Systems and methods for tuning hyperparameters of a model and advanced curtailment of a training of the model

Patent metadata

Field	Value
Publication number	US-12450479-B2
Application number	US-202117508665-A
Country	US
Kind code	B2
Filing date	Oct 22, 2021
Priority date	Apr 15, 2019
Publication date	Oct 21, 2025
Grant date	Oct 21, 2025

Abstract

A system and method for tuning hyperparameters and training a model includes implementing a hyperparameter tuning service that tunes hyperparameters of a model that includes receiving, via an API, a tuning request that includes: (i) a first part comprising tuning parameters for generating tuned hyperparameter values for hyperparameters of the model; and (ii) a second part comprising model training control parameters for monitoring and controlling a training of the model, wherein the model training control parameters include criteria for generating instructions for curtailing a training run of the model; monitoring the training run for training the model based on the second part of the tuning request, wherein the monitoring of the training run includes periodically collecting training run data; and computing an advanced training curtailment instruction based on the training run data that automatically curtails the training run prior to a predefined maximum training schedule of the training run.

First claim

We claim:

1. An apparatus comprising:
at least one memory;
instructions in the apparatus; and
processor circuitry to execute the instructions to:
compute an average metric value for a plurality of first training runs;
compute a metric value for a second training run at an evaluation interval;
evaluate the metric value for the evaluation interval of the second training run relative to the average metric value based on a metric goal; and
stop the second training run based on an early termination policy and the evaluation of the metric value relative to the average metric value.

2. The apparatus of claim 1 , wherein the metric value corresponds to a performance metric for tuning a hyperparameter of a model.

3. The apparatus of claim 2 , wherein the processor circuitry is to execute the instructions to define the metric goal, the metric goal to optimize the performance metric of the model based on the plurality of first training runs.

4. The apparatus of claim 2 , wherein the metric value corresponds to an accuracy metric of the model and the metric goal is to maximize accuracy of the model, the processor to execute the instructions to stop the second training run in response to the metric value being less than the average metric value.

5. The apparatus of claim 1 , wherein the evaluation interval defines a frequency of applying the early termination policy, the processor circuitry to execute the instructions to apply the early termination policy based on the evaluation interval.

6. The apparatus of claim 1 , wherein the processor circuitry is to execute the instructions to select an evaluation delay, the evaluation delay to cause a minimum number of evaluation intervals to complete before applying the early termination policy at the evaluation interval.

7. The apparatus of claim 1 , wherein the processor circuitry is to execute the instructions to use the early termination policy to stop the second training run based on the second training run being a low-performance run.

8. The apparatus of claim 1 , wherein the processor circuitry is to execute the instructions to stop the second training run based on the early termination policy prior to completion of the second training run.

9. At least one non-transitory computer readable medium comprising instructions that, when executed, cause at least one processor to at least:
compute an average metric value for a plurality of first training runs;
compute a metric value for a second training run at an evaluation interval;
evaluate the metric value for the evaluation interval of the second training run relative to the average metric value based on a metric goal; and
stop the second training run based on an early termination policy and the evaluation of the metric value relative to the average metric value.

10. The at least one non-transitory computer readable medium of claim 9 , wherein the metric value and the average metric value are based on a performance metric corresponding to a tuning of a hyperparameter of a model.

11. The at least one non-transitory computer readable medium of claim 10 , wherein the instructions cause the at least one processor to define the metric goal, the metric goal to optimize the performance metric of the model based on the plurality of first training runs.

12. The at least one non-transitory computer readable medium of claim 10 , wherein the metric value corresponds to an accuracy metric of the model and the metric goal is to maximize accuracy of the model, the instructions to cause the at least one processor to stop the second training run in response to the metric value being less than the average metric value.

13. The at least one non-transitory computer readable medium of claim 9 , wherein the evaluation interval defines a frequency of applying the early termination policy, the instructions to cause the at least one processor to apply the early termination policy based on the evaluation interval.

14. The at least one non-transitory computer readable medium of claim 9 , wherein the instructions cause the at least one processor to select an evaluation delay, the evaluation delay to cause a minimum number of evaluation intervals to complete before applying the early termination policy at the evaluation interval.

15. The at least one non-transitory computer readable medium of claim 9 , wherein the instructions cause the at least one processor to use the early termination policy to stop the second training run based on the second training run being a low-performance run.

16. The at least one non-transitory computer readable medium of claim 9 , wherein the instructions cause the at least one processor to stop the second training run based on the early termination policy prior to completion of the second training run.

17. A method to tune a hyperparameter of a model, the method comprising:
computing an average metric value for a plurality of first training runs;
computing a metric value for a second training run at an evaluation interval;
evaluating the metric value for the evaluation interval of the second training run relative to the average metric value based on a metric goal; and
stopping the second training run based on an early termination policy and the evaluation of the metric value relative to the average metric value.

18. The method of claim 17 , further including defining the metric goal, the metric goal to optimize a metric of the model based on the plurality of first training runs, the metric corresponding to a performance of the model.

19. The method of claim 17 , wherein the metric value corresponds to an accuracy metric of a model, the metric goal to maximize accuracy of the model, the method further including stopping the second training run in response to the metric value being less than the average metric value.

20. The method of claim 17 , wherein the evaluation interval defines a frequency of applying the early termination policy, the method further including applying the early termination policy based on the evaluation interval.

21. The method of claim 17 , wherein the evaluation interval is a first evaluation interval, the method further including selecting an evaluation delay, the evaluation delay to cause a minimum number of evaluation intervals to complete before the first evaluation interval.

22. The method of claim 17 , further including stopping the second training run based on the early termination policy prior to completion of the second training run.

Assignees

• Intel Corp

Inventors

• Mccourt Michael
• Jackle-Spriggs Taylor
• Hsu Ben
• Howey Simon
• Vance Halley Nicki
• Blomo James
• Hayes Patrick
• Clark Scott

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06N3/0985

  Hyperparameter optimisation; Meta-learning; Learning-to-learn · CPC title

• G06N3/092

  Reinforcement learning · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title