阅读说明：本技术 用于估计由监督机器学习器做出的决策的不确定性的系统和方法 (System and method for estimating uncertainty of decisions made by a supervised machine learner ) 是由 S·科洛瑞 H·霍夫曼 于 2019-01-30 设计创作，主要内容包括：描述了一种用于对自主平台进行控制的系统。基于输入图像,系统生成针对自主平台的电动机控制命令决策。确定输入图像属于训练图像集合的概率,并且使用所确定的概率生成针对电动机控制命令决策的可靠性度量。当可靠性度量高于预定阈值时,执行探索动作。否则,当可靠性度量低于预定阈值时,执行与电动机控制命令决策相对应的利用动作。(A system for controlling an autonomous platform is described. Based on the input images, the system generates motor control command decisions for the autonomous platform. A probability that the input image belongs to the set of training images is determined, and a reliability metric for the motor control command decision is generated using the determined probability. When the reliability measure is above a predetermined threshold, a heuristic action is performed. Otherwise, when the reliability metric is below a predetermined threshold, a utilization action corresponding to the motor control command decision is performed.)

1. A system for controlling an autonomous platform, the system comprising:

a non-transitory computer-readable medium encoded with executable instructions and one or more processors such that, when executed, the one or more processors perform operations comprising:

generating a motor control command decision for the autonomous platform based on an input image;

determining a probability that the input image belongs to a set of training images;

generating a reliability metric for the motor control command decision using the determined probability; and

performing a heuristic action when the reliability measure is above a predetermined threshold; otherwise

Performing a utilization action corresponding to the motor control command decision when the reliability metric is below a predetermined threshold.

2. The system of claim 1, wherein the one or more processors further perform an operation of learning an embedding for a distribution of the input data in determining a probability that the input image belongs to a class of images that was seen.

3. The system of claim 2, wherein a slice Wtherstein clustering technique is used in learning the embeddings for the distribution of the input data.

4. The system of claim 1, wherein the autonomous platform collects additional input data in an area of an environment surrounding the autonomous platform while performing the exploratory action.

5. The system of claim 1, wherein in performing the utilization action, the autonomous platform executes motor control commands that result in a pre-specified target.

6. The system of claim 1, wherein the one or more processors further perform the following:

training a decision module with the set of training images via supervised learning or reinforcement learning, wherein the decision module generates the motor control command decision; and

training an uncertainty module with the set of training images in an unsupervised manner, wherein the uncertainty module determines the probability that the input image belongs to the set of training images, and

wherein the uncertainty module is trained in parallel with the decision module.

7. The system of claim 1, wherein in determining the probability that the input image belongs to the set of training images, the one or more processors further perform the following:

embedding the input image into a low-dimensional Hilbert space using an deconvolution auto-encoder; and

modeling a distribution of the input data in the low-dimensional Hilbert space via a Gaussian mixture model.

8. A computer-implemented method of controlling an autonomous platform, the computer-implemented method comprising acts of:

causing one or more processors to execute instructions encoded on a non-transitory computer-readable medium such that, when executed, the one or more processors perform the following:

generating a motor control command decision for the autonomous platform based on an input image;

determining a probability that the input image belongs to a set of training images;

generating a reliability metric for the motor control command decision using the determined probability; and

performing a heuristic action when the reliability measure is above a predetermined threshold; otherwise

Performing a utilization action corresponding to the motor control command decision when the reliability metric is below a predetermined threshold.

9. The method of claim 8, wherein in determining the probability that the input image belongs to a class of images that were seen, the one or more processors further perform an operation of learning an embedding for a distribution of the input data.

10. The method of claim 9, wherein a slice-wotherstein clustering technique is used in learning the embeddings for the distribution of the input data.

11. The method of claim 8, wherein the autonomous platform collects additional input data in an area of an environment surrounding the autonomous platform while performing the exploratory action.

12. The method of claim 8, wherein in performing the utilization action, the autonomous platform executes motor control commands that result in a pre-specified target.

13. The method of claim 8, wherein the one or more processors further perform the following:

training a decision module with the set of training images via supervised learning or reinforcement learning, wherein the decision module generates the motor control command decision; and

training an uncertainty module with the set of training images in an unsupervised manner, wherein the uncertainty module determines the probability that the input image belongs to the set of training images, and

wherein the uncertainty module is trained in parallel with the decision module.

14. The method of claim 8, wherein, in determining the probability that the input image belongs to the set of training images, the one or more processors further perform the following:

embedding the input image into a low-dimensional Hilbert space using an deconvolution auto-encoder; and

modeling a distribution of the input data in the low-dimensional Hilbert space via a Gaussian mixture model.

15. A computer program product for controlling an autonomous platform, the computer program product comprising:

computer-readable instructions stored on a non-transitory computer-readable medium, the computer-readable instructions executable by a computer having one or more processors to cause the processors to:

generating a motor control command decision for the autonomous platform based on an input image;

determining a probability that the input image belongs to a set of training images;

generating a reliability metric for the motor control command decision using the determined probability; and

performing a heuristic action when the reliability measure is above a predetermined threshold; otherwise

Performing a utilization action corresponding to the motor control command decision when the reliability metric is below a predetermined threshold.

16. The computer program product of claim 15, wherein the one or more processors further perform an operation of learning an embedding for a distribution of the input data in determining a probability that the input image belongs to the class of viewed images.

17. The computer program product of claim 16, wherein a slice-wotherstein clustering technique is used in learning the embedding for the distribution of the input data.

18. The computer program product of claim 15, wherein the autonomous platform collects additional input data in an area of an environment surrounding the autonomous platform while performing the exploring.

19. The computer program product of claim 15, wherein in performing the utilization action, the autonomous platform executes motor control commands that result in a pre-specified goal.

20. The computer program product of claim 15, wherein the one or more processors further perform the following:

training a decision module with the set of training images via supervised learning or reinforcement learning, wherein the decision module generates the motor control command decision; and

training an uncertainty module with the set of training images in an unsupervised manner, wherein the uncertainty module determines the probability that the input image belongs to the set of training images, and

wherein the uncertainty module is trained in parallel with the decision module.

21. The computer program product of claim 15, wherein in determining the probability that the input image belongs to the set of training images, the one or more processors further perform the following:

embedding the input image into a low-dimensional Hilbert space using an deconvolution auto-encoder; and

modeling a distribution of the input data in the low-dimensional Hilbert space via a Gaussian mixture model.

22. The system of claim 1, wherein the autonomous platform comprises a robotic arm, and wherein the exploratory action is a new way to attempt to grasp an object with the robotic arm.

23. The system of claim 1, wherein the autonomous platform comprises a robotic arm, and wherein the utilization action is a known action robotic arm comprising grasping and processing an object with the robotic arm.

24. The system of claim 1, wherein the autonomous platform is a vehicle, and wherein the leveraging action is collision avoidance.