Forecasting uncertainty in machine learning models
Provided are systems and methods for generating a score for any model which can be updated online, regardless of model type architecture and parameters, leveraging relations between regret and uncertainty.
1. A computer-implemented method, the method comprising:
for each of a plurality of examples, each example comprising an input and a label:
processing, by a computing system comprising one or more computing devices, the input with the machine-learned model to generate a model score and a prediction with respect to the label;
updating, by the computing system, one or more parameter values of the machine-learned model based on a gradient associated with the prediction of the machine-learned model with respect to the label, thereby generating an updated machine-learned model;
processing, by the computing system, the input with the updated machine-learned model to generate an updated score;
determining, by the computing system, a difference between the model score and the updated score; and
determining, by the computing system, a measure of an effective number of similar examples that the machine-learned model observed in a training dataset based on the difference between the model score and the updated score;
determining, by the computing system, an uncertainty score for the input relative to each of a number of possible labels; and
generating, by the computing system, a weighted average of the uncertainty score over all of the number of possible labels and
initiating, by the computing system, an action based on the measure determined for each of the plurality of examples.
2. The computer-implemented method of claim 1 , wherein the action comprises an explore or exploit decision.
3. The computer implemented method of claim 1 , wherein the model used to compute the measure is an auxiliary model to a model used to perform inference.
4. The computer-implemented method of claim 1 , wherein the measure of the effective number of similar examples is used to determine a shrinkage of the prediction.
5. The computer-implemented method of claim 1 , wherein the label for each example comprises a ground truth label for the input.
6. The computer-implemented method of claim 1 , wherein the label for each example comprises a binary label.
7. The computer-implemented method of claim 1 , wherein the label for each example comprises a label from a multi-label setting.
8. The computer-implemented method of claim 1 , wherein the model score comprises a dot product or an output of a neural network.
9. The computer-implemented method of claim 1 , wherein the model score comprises a logit score.
10. The computer-implemented method of claim 1 , wherein the prediction comprises a probability prediction for the input relative to the label.
11. The computer-implemented method of claim 1 , wherein the measure comprises one divided by the prediction times an absolute value of the difference between the model score and the updated score.
12. The computer-implemented method of claim 1 , wherein updating, by the computing system, the one or more parameter values of the machine-learned model based on the gradient comprises performing a gradient descent step.
13. The computer-implemented method of claim 1 , further comprising, for each example: reverting the updating of the one or more parameter values of the machine-learned model so as to return to the machine-learned model prior to said updating to perform inference on the score of the ground truth label of an example.
14. The computer-implemented method of claim 1 , wherein the machine-learned model is updated in an online fashion.
15. A computer system configured to perform operations, the operations comprising:
for at least one of a plurality of examples, each example comprising an input and a label:
processing, by a computing system comprising one or more computing devices, the input with the machine-learned model to generate a model score and a prediction with respect to the label;
updating, by the computing system, one or more parameter values of the machine-learned model based on a gradient associated with the prediction of the machine-learned model with respect to the label, thereby generating an updated machine-learned model;
processing, by the computing system, the input with the updated machine-learned model to generate an updated score;
determining, by the computing system, a difference between the model score and the updated score; and
determining, by the computing system, an uncertainty score for the example based on the difference between the model score and the updated score;
determining, by the computing system, the uncertainty score for the input relative to each of a number of possible labels; and
generating, by the computing system, a weighted average of the uncertainty scores over all of the number of possible labels; and
initiating, by the computing system, an action based on the uncertainty score for the at least one of plurality of examples.
16. One or more non-transitory computer-readable media that store instructions that, when executed by one or more processors of a computing system, cause the computing system to perform operations, the operations comprising:
for each of a plurality of examples, each example comprising an input and a label:
processing, by a computing system comprising one or more computing devices, the input with the machine-learned model to generate a model score and a prediction with respect to the label;
updating, by the computing system, one or more parameter values of the machine-learned model based on a gradient associated with the prediction of the machine-learned model with respect to the label, thereby generating an updated machine-learned model;
processing, by the computing system, the input with the updated machine-learned model to generate an updated score;
determining, by the computing system, a difference between the model score and the updated score; and
determining, by the computing system, an uncertainty score for the example based on the difference between the model score and the updated score;
determining, by the computing system, the uncertainty score for the input relative to each of a number of possible labels; and
generating, by the computing system, a weighted average of the uncertainty scores over all of the number of possible labels; and
initiating, by the computing system, an action based on the uncertainty score determined for each of the plurality of examples.
17. The one or more non-transitory computer-readable media of claim 16 , wherein the uncertainty score comprises one minus the prediction times the prediction squared times an absolute value of the difference between the model score and the updated score.
18. The one or more non-transitory computer-readable media of claim 16 , wherein the action comprises an explore or exploit decision.