Active learning via a surrogate machine learning model using knowledge distillation
View Patent ↗Systems and methods of training a model is provided. The system can identify an unlabeled data set with phrases received by a virtual assistant that interfaces with one or more virtual applications to execute one or more functions. The system can query the unlabeled data set to select a first set of phrases based at least on one or more confidence scores output by a surrogate model that corresponds to a third-party model maintained by a third-party system. The system can receive, via a user interface, indications of functions to be executed by the one or more virtual applications responsive to the selected first set of phrases. The system can provide, to the third-party system, the indications of functions for the selected first set of phrases to train the third-party model and configure the virtual assistant to execute a function responsive to a phrase in the first set of phrases.
1. A method of training a model for a Virtual assistant that interfaces with one or more virtual applications hosted on one or more servers, comprising:
identifying, by one or more processors, an unlabeled data set comprising a plurality of phrases received by a virtual assistant that interfaces with one or more virtual applications to execute one or more functions;
querying, by the one or more processors, the unlabeled data set to select a first set of phrases from the plurality of phrases based at least on one or more confidence scores output by a surrogate model that corresponds to a third-party model maintained by a third-party system;
receiving, by the one or more processors via a user interface, indications of functions to be executed by the one or more virtual applications responsive to the selected first set of phrases;
providing, by the one or more processors to the third-party system, the indications of functions for the selected first set of phrases to train the third-party model and configure the virtual assistant to execute a function responsive to a phrase in the first set of phrases;
determining, by the one or more processors, a level of performance of the virtual assistant in causing the one or more virtual applications to execute the one or more functions responsive to input phrases;
selecting, by the one or more processors responsive to the level of performance being less than or equal to a threshold, a second set of phrases from the plurality of phrases in the unlabeled data set to receive second indications of functions for provision to the third-party system to train the third-party model;
determining, by the one or more processors, a change in the level of performance of the virtual assistant in causing the one or more virtual applications to execute the one or more functions responsive to input phrases; and
preventing, by the one or more processors responsive to the change in the level of performance being less than or equal to a second threshold, selection of a subsequent set of phrases from the unlabeled data set to complete a labeling process for the unlabeled data set.
2. The method of claim 1 , comprising:
providing, by the one or more processors, a labeled data set to the third-party system to train the third-party model, the labeled data set comprising phrases configured for input into the virtual assistant and indications of corresponding functions to be executed by the one or more virtual applications.
3. The method of claim 1 , comprising:
training, by the one or more processors, the surrogate model with a labeled data set, the labeled data set comprising phrases configured for input into the virtual assistant and indications of corresponding functions to be executed by the one or more virtual applications.
4. The method of claim 3 , comprising:
inputting, by the one or more processors, the unlabeled data set into the surrogate model trained with the labeled data set to generate the predictions for the unlabeled data set.
5. The method of claim 1 , comprising:
constructing, by the one or more processors, a query to select the first set of phrases based on at least one of an uncertainty sampling technique or a query-by-committee technique.
6. The method of claim 1 , comprising:
providing, by the one or more processors, the first set of phrases selected from the plurality of phrases in the unlabeled data set to the third-party system; and
receiving, by the one or more processors, soft targets for the first of phrases output by the third-party model, the soft targets comprising predictions for functions responsive to the first set of phrases.
7. The method of claim 6 , comprising:
adjusting, by the one or more processors, the soft targets using a smoothing technique comprising at least one of probability clipping, a probability assignment, or a softmax temperature; and
training, by the one or more processors, the surrogate model based on the adjusted soft targets and the indications of functions received Via the user interface.
8. The method of claim 1 , comprising:
providing, by the one or more processors responsive to the selection of the first set of phrases, a prompt via the user interface for the indications of functions responsive to the first set of phrases.
9. A system to train a model for a virtual assistant that interfaces with one or more virtual applications hosted on one or more servers, comprising:
memory and one or more processors to:
identify an unlabeled data set comprising a plurality of phrases received by a virtual assistant that interfaces with one or more virtual applications to execute one or more functions;
query the unlabeled data set to select a first set of phrases from the plurality of phrases based at least on one or more confidence scores output by a surrogate model that corresponds to a third-party model maintained by a third-party system;
receive, via a user interface, indications of functions to be executed by the one or more virtual applications responsive to the selected first set of phrases;
provide, to the third-party system, the indications of functions for the selected first set of phrases to train the third-party model and configure the virtual assistant to execute a function responsive to a phrase in the first set of phrases;
determine a level of performance of the virtual assistant in causing the one or more virtual applications to execute the one or more functions responsive to input phrases;
select, in response to the level of performance being less than or equal to a threshold, a second set of phrases from the plurality of phrases in the unlabeled data set to receive second indications of functions for provision to the third-party system to train the third-party model;
determine a change in the level of performance of the virtual assistant in causing the one or more virtual applications to execute the one or more functions responsive to input phrases; and
prevent, in response to the change in the level of performance being less than or equal to a second threshold, selection of a subsequent set of phrases from the unlabeled data set to complete a labeling process for the unlabeled data set.
10. The system of claim 9 , wherein the one or more processors are further configured to:
provide a labeled data set to the third-party system to train the third-party model, the labeled data set comprising phrases configured for input into the virtual assistant and indications of corresponding functions to be executed by the one or more virtual applications.
11. The system of claim 9 , wherein the one or more processors are further configured to:
train the surrogate model with a labeled data set, the labeled data set comprising phrases configured for input into the virtual assistant and indications of corresponding functions to be executed by the one or more virtual applications.
12. The system of claim 11 , wherein the one or more processors are further configured to:
input the unlabeled data set into the surrogate model trained with the labeled data set to generate the predictions for the unlabeled data set.
13. The system of claim 9 , wherein the one or more processors are further configured to:
construct a query to select the first set of phrases based on at least one of an uncertainty sampling technique or a query-by-committee technique.
14. The system of claim 9 , wherein the one or more processors are further configured to:
provide the first set of phrases selected from the plurality of phrases in the unlabeled data set to the third-party system; and
receive soft targets for the first of phrases output by the third-party model, the soft targets comprising predictions for functions responsive to the first set of phrases.
15. The system of claim 14 , wherein the one or more processors are further configured to:
adjust the soft targets using a smoothing technique comprising at least one of probability clipping, a probability assignment, or a softmax temperature; and
train the surrogate model based on the adjusted soft targets and the indications of functions received via the user interface.
16. The system of claim 9 , wherein the one or more processors are further configured to:
provide, responsive to the selection of the first set of phrases, a prompt via the user interface for the indications of functions responsive to the first set of phrases.
17. A non-transitory computer-readable medium comprising instructions that, when executed by one or more processors, train a model for a Virtual assistant that interfaces with one or more virtual applications hosted on one or more sewers, wherein the instructions comprise instructions to:
identify an unlabeled data set comprising a plurality of phrases received by a virtual assistant that interfaces with one or more Virtual applications to execute one or more functions;
query the unlabeled data set to select a first set of phrases from the plurality of phrases based at least on one or more confidence scores output by a surrogate model that corresponds to a third-party model maintained by a third-party system;
receive, via a user interface, indications of functions to be executed by the one or more virtual applications responsive to the selected first set of phrases;
provide, to the third-party system, the indications of functions for the selected first set of phrases to train the third-party model and configure the virtual assistant to execute a function responsive to a phrase in the first set of phrases;
determine a level of performance of the virtual assistant in causing the one or more virtual applications to execute the one or more functions responsive to input phrases;
selecting, in response to the level of performance being less than or equal to a threshold, a second set of phrases from the plurality of phrases in the unlabeled data set to receive second indications of functions for provision to the third-party system to train the third-party model;
determine a change in the level of performance of the virtual assistant in causing the one or more virtual applications to execute the one or more functions responsive to input phrases; and
prevent, in response to the change in the level of performance being less than or equal to a second threshold, selection of a subsequent set of phrases from the unlabeled data set to complete a labeling process for the unlabeled data set.
18. The computer-readable medium of claim 17 , wherein the instructions further comprise instructions to:
provide a labeled data set to the third-party system to train the third-party model, the labeled data set comprising phrases configured for input into the Virtual assistant and indications of corresponding functions to be executed by the one or more Virtual applications.