Training of student neural network with teacher neural networks

1. A computer-implemented method, comprising:

inputting input data to each teacher neural network among a plurality of teacher neural networks to obtain a soft label output among a plurality of soft label outputs from each teacher neural network among the plurality of teacher neural networks, wherein the plurality of teacher neural networks includes two or more different types of teacher neural networks;

evaluating an accuracy of each of the plurality of teacher neural networks using test data;

iteratively training a student neural network with the input data and the plurality of soft label outputs by selecting two or more of the plurality of teacher neural networks in a predetermined order for each of two or more training iterations;

iteratively increasing a frequency of selecting a particular teacher neural network among the plurality of teacher neural networks based on a comparison of an accuracy of the soft label output from the particular teacher neural network with corresponding correct training data; and

adjusting one or more of a plurality of weights between nodes in the selected teacher neural network based on the comparison as a number of iterations increases.

2. The method of claim 1 , wherein training the student neural network includes

repeating the iterations, each iteration including:

inputting the input data into the student neural network,

comparing an output data of the student neural network with the soft label output among the plurality of soft label outputs,

adjusting a plurality of weights in the student neural network based on the comparison, and

using the plurality of soft label outputs for the each iteration of the training the student neural network.

3. The method of claim 2 , wherein using the plurality of soft label outputs includes:

using the soft label output of a less accurate teacher neural network among the plurality of teacher neural networks earlier than the soft label output of an accurate teacher neural network among the plurality of teacher neural networks.

4. The method of claim 1 , wherein training the student neural network includes,

using the soft label output of the most accurate teacher neural network among the plurality of teacher neural networks for the input data.

5. The method of claim 1 , wherein the input data is audio data and each soil label output is a classification of the audio data.

6. The method of claim 5 , wherein the classification of the audio data identifies phonemes.

7. The method of claim 1 , wherein training the student neural network includes;

training the student neural network with at least the input data, the plurality of soft label outputs, and a correct data corresponding to the input data.

8. The method of claim 1 , further comprising:

training the plurality of teacher neural networks by a server computer, and

receiving, by the server computer, the input data from a client computer,

wherein inputting an input data to each teacher neural network among a plurality of teacher neural networks includes:

inputting, by the server computer, the input data to each teacher neural network to obtain the soft label output from each teacher neural network, and

transmitting, by the server computer, the plurality of soft label outputs to the client computer,

wherein training the student neural network includes:

training, by the client computer, the student neural network with at least the input data and the plurality of soft label outputs.

9. A neural network training apparatus comprising:

a processor or a programmable circuitry; and

one or more computer readable mediums collectively including instructions that, when executed by the processor or the programmable circuitry, cause the processor or the programmable circuitry to perform operations including:

inputting an input data to each teacher neural network among a plurality of teacher neural networks to obtain a soft label output among a plurality of soft label outputs from each teacher neural network among the plurality of teacher neural networks, wherein the plurality of teacher neural networks includes two or more different types of teacher neural networks;

evaluating an accuracy of each of the plurality of teacher neural networks using test data;

iteratively training a student neural network with the input data and the plurality of soft label outputs by selecting two or more of the plurality of teacher neural networks in a predetermined order for each of two or more training iterations;

iteratively increasing a frequency of selecting a particular teacher neural network among the plurality of teacher neural networks based on a comparison of an accuracy of the soft label output from the particular teacher neural network with corresponding correct training data; and

adjusting one or more of a plurality of weights between nodes in the selected teacher neural network based on the comparison as a number of iterations increases.

10. The apparatus of claim 9 , wherein training the student neural network includes:

repeating the iterations, each iteration including:

inputting the input data into the student neural network,

comparing an output data of the student neural network with the soft label output among the plurality of soft label outputs,

adjusting a plurality of weights in the student neural network based on the comparison, and

using the plurality of soft label outputs for the each iteration of the training the student neural network.

11. The apparatus of claim 10 , wherein using the soft label outputs includes:

using the soft label output of a less accurate teacher neural network among the plurality of teacher neural networks in an earlier iteration than the soft label output of an accurate teacher neural network among the plurality of teacher neral networks.

12. The apparatus of claim 9 , wherein training the student neural network includes:

using the soft label output of the most accurate teacher neural network among the plurality of teacher neural networks in a last iteration.

13. The apparatus of claim 9 , wherein the input data is audio data and each soft label output is a classification of the audio data.

14. The apparatus of claim 13 , wherein the classification of the audio data identifies phonemes.

15. A computer program product including one or more computer readable storage mediums collectively storing program instructions for improving neural network training that are executable by a processor or programmable circuitry to cause the processor or programmable circuitry to perform operations comprising:

inputting an input data to each teacher neural network among a plurality of teacher neural networks to obtain a soft label output among a plurality of soft label outputs from each teacher neural network among the plurality of teacher neural networks, wherein the plurality of teacher neural networks includes two or more different types of teacher neural networks;

evaluating an accuracy of each of the plurality of teacher neural networks using test data;

iteratively training a student neural network with the input data and the plurality of soft label outputs by selecting two or more of the plurality of teacher neural networks in a predetermined order for each of two or more training iterations;

iteratively increasing a frequency of selecting a particular teacher neural network among the plurality of teacher neural networks based on a comparison of an accuracy of the soft label output from the particular teacher neural network with corresponding correct training data; and

adjusting one or more of a plurality of weights between nodes in the selected teacher neural network based on the comparison as a number of iterations increases.

16. The computer program product of claim 15 , wherein training the student neural network includes:

repeating the iterations, each iteration including:

inputting the input data into the student neural network,

comparing an output data of the student neural network with the soft label output among the plurality of soft label outputs,

adjusting a plurality of weights in the student neural network based on the comparison, and

using the soft label outputs for the each iteration of the training the student neural network.

17. The computer program product of claim 16 , wherein using the soft label outputs includes:

using the soft label output of a less accurate teacher neural network among the plurality of teacher neural networks in an earlier iteration than the soft label output of an accurate teacher neural network among the plurality of teacher neural networks.

18. The computer program product of claim 15 , wherein training the student neural network includes:

using the soft label output of the most accurate teacher neural network among the plurality of teacher neural networks in the last iteration.

19. The computer program product of claim 15 , wherein the input data is audio data and each soft label output is a classification of the audio data.

20. The computer program product of claim 19 , wherein the classification of the audio data identifies phonemes.