Batch normalization layers

• US 10,417,562 B2
• Filed: 01/28/2016
• Issued: 09/17/2019
• Est. Priority Date: 01/28/2015
• Status: Active Grant

• Current Assignee: Google LLC (Alphabet Inc.)
• Original Assignee: Google LLC (Alphabet Inc.)
• Inventors: Ioffe, Sergey, Cortes, Corinna
• Primary Examiner(s): Waldron, Scott A.
• Assistant Examiner(s): Tamulonis, Fen Christopher
• Application Number:
  US15/009,647
• Publication Number:
  US 20160217368A1
• Time in Patent Office: 1,328 Days
• CPC Class Codes: G06F 18/10 Pre-processing; Data cleansing
• CPC Class Codes: G06F 18/2415 based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06N 3/084 Backpropagation, e.g. using gradient descent
• CPC Class Codes: G06T 2207/20081 Training; Learning
• CPC Class Codes: G06V 10/70 using pattern recognition or machine learning optical pattern recognition or electronic computations therefor G06V10/88
• CPC Class Codes: G06V 10/82 using neural networks

Abstract

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for processing inputs using a neural network system that includes a batch normalization layer. One of the methods includes receiving a respective first layer output for each training example in the batch; computing a plurality of normalization statistics for the batch from the first layer outputs; normalizing each component of each first layer output using the normalization statistics to generate a respective normalized layer output for each training example in the batch; generating a respective batch normalization layer output for each of the training examples from the normalized layer outputs; and providing the batch normalization layer output as an input to the second neural network layer.

Claims

1. A neural network system implemented by one or more computers, the neural network system comprising:

instructions for implementing a batch normalization layer between a first neural network layer and a second neural network layer in a neural network, wherein the first neural network layer generates first layer outputs having a plurality of components, and wherein the instructions cause the one or more computers to perform operations comprising;

during training of the neural network on a plurality of batches of training data, each batch comprising a respective plurality of training examples and for each of the batches;

receiving a respective first layer output for each of the plurality of training examples in the batch;

computing a plurality of normalization statistics for the batch from the first layer outputs, comprising;

determining, for each of a plurality of subsets of the plurality of the components of the first layer outputs, a mean of the components of the first layer outputs for each of the plurality of training examples in the batch that are in the respective subset, and

determining, for each of the plurality of subsets of the plurality of the components of the first layer outputs, a standard deviation of the components of the first layer outputs for each of the plurality of training examples in the batch that are in the respective subset;

normalizing each of the plurality of the components of each first layer output using the normalization statistics to generate a respective normalized layer output for each training example in the batch, comprising;

for each first layer output and for each of the plurality of subsets, normalizing the components of the first layer output that are in the respective subset using the mean for the respective subset and the standard deviation for the respective subset;

generating a respective batch normalization layer output for each of the training examples from the normalized layer outputs; and

providing the batch normalization layer output as an input to the second neural network layer.

2. The neural network system of claim 1, wherein the plurality of the components of the first layer output are indexed by dimension, and wherein computing a plurality of normalization statistics for the first layer outputs comprises:

computing, for each of the dimensions, a mean of the components of the first layer outputs in the dimension; and

computing, for each of the dimensions, a standard deviation of the components of the first layer outputs in the dimension.

3. The neural network system of claim 2, wherein normalizing each of the plurality of the components of each first layer output comprises:

normalizing the component using the computed mean and computed standard deviation for the dimension corresponding to the component.

4. The neural network system of claim 2, wherein generating the respective batch normalization layer output for each of the training examples from the normalized layer outputs comprises:

transforming, for each dimension, the component of the normalized layer output for the training example in the dimension in accordance with current values of a set of parameters for the dimension.

5. The neural network system of claim 4, wherein the operations further comprise, after the neural network system has been trained to determine trained values of the parameters for each of the dimensions:

receiving a new first layer output generated by the first neural network layer for a new neural network input;

normalizing each component of the new first layer output using pre-computed mean and standard deviation statistics for the dimensions to generate a new normalized layer output;

generating a new batch normalization layer output by transforming, for each dimension, the component of the new normalized layer output for the training example in the dimension in accordance with trained values of the set of parameters for the dimension; and

providing the batch normalization layer output as a new layer input to the second neural network layer.

6. The neural network system of claim 5, wherein the pre-computed mean and standard deviation statistics for the dimensions are computed from first layer outputs generated by the first neural network layer during the training of the neural network.

7. The neural network system of claim 5, wherein the pre-computed mean and standard deviation statistics for the dimensions are computed from new first layer outputs generated by the first neural network layer after the neural network has been trained.

8. The neural network system of claim 7, wherein new neural network inputs processed by the neural network system after the neural network has been trained are a different type of input than the training examples used to train the neural network.

9. The neural network system of claim 1, wherein the first neural network layer is a convolutional layer, wherein the plurality of the components of the first layer output are indexed by feature index and spatial location index, and wherein computing a plurality of normalization statistics for the first layer outputs comprises:

computing, for each combination of feature index and spatial location index, a mean of the components of the first layer outputs having the feature index and spatial location index;

computing, for each feature index, an average of the means for combinations that include the feature index;

computing, for each combination of feature index and spatial location index, a variance of the components of the first layer outputs having the feature index and spatial location index; and

computing, for each feature index, an average of the variances for combinations that include the feature index.

10. The neural network system of claim 9, wherein normalizing each of the plurality of the components of each layer output comprises:

normalizing the component using the average of the means and the average of the variances for the feature index corresponding to the component.

11. The neural network system of claim 9, wherein generating the respective batch normalization layer output for each of the training examples from the normalized layer outputs comprises:

transforming each of the plurality of the components of the normalized layer output in accordance with current values of a set of parameters for the feature index corresponding to the component.

12. The neural network system of claim 11, wherein the operations further comprise, after the neural network has been trained to determine trained values of the parameters for each of the dimensions:

receive a new first layer input generated from a new neural network input;

normalize each of the plurality of the components of the new first layer output using pre-computed mean and standard deviation statistics for the feature indices to generate a new normalized layer output;

generate a new batch normalization layer output by transforming each of the plurality of the components of the normalized layer output in accordance with trained values of the set of parameters for the feature index corresponding to the component; and

provide the new batch normalization layer output as a new layer input to the second neural network layer.

13. The neural network system of claim 1, wherein the first neural network layer is a convolutional layer, wherein the plurality of the components of the first layer output are indexed by feature index and spatial location index, and wherein computing a plurality of normalization statistics for the first layer outputs comprises, for each of the feature indices:

computing a mean of the components of the first layer outputs that correspond to the feature index; and

computing a variance of the components of the first layer outputs that correspond to the feature index.

14. The neural network system of claim 13, wherein normalizing each of the plurality of the components of each layer output comprises:

normalizing the component using the mean and the variance for the feature index corresponding to the component.

15. The neural network system of claim 13, wherein generating the respective batch normalization layer output for each of the training examples from the normalized layer outputs comprises:

transforming each of the plurality of the components of the normalized layer output in accordance with current values of a set of parameters for the feature index corresponding to the component.

16. The neural network system of claim 15, wherein the operations further comprise, after the neural network has been trained to determine trained values of the parameters for each of the dimensions:

receive a new first layer input generated from a new neural network input;

normalize each component of the new first layer output using pre-computed mean and standard deviation statistics for the feature indices to generate a new normalized layer output;

generate a new batch normalization layer output by transforming each of the plurality of the components of the normalized layer output in accordance with trained values of the set of parameters for the feature index corresponding to the component; and

provide the new batch normalization layer output as a new layer input to the second neural network layer.

17. The neural network system of claim 1, wherein the first neural network layer generates the first layer outputs by modifying first layer inputs in accordance with current values of a set of parameters for the first neural network layer.

18. The neural network system of claim 17, wherein the second neural network layer generates second layer outputs by applying a non-linear operation to the batch normalization layer outputs.

19. The neural network system of claim 1, wherein the first neural network layer generates the first layer outputs by modifying first layer inputs in accordance with current values of a set of parameters to generate modified first layer inputs and then applying a non-linear operation to the modified first layer inputs.

20. The neural network system of claim 1, wherein, during the training of the neural network, the neural network system is configured to backpropagate the normalization statistics as part of adjusting values of parameters of the neural network.

21. A method performed by one or more computers implementing a batch normalization layer that is between a first neural network layer and a second neural network layer in a neural network, wherein the first neural network layer generates first layer outputs having a plurality of components, and wherein the method comprises;

during training of the neural network on a plurality of batches of training data, each batch comprising a respective plurality of training examples and for each of the batches;

receiving a respective first layer output for each of the plurality of training examples in the batch;

computing a plurality of normalization statistics for the batch from the first layer outputs, comprising;

determining, for each of a plurality of subsets of the plurality of the components of the first layer outputs, a mean of the components of the first layer outputs for each of the plurality of training examples in the batch that are in the respective subset, and

determining, for each of the plurality of subsets of the plurality of the components of the first layer outputs, a standard deviation of the components of the first layer outputs for each of the plurality of training examples in the batch that are in the respective subset;

normalizing each of the plurality of the components of each first layer output using the normalization statistics to generate a respective normalized layer output for each training example in the batch, comprising;

for each first layer output and for each of the plurality of subsets, normalizing the components of the first layer output that are in the respective subset using the mean for the respective subset and the standard deviation for the respective subset;

generating a respective batch normalization layer output for each of the training examples from the normalized layer outputs; and

providing the batch normalization layer output as an input to the second neural network layer.

29. The method of claim 21, wherein the plurality of the components of the first layer output are indexed by dimension, and wherein computing a plurality of normalization statistics for the first layer outputs comprises:

computing, for each of the dimensions, a mean of the components of the first layer outputs in the dimension; and

computing, for each of the dimensions, a standard deviation of the components of the first layer outputs in the dimension.

30. The method of claim 29, wherein normalizing each of the plurality of the components of each first layer output comprises:

normalizing the component using the computed mean and computed standard deviation for the dimension corresponding to the component.

31. The method of claim 29, wherein generating the respective batch normalization layer output for each of the training examples from the normalized layer outputs comprises:

transforming, for each dimension, the component of the normalized layer output for the training example in the dimension in accordance with current values of a set of parameters for the dimension.

32. The method of claim 21, wherein the first neural network layer is a convolutional layer, wherein the plurality of the components of the first layer output are indexed by feature index and spatial location index, and wherein computing a plurality of normalization statistics for the first layer outputs comprises, for each of the feature indices:

computing a mean of the components of the first layer outputs that correspond to the feature index; and

computing a variance of the components of the first layer outputs that correspond to the feature index.

33. The method of claim 32, wherein normalizing each of the plurality of the components of each layer output comprises:

normalizing the component using the mean and the variance for the feature index corresponding to the component.

34. The method of claim 32, wherein generating the respective batch normalization layer output for each of the training examples from the normalized layer outputs comprises:

transforming each of the plurality of the components of the normalized layer output in accordance with current values of a set of parameters for the feature index corresponding to the component.

22. One or more non-transitory computer-readable storage media encoded with a computer program, the computer program comprising instructions that when executed by one or more computers cause the one or more computers to implement a neural network system, the neural network system comprising:

batch normalization instructions for implementing a batch normalization layer between a first neural network layer and a second neural network layer in a neural network, wherein the first neural network layer generates first layer outputs having a plurality of components, and wherein the batch normalization instructions cause the one or more computers to perform operations comprising;

during training of the neural network on a plurality of batches of training data, each batch comprising a respective plurality of training examples and for each of the batches;

receiving a respective first layer output for each of the plurality of training examples in the batch;

computing a plurality of normalization statistics for the batch from the first layer outputs, comprising;

determining, for each of a plurality of subsets of the plurality of the components of the first layer outputs, a mean of the components of the first layer outputs for each of the plurality of training examples in the batch that are in the respective subset, and

determining, for each of the plurality of subsets of the plurality of the components of the first layer outputs, a standard deviation of the components of the first layer outputs for each of the plurality of training examples in the batch that are in the respective subset;

normalizing each of the plurality of the components of each first layer output using the normalization statistics to generate a respective normalized layer output for each training example in the batch, comprising;

for each first layer output and for each of the plurality of subsets, normalizing the components of the first layer output that are in the respective subset using the mean for the respective subset and the standard deviation for the respective subset;

generating a respective batch normalization layer output for each of the training examples from the normalized layer outputs; and

providing the batch normalization layer output as an input to the second neural network layer.

23. The non-transitory computer-readable storage media of claim 22, wherein the plurality of the components of the first layer output are indexed by dimension, and wherein computing a plurality of normalization statistics for the first layer outputs comprises:

computing, for each of the dimensions, a mean of the components of the first layer outputs in the dimension; and

computing, for each of the dimensions, a standard deviation of the components of the first layer outputs in the dimension.

24. The non-transitory computer-readable storage media of claim 23, wherein normalizing each of the plurality of the components of each first layer output comprises:

normalizing the component using the computed mean and computed standard deviation for the dimension corresponding to the component.

25. The non-transitory computer-readable storage media of claim 23, wherein generating the respective batch normalization layer output for each of the training examples from the normalized layer outputs comprises:

transforming, for each dimension, the component of the normalized layer output for the training example in the dimension in accordance with current values of a set of parameters for the dimension.

26. The non-transitory computer-readable storage media of claim 22, wherein the first neural network layer is a convolutional layer, wherein the plurality of the components of the first layer output are indexed by feature index and spatial location index, and wherein computing a plurality of normalization statistics for the first layer outputs comprises, for each of the feature indices:

computing a mean of the components of the first layer outputs that correspond to the feature index; and

computing a variance of the components of the first layer outputs that correspond to the feature index.

27. The non-transitory computer-readable storage media of claim 26, wherein normalizing each of the plurality of the components of each layer output comprises:

normalizing the component using the mean and the variance for the feature index corresponding to the component.

28. The non-transitory computer-readable storage media of claim 26, wherein generating the respective batch normalization layer output for each of the training examples from the normalized layer outputs comprises:

transforming each of the plurality of the components of the normalized layer output in accordance with current values of a set of parameters for the feature index corresponding to the component.