particular node by an incoming edge in the DAG.

11. The method of claim 9 , wherein, for each particular node of a plurality of nodes in the DAG, the controller neural network generates, at a first time step corresponding to the node, a respective independent probability for each node that is connected to the particular node by an incoming edge in the DAG that defines a likelihood that an edge will be designated as active.

12. The method of claim 9 , for each particular node of a plurality of nodes in the DAG, the controller neural network generates, at a second time step corresponding to the particular node, a probability distribution over possible computations performed by the particular node.

13. The method of claim 1 , wherein the large neural network is a recurrent neural network.

14. The method of claim 1 , wherein the large neural network is a convolutional neural network.

15. The method of claim 1 , further comprising:
generating, in accordance with the adjusted values of the controller parameters, a final output sequence that defines a final subset of components.

16. The method of claim 15 ,
performing the particular neural network task for received network inputs by processing the received network inputs with only the final subset of components active.

17. A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform operations for determining an architecture for a neural network for performing a particular neural network task, the operations comprising:
generating, using a controller neural network having a plurality of controller parameters and in accordance with current values of the controller parameters, a batch of output sequences, each output sequence in the batch specifying a respective subset of a plurality of components of a large neural network, wherein the large neural network has a plurality of large network parameters;
for each output sequence in the batch:
determining a performance metric of the large neural network on the particular neural network task (i) in accordance with current values of the large network parameters and (ii) with only the subset of the plurality of components, specified by the output sequence being active; and

using the performance metrics for the output sequences in the batch to adjust the current values of the controller parameters of the controller neural network.

18. One or more non-transitory computer-readable storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations for determining an architecture for a neural network for performing a particular neural network task, the operations comprising:
generating, using a controller neural network having a plurality of controller parameters and in accordance with current values of the controller parameters, a batch of output sequences, each output sequence in the batch specifying a respective subset of a plurality of components of a large neural network, wherein the large neural network has a plurality of large network parameters;
for each output sequence in the batch:
determining a performance metric of the large neural network on the particular neural network task (i) in accordance with current values of the large network parameters and (ii) with only the subset of the plurality of components, specified by the output sequence being active; and

using the performance metrics for the output sequences in the batch to adjust the current values of the controller parameters of the controller neural network.

19. The system of claim 17 , the operations further comprising:
generating, using the controller neural network and in accordance with the adjusted values of the controller parameters, a new output sequence; and
training the large neural network with only the subset of the plurality of components specified by the new output sequence active on training data to determine adjusted values of the large network parameters.

20. The system of claim 17 , wherein using the performance metrics for the output seq

Assignees

• Google Llc

Inventors

• Zoph Barret
• Guan Yun Jia
• Pham Hieu Hy
• Le Quoc V

Classifications

• G06N3/047

  Probabilistic or stochastic networks · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/044 Primary

  Recurrent networks, e.g. Hopfield networks · CPC title

• G06N3/092

  Reinforcement learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title