Training machine learning models to be robust against label noise

1. A method of training a machine learning model having a plurality of model parameters and configured to receive a model input and to process the model input in accordance with the model parameters to generate a model output based on the model input, the method comprising:

obtaining a training input and a corresponding ground truth output;

processing the training input using the machine learning model and in accordance with current values of the model parameters to generate a training output based on the training input;

computing a loss for the training output by evaluating an objective function that measures a difference between the training output and the ground truth output, wherein the objective function is composed of a base loss and a link function; and

determining an update to current values of the model parameters, comprising:

determining, with respect to the model parameters, a first partial gradient of the loss with respect to the base loss and a second partial gradient of the loss with respect to the link function;

regularizing the first partial gradient of the loss to generate a regularized first partial gradient of the loss;

generating a recomposition of the regularized partial first gradient of the loss and the second partial gradient of the loss; and

computing the update from the generated recomposition.

2. The method of claim 1 , wherein regularizing the first partial gradient of the loss comprises:

determining whether a norm of the first partial gradient of the loss exceeds a particular threshold; and

when the norm of the first partial gradient of the loss exceeds the particular threshold:

clipping the first partial gradient of the loss to equal the particular threshold.

3. The method of claim 1 , wherein:

the machine learning model is a neural network; and

computing the update from the generated recomposition comprises backpropagating the generated recomposition through network parameters.

4. The method of claim 1 , wherein the base loss comprises a log loss, a square loss, a Matsushita loss, a boosting loss, or a Q-loss.

5. The method of claim 1 , wherein the link function comprises a sigmoid function, a probit function, a capped linear function, or a Matsushita function.

6. The method of claim 1 , wherein the model input includes one or more input variables, and the model output specifies a respective predicted classification for each input variable.

7. The method of claim 1 , further comprising:

providing data specifying the trained machine learning model.

8. The method of claim 1 , further comprising:

using the trained machine learning model to process new training inputs.

9. A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform operations for training a machine learning model having a plurality of model parameters and configured to receive a model input and to process the model input in accordance with the model parameters to generate a model output based on the model input, the operations comprising:

obtaining a training input and a corresponding ground truth output;

processing the training input using the machine learning model and in accordance with current values of the model parameters to generate a training output based on the training input;

computing a loss for the training output by evaluating an objective function that measures a difference between the training output and the ground truth output, wherein the objective function is composed of a base loss and a link function; and

determining an update to current values of the model parameters, comprising:

determining, with respect to the model parameters, a first partial gradient of the loss with respect to the base loss and a second partial gradient of the loss with respect to the link function;

regularizing the first partial gradient of the loss to generate a regularized first partial gradient of the loss;

generating a recomposition of the regularized partial first gradient of the loss and the second partial gradient of the loss; and

computing the update from the generated recomposition.

10. The system of claim 9 , wherein regularizing the first partial gradient of the loss comprises:

determining whether a norm of the first partial gradient of the loss exceeds a particular threshold; and

when the norm of the first partial gradient of the loss exceeds the particular threshold:

clipping the first partial gradient of the loss to equal the particular threshold.

11. The system of claim 9 , wherein:

the machine learning model is a neural network; and

computing the update from the generated recomposition comprises backpropagating the generated recomposition through network parameters.

12. The system of claim 9 , wherein the base loss comprises a log loss, a square loss, a Matsushita loss, a boosting loss, or a Q-loss.

13. The system of claim 9 , wherein the link function comprises a sigmoid function, a probit function, a capped linear function, or a Matsushita function.

14. The system of claim 9 , wherein the model input includes one or more input variables, and the model output specifies a respective predicted classification for each input variable.

15. One or more non-transitory computer-readable storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations for training a machine learning model having a plurality of model parameters and configured to receive a model input and to process the model input in accordance with the model parameters to generate a model output based on the model input, the operations comprising: obtaining a training input and a corresponding ground truth output;

processing the training input using the machine learning model and in accordance with current values of the model parameters to generate a training output based on the training input;

computing a loss for the training output by evaluating an objective function that measures a difference between the training output and the ground truth output, wherein the objective function is composed of a base loss and a link function; and

determining an update to current values of the model parameters, comprising:

determining, with respect to the model parameters, a first partial gradient of the loss with respect to the base loss and a second partial gradient of the loss with respect to the link function;

regularizing the first partial gradient of the loss to generate a regularized first partial gradient of the loss;

generating a recomposition of the regularized partial first gradient of the loss and the second partial gradient of the loss; and

computing the update from the generated recomposition.

16. The computer-readable storage media of claim 15 , wherein regularizing the first partial gradient of the loss comprises:

determining whether a norm of the first partial gradient of the loss exceeds a particular threshold; and

when the norm of the first partial gradient of the loss exceeds the particular threshold:

clipping the first partial gradient of the loss to equal the particular threshold.

17. The computer-readable storage media of claim 15 , wherein:

the machine learning model is a neural network; and

computing the update from the generated recomposition comprises backpropagating the generated recomposition through network parameters.

18. The computer-readable storage media of claim 15 , wherein the base loss comprises a log loss, a square loss, a Matsushita loss, a boosting loss, or a Q-loss.

19. The computer-readable storage media of claim 15 , wherein the link function comprises a sigmoid function, a probit function, a capped linear function, or a Matsushita function.

20. The computer-readable storage media of claim 15 , wherein the model input includes one or more input variables, and the model output specifies a respective predicted classification for each input variable.