Interface neural network

Patent metadata

Field	Value
Publication number	US-2017300813-A1
Application number	US-201615247160-A
Country	US
Kind code	A1
Filing date	Aug 25, 2016
Priority date	Apr 15, 2016
Publication date	Oct 19, 2017
Grant date	—

Abstract

An operation method of a neural network, a training method, and a signal processing apparatus are provided. The operation method includes receiving an output signal from a first neural network, and converting a first feature included in the output signal to a second feature configured to be input to a second neural network, based on a conversion rule controlling conversion between a feature to be output from the first neural network and a feature to be input to the second neural network. The operation method further includes generating an input signal to be input to the second neural network, based on the second feature, and transmitting the input signal to the second neural network.

First claim

What is claimed is:

1 . An operation method of a neural network, the operation method comprising:
receiving an output signal from a first neural network; converting a first feature included in the output signal to a second feature configured to be input to a second neural network, based on a conversion rule controlling conversion between a feature to be output from the first neural network and a feature to be input to the second neural network; generating an input signal to be input to the second neural network, based on the second feature; and transmitting the input signal to the second neural network.

2 . The operation method of claim 1 , wherein the first feature comprises a first feature vector, and
the second feature comprises a second feature vector.

3 . The operation method of claim 1 , wherein the neural network comprises an interface neural network to which the conversion rule is applied.

4 . The operation method of claim 1 , wherein an input dimension of the neural network corresponds to an output dimension of the first neural network, and
an output dimension of the neural network corresponds to an input dimension of the second neural network.

5 . The operation method of claim 1 , wherein the conversion rule comprises parameters of the neural network that are optimized.

6 . The operation method of claim 1 , further comprising:
in response to the first neural network being replaced with a third neural network, updating the conversion rule to control conversion between a feature to be output from the third neural network and the feature to be input to the second neural network; and in response to the second neural network being replaced with a fourth neural network, updating the conversion rule to control conversion between the feature to be output from the first neural network and a feature to be input to the fourth neural network.

7 . The operation method of claim 6 , wherein the updating the conversion rule to control the conversion between the feature to be output from the third neural network and the feature to be input to the second neural network comprises adjusting parameters of the neural network, based on a relationship between the feature to be output from the third neural network and the feature to be input to the second neural network, and
the updating the conversion rule to control the conversion between the feature to be output from the first neural network and the feature to be input to the fourth neural network comprises adjusting parameters of the neural network, based on a relationship between the feature to be output from the first neural network and the feature to be input to the fourth neural network.

8 . The operation method of claim 6 , wherein the third neural network and the first neural network are distinguished with respect to any one or any combination of an input modality, an output modality, an input dimension, an output dimension, an input feature, and an output feature, and
the fourth neural network and the second neural network are distinguished with respect to any one or any combination of an input modality, an output modality, an input dimension, an output dimension, an input feature, and an output feature.

9 . The operation method of claim 6 , wherein, in response to the first neural network being replaced with the third neural network, a type of an input signal based on the updated conversion rule is identical to a type of the input signal based on the conversion rule, and
in response to the second neural network being replaced with the fourth neural network, a type of an output signal based on the updated conversion rule is identical to a type of the output signal based on the conversion rule.

10 . The operation method of claim 1 , further comprising, in response to a third neural network being additionally connected to the neural network, generating a conversion rule controlling conversion between a feature to be output from the third neural network and the feature to be input to the second neural network.

11 . The operation method of claim 1 , wherein the first neural network is configured to extract, as the first feature, a feature vector from an object, and
the second neural network is configured to identify the object, based on the input signal.

12 . The operation method of claim 1 , wherein the first neural network is configured to determine, as the first feature, a command vector of an actuator, and
the second neural network is configured to control the actuator, based on the input signal.

13 . A training method comprising:
connecting an input layer of an interface neural network to an output layer of a first neural network; connecting an output layer of the interface neural network to an input layer of a second neural network; inputting a training sample to an input layer of the first neural network; obtaining an output signal from an output layer of the second neural network in response to the inputting of the training sample; and training the interface neural network, based on the output signal and a label of the training sample.

14 . The training method of claim 13 , wherein an input dimension of the interface neural network corresponds to an output dimension of the first neural network, and
an output dimension of the interface neural network corresponds to an input dimension of the second neural network.

15 . The training method of claim 13 , wherein the interface neural network is trained to:
convert a first feature included in an output signal that is output from the first neural network to a second feature configured to be input to the second neural network; generate an input signal to be input to the second neural network, based on the second feature; and transmit the input signal to the second neural network.

16 . The training method of claim 15 , wherein the first neural network is configured to extract, as the first feature, a feature vector from an object, and
the second neural network is configured to identify the object, based on the input signal.

17 . The training method of claim 15 , wherein the first neural network is configured to determine, as the first feature, a command vector of an actuator, and
the second neural network is configured to control the actuator, based on the input signal.

18 . A non-transitory computer-readable medium storing a program comprising instructions to control a processor to perform the method of claim 1 .

19 . A signal processing apparatus comprising:
a processor configured to:
receive an output signal from a first neural network;
convert a first feature included in the output signal to a second feature configured to be input to a second neural network, based on a conversion rule controlling conversion between a feature to be output from the first neural network and a feature to be input to the second neural network;
generate an input signal to be input to the second neural network, based on the second feature; and
transmit the input signal to the second neural network.

20 . An operation method of a neural network, the operation method comprising:
receiving an output signal from a first neural network; converting a first feature of an object and included in the output signal to a second feature of the object and configured to be input to a second neural network; generating an input signal to be input to the second neural network, the input signal comprising the second feature of the object; and transmitting the input signal to

Assignees

• Samsung Electronics Co Ltd
• Univ Zuerich

Inventors

• Lee Jun Haeng
• Delbruck Tobi
• Ryu Eric Hyunsurk

Classifications

• G06N3/045 Primary

  Combinations of networks · CPC title

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/082

  modifying the architecture, e.g. adding, deleting or silencing nodes or connections · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/09

  Supervised learning · CPC title