Autonomous vehicle system for intelligent on-board selection of data for building a remote machine learning model
Systems and methods for on-board selection of data logs for training a machine learning model are provided. The system includes an autonomous vehicle having a plurality of sensors and a processor. The processor receives a plurality of unlabeled images from the plurality of sensors, a machine learning model, and a loss function corresponding to the machine learning model. For each of the plurality of images, the processor then determines one or more predictions using the machine learning model, compute an importance function based on the loss function and the one or more predictions, and transmit that image to a remote server for updating the machine learning model when a value of the importance function is greater than a threshold.
1. A method for active selection of data for training a machine learning model, comprising, by an on-board computing device of an autonomous vehicle:
receiving, from a plurality of sensors of the autonomous vehicle, a plurality of unlabeled data logs;
receiving, from a remote server:
the machine learning model, and
a loss function corresponding to the machine learning model; and
for each of the plurality of unlabeled data logs:
determining, using the machine learning model, one or more predictions for that data log,
computing, based on the loss function and the one or more predictions for that data log, an importance function for that data log, and
transmitting that data log to the remote server for updating the machine learning model when a value of the importance function is greater than a threshold.
2. The method of claim 1 , wherein the importance function is a gradient based importance function computed with respect to a parameter of the machine learning model.
3. The method of claim 2 , wherein the machine learning model utilizes a differentiable loss function.
4. The method of claim 2 , further comprising determining the loss function with respect to the parameter of the machine learning model based on all previously collected training data that is used to build the machine learning model.
5. The method of claim 4 , further comprising computing, based on the loss function and the one or more predictions for that data log, the importance function by taking a derivative of the loss function with respect to the parameter of the machine learning model for that data log.
6. The method of claim 5 , further comprising determining the threshold such that use of a data log, having the value of the importance function that is greater than the threshold, for updating the machine learning model, causes the parameter of the machine learning model to be updated in a direction of a gradient for the parameter.
7. The method of claim 6 , wherein the threshold is determined empirically.
8. The method of claim 5 , wherein the value of the importance function is determined based on a confidence level associated with an output of the one or more predictions for that data log.
9. The method of claim 1 , further comprising discarding that data log when the value of the importance function is less than the threshold.
10. The method of claim 1 , wherein the machine learning model is an object detection model and that data log comprises an image.
11. The method of claim 10 , wherein determining the one or more predictions comprises using the object detection model to:
detect an object in the image; and
determine a bounding box and label corresponding to the detected object.
12. A system for on-board selection of data logs for training a machine learning model comprising:
an autonomous vehicle comprising:
a plurality of sensors,
a processor, and
a non-transitory computer-readable medium comprising one or more programming instructions that when executed by the processor, will cause the processor to:
receive, from the plurality of sensors, a plurality of unlabeled data logs,
receive, from a remote server:
the machine learning model, and
a loss function corresponding to the machine learning model, and for each of the plurality of unlabeled data logs:
determine, using the machine learning model, one or more predictions for that data log,
compute, based on the loss function and the one or more predictions for that data log, an importance function for that data log, and
transmit that data log to the remote server for updating the machine learning model when a value of the importance function is greater than a threshold.
13. The system of claim 12 , wherein the importance function is a gradient based importance function computed with respect to a parameter of the machine learning model.
14. The system of claim 13 , wherein the machine learning model utilizes a differentiable loss function.
15. The system of claim 13 , further comprising programming instructions that when executed by the processor, will cause the processor to determine the loss function with respect to the parameter of the machine learning model based on all previously collected training data that is used to build the machine learning model.
16. The system of claim 15 , further comprising programming instructions that when executed by the processor, will cause the processor to compute, based on the loss function and the one or more predictions for that data log, the importance function by taking a derivative of the loss function with respect to the parameter of the machine learning model for that data log.
17. The system of claim 16 , further comprising programming instructions that when executed by the processor, will cause the processor to determine the threshold such that use of the data log, having the value of the importance function that is greater than the threshold, for updating the machine learning model, causes the parameter of the machine learning model to be updated in a direction of a gradient for the parameter.
18. The system of claim 17 , wherein the threshold is determined empirically.
19. The system of claim 16 , wherein the value of the importance function is determined based on a confidence level associated with an output of the one or more predictions for that data log.
20. The system of claim 12 , further comprising programming instructions that when executed by the processor, will cause the processor to discard that data log when the value of the importance function is less than the threshold.
21. The system of claim 12 , wherein the machine learning model is an object detection model and that data log comprises an image.
22. The system of claim 21 , wherein the programming instructions that when executed by the processor, will cause the processor to determine the one or more predictions comprise programming instructions that when executed by the processor, will cause the processor to use the object detection model to:
detect an object in the image; and
determine a bounding box and label corresponding to the detected object.