Method for reproducibility of deep learning classifiers using ensembles

Patent metadata

Field	Value
Publication number	US-11574166-B2
Application number	US-202016886344-A
Country	US
Kind code	B2
Filing date	May 28, 2020
Priority date	May 28, 2020
Publication date	Feb 7, 2023
Grant date	Feb 7, 2023

Abstract

Example implementations described herein involve systems and methods for generating an ensemble of deep learning or neural network models, which can involve, for a training set of data, generating a plurality of model samples for the training set of data, the plurality of model samples generated from deep learning or neural network methods; and aggregating output of the model samples to generate an output of the ensemble models.

First claim

What is claimed is:

1. A method of generating an ensemble of deep learning or neural network models, the method comprising:
for a training set of data:
generating a plurality of model samples for the training set of data, the plurality of model samples generated from deep learning or neural network methods; and
aggregating output of the model samples to generate an output of the ensemble models,
wherein the generating the plurality of model samples for the training set of data comprises:
executing an inference process on the plurality of model samples to generate additional model samples through passing inputs on the plurality of model samples; and
employing a dropout process on the plurality of model samples and the additional model samples to obtain a subset of model samples as the plurality of model samples, the dropout process configured to reduce the plurality of model samples and the additional model samples to the subset of the model samples based on validation accuracy against the training set of data.

2. The method of claim 1 , wherein the generating the plurality of model samples for the training set of data further comprises executing a distributed training process during a training phase of the plurality of model samples, the distributed training process training each of the plurality of model samples across separate servers, each of the separate servers configured to generate a model sample.

3. The method of claim 1 , wherein the plurality of model samples are predictive maintenance models, and wherein the output is a maintenance recommendation.

4. A non-transitory computer readable medium, storing instructions of generating an ensemble of deep learning or neural network models, the instructions comprising:
for a training set of data:
generating a plurality of model samples for the training set of data, the plurality of model samples generated using sampling of trained learners generated from deep learning or neural network methods; and
aggregating output of the model samples to generate an output of the ensemble models,
wherein generating the plurality of model samples for the training set of data comprises:
executing an inference process on the plurality of model samples to generate additional model samples through passing inputs on the plurality of model samples; and
employing a dropout process on the plurality of model samples and the additional model samples to obtain a subset of model samples as the plurality of model samples, the dropout process configured to reduce the plurality of model samples and the additional model samples to the subset of the model samples based on validation accuracy against the training set of data.

5. The non-transitory computer readable medium of claim 4 , wherein the generating the plurality of model samples for the training set of data further comprises executing a distributed training process during a training phase of the plurality of model samples, the distributed training process training each of the plurality of model samples across separate servers, each of the separate servers configured to generate a model sample.

6. The non-transitory computer readable medium of claim 4 , wherein the plurality of model samples are predictive maintenance models, and wherein the output is a maintenance recommendation.

7. An apparatus configured to generate an ensemble of deep learning or neural network models, the apparatus comprising:
a processor, configured to:
for a training set of data:
generate a plurality of model samples for the training set of data, the plurality of model samples generated from deep learning or neural network methods; and
aggregate output of the model samples to generate an output of the ensemble models,
wherein the processor is configured to generate the plurality of model samples for the training set of data by,
executing an inference process on the plurality of model samples to generate additional model samples through passing inputs on the plurality of model samples; and
employing a dropout process on the plurality of model samples and the additional model samples to obtain a subset of model samples as the plurality of model samples, the dropout process configured to reduce the plurality of model samples and the additional model samples to the subset of the model samples based on validation accuracy against the training set of data.

8. The apparatus of claim 7 , wherein the processor is configured to generate the plurality of model samples for the training set of data by further executing a distributed training process during a training phase of the plurality of model samples, the distributed training process training each of the plurality of model samples across separate servers, each of the separate servers configured to generate a model sample.

9. The apparatus of claim 7 , wherein the plurality of model samples are predictive maintenance models, and wherein the output is a maintenance recommendation.

Assignees

• Hitachi Ltd

Inventors

• Ghosh Dipanjan
• Gonzalez Diaz Maria Teresa
• Alam Mahbubul
• Farahat Ahmed
• Gupta Chetan
• Wang Lijing

Classifications

• G06N3/098

  Distributed learning, e.g. federated learning · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · 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