Back to results
US 10,540,958 B2Granted
Neural network training method and apparatus using experience replay sets for recognition
AssigneeSamsung Electronics Co., Ltd.
InventorYunhong Min
Life of this patent
Estimating the term…
Abstract
A training method and apparatus for speech recognition is disclosed, where an example of the training method includes determining whether a current iteration for training a neural network is performed by an experience replay iteration using an experience replay set, selecting a sample from at least one of the experience replay set and a training set based on a result of the determining, and training the neural network based on the selected sample.
Claims31
A machine learning method of a speech recognition neural network, the method comprising:
- determining whether a current iteration for training the neural network, subsequent to a prior training iteration of the neural network, is performed by an experience replay iteration using an experience replay set or a training set;
- selecting a sample from at least one of the experience replay set or the training set based on a result of the determining;
- training the speech recognition neural network based on the selected sample;
- calculating a quality of the selected sample based on a difference between a training error reduction corresponding to the iteration prior to the current iteration and a training error reduction corresponding to the current iteration;
- comparing the calculated quality to qualities of the samples in the experience replay set;
- and
- determining whether to update the experience replay set based on a result of the comparing.
4 more elements
A machine learning method of a speech recognition neural network, the method comprising:
- training the speech recognition neural network based on a sample selected from at least one of an experience replay set or a training set for a current iteration for training the neural network;
- calculating a quality of the selected sample based on a difference between a training error reduction corresponding to an iteration prior to the current iteration and a training error reduction corresponding to the current iteration;
- comparing the calculated quality to qualities of the samples in the experience replay set;
- determining whether to update the experience replay set based on a result of the training and the comparing;
- and
- updating the experience replay set based on a result of the determining.
3 more elements
A machine learning apparatus for a speech recognition neural network comprising:
- a processor;
- a memory configured to store an instruction readable by the processor,
- wherein, when the instruction is executed by the processor, the processor is configured to:
- determine whether a current iteration for training the neural network, subsequent to a prior training iteration of the neural network, is performed by an experience replay iteration using an experience replay set or a training set;
- select a sample from at least one of the experience replay set or the training set based on a result of the determining;
- train the speech recognition neural network based on the selected sample;
- calculate a quality of the selected sample based on a difference between a training error reduction corresponding to the iteration prior to the current iteration and a training error reduction corresponding to the current iteration;
- compare the calculated quality to qualities of the samples in the experience replay set;
- and
- determine whether to update the experience replay set based on a result of the comparing.
7 more elements
A machine learning apparatus for a speech recognition neural network comprising:
- a processor;
- a memory configured to store an instruction readable by the processor,
- wherein, when the instruction is executed by the processor, the processor is configured to:
- train the speech recognition neural network based on a sample selected from at least one of an experience replay set or a training set in association with a current iteration for training the neural network;
- calculate a quality of the selected sample based on a difference between a training error reduction corresponding to an iteration prior to the current iteration and a training error reduction corresponding to the current iteration;
- compare the calculated quality to qualities of the samples in the experience replay set;
- determine whether to update the experience replay set based on a result of the training and the comparing;
- and
- update the experience replay set based on a result of the determining.
6 more elements
Dependencies are inferred from the claim text.