Uncertainty prediction based deep learning
According to one aspect, uncertainty prediction based deep learning may include receiving, using a memory, a trained neural network policy π trained based on a first dataset in a first environment, implementing, via a controller, the trained neural network policy π in a second environment by receiving an input and generating an output y, calculating an uncertainty array U[T] for a time window T, wherein the uncertainty array is indicative of a level of uncertainty associated with an output sample distribution of the output across the time window T based on a temporal divergence, an entropy H, a variational ratio VR, and a standard deviation SD of the output y, and executing, via the controller and one or more systems, an action based on the uncertainty array U[T], such as discontinuing use of the trained neural network policy π.
1. An uncertainty prediction based deep learning controller, comprising:
a memory receiving a trained neural network policy π trained based on a first dataset in a first environment;
a controller implementing the trained neural network policy π in a second environment by receiving an input and generating an output y; and
an uncertainty evaluator, implemented via a processor, calculating an uncertainty array U[T] for a time window T, wherein the uncertainty array is indicative of a level of uncertainty associated with an output sample distribution of the output across the time window T, wherein the uncertainty evaluator calculates the uncertainty array U[T] based on a temporal divergence of the output y, an entropy H of the output y, a variational ratio VR of the output y, and a standard deviation SD of the output y,
wherein the controller executes an action based on the uncertainty array U[T],
wherein the action is a disengage autonomous driving action including disabling an autonomous driving module of an autonomous vehicle based on an element of the uncertainty array U[T] exceeding an uncertainty threshold.
2. The uncertainty prediction based deep learning controller of claim 1 , wherein the first dataset from the first environment is generated by a human driver.
3. The uncertainty prediction based deep learning controller of claim 1 , wherein the controller implements the trained neural network policy π in an autonomous vehicle and the second environment is a real-world environment.
4. The uncertainty prediction based deep learning controller of claim 3 , wherein the first environment is a simulated environment.
5. The uncertainty prediction based deep learning controller of claim 1 , wherein the action is a notification action including providing a notification to an occupant of an autonomous vehicle where the trained neural network policy π is implemented.
6. The uncertainty prediction based deep learning controller of claim 1 , wherein the controller executes a second action based on the element of the uncertainty array U[T] exceeding the uncertainty threshold, the second action including updating the trained neural network policy π.
7. The uncertainty prediction based deep learning controller of claim 6 , wherein the second action of updating the trained neural network policy π includes:
receiving image capture data from an image capture device of the autonomous vehicle;
receiving one or more steering inputs from one or more vehicle systems; and
generating the updated trained neural network policy π based on the image capture data and one or more of the steering inputs.
8. The uncertainty prediction based deep learning controller of claim 1 , wherein the uncertainty evaluator calculates the uncertainty array U[T] by multiplying the temporal divergence of the output y, the entropy H of the output y, the variational ratio VR of the output y, and the standard deviation SD of the output y.
9. The uncertainty prediction based deep learning controller of claim 8 , wherein the controller applies a quadratic filter to the uncertainty array U[T].
10. A computer-implemented method for uncertainty prediction based deep learning, comprising:
receiving, via a processor, a trained neural network policy π trained based on a first dataset in a first environment;
implementing, via the processor, the trained neural network policy π in a second environment by receiving an input and generating an output y;
calculating, via the processor, an uncertainty array U[T] for a time window T, wherein the uncertainty array is indicative of a level of uncertainty associated with an output sample distribution of the output across the time window T, wherein the calculating the uncertainty array U[T] is based on a temporal divergence of the output y, an entropy H of the output y, a variational ratio VR of the output y, and a standard deviation SD of the output y; and
executing, via the processor, an action based on the uncertainty array U[T],
wherein the action is a disengage autonomous driving action including disabling an autonomous driving module of an autonomous vehicle based on an element of the uncertainty array U[T] exceeding an uncertainty threshold.
11. The computer-implemented method for uncertainty prediction based deep learning of claim 10 , comprising implementing the trained neural network policy π in an autonomous vehicle and the second environment is a real-world environment.
12. The computer-implemented method for uncertainty prediction based deep learning of claim 11 , wherein the first environment is a simulated environment.
13. The computer-implemented method for uncertainty prediction based deep learning of claim 10 , wherein the action is a notification action including providing a notification to an occupant of an autonomous vehicle where the trained neural network policy π is implemented.
14. The computer-implemented method for uncertainty prediction based deep learning of claim 10 , comprising executing a second action based on the element of the uncertainty array U[T] exceeding the uncertainty threshold, the second action including updating the trained neural network policy π.
15. The computer-implemented method for uncertainty prediction based deep learning of claim 14 , wherein the second action of updating the trained neural network policy π includes:
receiving image capture data from an image capture device of the autonomous vehicle;
receiving one or more steering inputs from one or more vehicle systems; and
generating the updated trained neural network policy π based on the image capture data and one or more of the steering inputs.
16. The computer-implemented method for uncertainty prediction based deep learning of claim 10 , comprising calculating the uncertainty array U[T] by multiplying the temporal divergence of the output y, the entropy H of the output y, the variational ratio VR of the output y, and the standard deviation SD of the output y.
17. The computer-implemented method for uncertainty prediction based deep learning of claim 10 , comprising applying a quadratic filter to the uncertainty array U[T].
18. A system for uncertainty prediction based deep learning, comprising:
a memory receiving a trained neural network policy π trained based on a first dataset in a first environment;
a controller implementing the trained neural network policy π within an autonomous vehicle travelling through a second, real-world environment by receiving an input and generating an output y via an autonomous driving module; and
an uncertainty evaluator, implemented via a processor, calculating an uncertainty array U[T] for a time window T, wherein the uncertainty array is indicative of a level of uncertainty associated with an output sample distribution of the output across the time window T, wherein the uncertainty evaluator calculates the uncertainty array U[T] based on a temporal divergence of the output y, an entropy H of the output y, a variational ratio VR of the output y, and a standard deviation SD of the output y,
wherein the controller executes an action based on the uncertainty array U[T], wherein the action is a disengage autonomous driving action including disabling the autonomous driving module based on an element of the uncertainty array U[T] exceeding an uncertainty threshold.