Neural networks for discovering latent factors from data

Patent metadata

Field	Value
Publication number	US-11468265-B2
Application number	US-201916389532-A
Country	US
Kind code	B2
Filing date	Apr 19, 2019
Priority date	May 15, 2018
Publication date	Oct 11, 2022
Grant date	Oct 11, 2022

Abstract

Aspects of the disclosure relate to a method for discovering latent factors from data in a neural network environment. Aspects include adding, by a data noise unit of the neural network environment, noise to a set of input data; computing, by the encoder model and a set of stochastic non-linear units, a set of latent code based on the set of input data; obtaining, by decoding the latent code with a decoder model, a set of reconstructed data.

First claim

What is claimed is:

1. A method for discovering latent factors from data in a neural network environment, the method comprising:
adding, by a data noise unit of the neural network environment, noise to a set of input data;
computing, by the encoder model and a set of stochastic non-linear units, a set of latent code based on the set of input data; and
obtaining, by decoding the latent code with a decoder model, a set of reconstructed data;
applying, with respect to the neural network, a loss function to obtain a code loss value;
calculating, from the set of reconstructed data, a reconstruction loss value that indicates a similarity of the set of reconstructed data with respect to the set of input data;
computing, using the reconstruction loss value and the code loss value, a total loss value; and
performing, with respect to the neural network, a feedback technique to reduce the total loss value.

2. The method of claim 1 , further comprising:
utilizing both the set of input data and the set of latent code computed by the encoder network for the set of input data to train a classification model.

3. The method of claim 1 , further comprising:
providing, to a code translation model, a set of pre-determined code;
providing the set of reconstructed data to a classification model configured to classify a plurality of desired feature attributes;
training the code translation model using the total loss value obtained from the classification model; and
generating, using the code translation network, a set of attribute code that is configured to produce a set of desired attributes when provided to the decoder model.

4. The method of claim 1 , further comprising:
receiving, from a first classification model, a set of misclassified data samples;
computing, by the encoder model, a set of latent code for the set of misclassified data samples;
generating, by the decoder model, a set of data samples based on the set of latent code for the set of misclassified data samples; and
performing, using the set of data samples based on the set of latent code for the set of misclassified data samples, a machine learning technique with respect to a second classification model.

5. The method of claim 1 , further comprising
performing, with respect to a classification model, a machine learning technique using a set of encrypted data;
computing, by an encoder model, a second set of latent code corresponding to a set of unencrypted data; and
performing, with respect to the classification model in response to both performing the machine learning technique using the set of encrypted data and computing the second set of latent code corresponding to the set of unencrypted data, a fine tuning technique using the set of unencrypted data and the set of latent code corresponding to the set of unencrypted data.

6. The method of claim 1 , further comprising:
receiving, by the neural network, a set of noisy data;
computing, by the encoder model and the set of stochastic non-linear units, a second set of latent code based on the set of noisy data; and
obtaining, by decoding the second set of latent code with the decoder model, a set of de-noised data.

7. The method of claim 1 , further comprising:
receiving a second set of input data;
computing, by the encoder model and the set of stochastic non-linear units, a second set of latent code for the second set of input data; and
assigning, to the second set of input data, a set of labels corresponding to the second set of latent code for the second set of input data.

8. The method of claim 1 , further comprising:
transmitting, to a client device, the set of latent code based on the set of input data; and
reconstructing, by the client device using the decoder model, the set of input data from the set of latent code.

9. A method for discovering latent factors from data in a neural network environment, the method comprising:
adding, by a data noise unit of the neural network environment, noise to a set of input data;
computing, by the encoder model and a set of stochastic non-linear units, a set of latent code based on the set of input data; and
obtaining, by decoding the latent code with a decoder model, a set of reconstructed data;
receiving a first set of latent code for a first set of data;
receiving a second set of latent code for a second set of data;
determining, by comparing the first set of latent code with the second set of latent code, whether the first set of data achieves a similarity threshold with respect to the second set of data; and
merging, in response to determining that the first set of data achieves the similarity threshold with respect to the second set of data, the first set of data and the second set of data to generate a combined data set.

10. A device for discovering latent factors from data in a neural network environment, the device comprising:
a memory having a set of computer readable computer instructions, and
a processor for executing the set of computer readable instructions, the set of computer readable instructions including:
adding, by a data noise unit of the neural network environment, noise to a set of input data;
computing, by the encoder model and a set of stochastic non-linear units, a set of latent code based on the set of input data;
obtaining, by decoding the latent code with a decoder model, a set of reconstructed data;
applying, with respect to the neural network, a loss function to obtain a code loss value;
calculating, from the set of reconstructed data, a reconstruction loss value that indicates a similarity of the set of reconstructed data with respect to the set of input data;
computing, using the reconstruction loss value and the code loss value, a total loss value; and
performing, with respect to the neural network, a feedback technique to reduce the total loss value.

11. The device of claim 10 , further comprising:
utilizing both the set of input data and the set of latent code computed by the encoder network for the set of input data to train a classification model.

12. A computer program product for data orchestration platform management in a network communication environment including a set of information sources, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, wherein the computer readable storage medium is not a transitory signal per se, the program instructions executable by a processor to cause the processor to perform a method comprising:
adding, by a data noise unit of the neural network environment, noise to a set of input data;
computing, by the encoder model and a set of stochastic non-linear units, a set of latent code based on the set of input data;
obtaining, by decoding the latent code with a decoder model, a set of reconstructed data;
providing, to a code translation model, a set of pre-determined code;
providing the set of reconstructed data to a classification model configured to classify a plurality of desired feature attributes;
training the code translation model using the total loss value obtained from the classification model; and
generating, using the code translation network, a set of attribute code that is configured to produce a set of desired attributes when provided to the decoder model.

Assignees

• Hitachi Ltd

Inventors

• Chhabra Mohit
• Klinkigt Martin
• Murakami Tomokazu

Classifications

• G06F18/2148

  characterised by the process organisation or structure, e.g. boosting cascade · CPC title

• G06N3/045 Primary

  Combinations of networks · CPC title

• G06F18/2155

  characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling · CPC title

• G06F18/2113

  by ranking or filtering the set of features, e.g. using a measure of variance or of feature cross-correlation · CPC title

• G06N3/0455 Primary

  Auto-encoder networks; Encoder-decoder networks · CPC title