Online learning by an instance of a deep learning model and sharing of learning with additional instances of the deep learning model

1. A computer-implemented method for enabling sharing of knowledge learned by an autonomous vehicle to a fleet of autonomous vehicles, the method comprising:

generating an update for a continuous deep learning neural network on-board the autonomous vehicle based on driving scenarios encountered by the autonomous vehicle during its deployment;

providing the update for the continuous deep learning neural network to additional vehicles in the fleet on autonomous vehicles, wherein the update for the continuous deep learning neural network is configured to be incorporated into a joint kernel for use by the additional vehicles in the fleet on autonomous vehicles; and

causing the autonomous vehicle to drive with the updated kernel applied; and

determining a learning budget wherein the learning budget is a maximum total change to the deep-learning neural network and calculating a learning limit by evenly distributing the learning budget across a me range, using the learning limit and setting the constant learning limit to zero when the learning budget is reached, wherein generating update occurs after reaching the learning limit.

2. The computer-implemented method of claim 1 , further comprising: applying the update for the continuous deep learning neural network to a kernel of the deep learning neural network on the autonomous vehicle to generate an updated kernel on the autonomous vehicle; and

performing, by the autonomous vehicle, a safety self-test on the updated kernel.

3. The computer-implemented method of claim 2 , wherein the safety self-test includes checks that determine accuracy of the updated kernel.

4. The computer-implemented method of claim 1 , wherein the continuous deep learning neural network is used in at least one of the perception stack, prediction stack, and planning stack.

5. The computer-implemented method of claim 1 , further comprising:

receiving an update over a network connection for the kernel of the continuous deep learning neural network from another autonomous vehicle in the fleet of autonomous vehicles;

applying the update received over the network connection to the kernel of the continuous deep learning neural network to generate a second updated kernel; and

performing, by the autonomous vehicle, a safety self-test on the second updated kernel.

6. The computer-implemented method of claim 1 , wherein the learning limit is based on at least one of a fleet size, a number of autonomous vehicles operating, a constant value, a threshold value, a type of deep-learning model, and a value determined by simulating a safety test and wherein the learning limit limits how much a single autonomous vehicle changes the deep learning neural network on-board the autonomous vehicle.

7. The computer-implemented method of claim 6 , wherein the learning budget is based on a number of autonomous vehicles in the fleet of autonomous vehicles.

8. The computer-implemented method of claim 1 , further comprising:

receiving an update over a network connection for the kernel from a first other autonomous vehicle in the fleet of autonomous vehicles, wherein the update received over the network connection is generated by the first other autonomous vehicle in the fleet of autonomous vehicles;

receiving an additional update over the network connection for the kernel from a second other autonomous vehicle in the fleet of autonomous vehicles, wherein the additional update received over a network connection is generated by the second other autonomous vehicle in the fleet of autonomous vehicles; and

applying the update received over the network connection from the first other autonomous vehicle and the additional update received over a network connection from the second other autonomous vehicle to the kernel of the continuous deep learning neural network to generate a second updated kernel.

9. The computer-implemented method of claim 8 , wherein applying the update comprises one of averaging the update and the additional update or using knowledge distillation on the update and the additional update.

10. A system comprising:

a storage configured to store instructions;

a processor configured to execute the instructions and cause the processor to:

generate an update for a continuous deep learning neural network on-board the autonomous vehicle based on driving scenarios encountered by the autonomous vehicle during its deployment,

provide the update for the continuous deep learning neural network to additional vehicles in the fleet on autonomous vehicles, wherein the update for the continuous deep learning neural network is configured to be incorporated into a joint kernel for use by the additional vehicles in the fleet on autonomous vehicles; and

cause the autonomous vehicle to drive with the updated kernel applied; and

determining a learning budget, wherein the learning budget is a maximum total change to the deep-learning neural network and calculating a learning limit by evenly distributing the learning budget across a time range, using the learning limit and setting the constant learning limit to zero when the earning budget is reached, wherein generating the update occurs after reaching the learning limit.

11. The system of claim 10 , wherein the processor is configured to execute the instructions and cause the processor to:

apply the update for the continuous deep learning neural network to a kernel of the deep learning neural network on the autonomous vehicle to generate an updated kernel on the autonomous vehicle; and

perform, by the autonomous vehicle, a safety self-test on the updated kernel.

12. The system of claim 10 , wherein the continuous deep learning neural network is used in at least one of the perception stack, prediction stack, and planning stack.

13. The system of claim 10 , wherein the processor is configured to execute the instructions and cause the processor to:

receive an update over a network connection for the kernel of the continuous deep learning neural network from another autonomous vehicle in the fleet of autonomous vehicles;

apply the update received over the network connection to the kernel of the continuous deep learning neural network to generate a second updated kernel; and

perform, by the autonomous vehicle, a safety self-test on the second updated kernel.

14. The system of claim 10 , wherein e learning limit is based on at least one of a fleet size, a number of autonomous vehicles operating, a constant value, a threshold value, a type deep-learning model, and a value determined by simulating a safety test, and wherein the learning limit limits how much a single autonomous vehicle changes the deep learning neural network on-board the autonomous vehicle.

15. The system of claim 10 , wherein the processor is configured to execute the instructions and cause the processor to:

receive an update over a network connection for the kernel from a first other autonomous vehicle in the fleet of autonomous vehicles, wherein the update received over the network connection is generated by the first other autonomous vehicle in the fleet of autonomous vehicles;

receive an additional update over the network connection for the kernel from a second other autonomous vehicle in the fleet of autonomous vehicles, wherein the additional update received over a network connection is generated by the second other autonomous vehicle in the fleet of autonomous vehicles; and

apply the update received over the network connection from the first other autonomous vehicle and the additional update received over a network connection from the second other autonomous vehicle to the kernel of the continuous deep learning neural network to generate a second updated kernel.

16. A non-transitory computer readable medium comprising instructions, the instructions, when executed by a computing system, cause the computing system to:

generate an update for a continuous deep learning neural network on-board the autonomous vehicle based on driving scenarios encountered by the autonomous vehicle during its deployment;

provide the update for the continuous deep learning neural network to additional vehicles in the fleet on autonomous vehicles, wherein the update for the continuous deep learning neural network is configured to be incorporated into a joint kernel for use by the additional vehicles in the fleet on autonomous vehicles; and

cause the autonomous vehicle to drive with the updated kernel applied, and

determining learning budget, wherein the learning budget is a maximum total change to the deep-learning neural network and calculating learning limit by evenly distributing the earning budget across a time range, using the learning limit and setting the constant learning limit to zero when the learning budget is reached, wherein generating the update occurs after reaching the learning limit.

17. The computer readable medium of claim 16 , wherein the computer readable medium further comprises instructions that, when executed by the computing system, cause the computing system to:

apply the update for the continuous deep learning neural network to a kernel of the deep learning neural network on the autonomous vehicle to generate an updated kernel on the autonomous vehicle; and

perform, by the autonomous vehicle, a safety self-test on the updated kernel.

18. The computer readable medium of claim 16 , the continuous deep learning neural network is used in at least one of the perception stack, prediction stack, and planning stack.

19. The computer readable medium of claim 16 , wherein the computer readable medium further comprises instructions that, when executed by the computing system, cause the computing system to:

receive an update over a network connection for the kernel of the continuous deep learning neural network from another autonomous vehicle in the fleet of autonomous vehicles;

apply the update received over the network connection to the kernel of the continuous deep learning neural network to generate a second updated kernel; and perform, by the autonomous vehicle, a safety self-test on the second updated kernel.

20. The computer readable medium of claim 16 , wherein the learning limit is based on at least one of a fleet size, a number of autonomous vehicles operating, a constant value, a threshold value, a type of deep-learning model, and a value determined by simulating a safety test and wherein the learning limit limits how much a single autonomous vehicle changes the deep learning neural network on-board the autonomous vehicle.