Balancing diversity and precision of generative models with complementary density estimators

1. A method for training and evaluating a deep generative model with an architecture consisting of two complementary density estimators, comprising:

receiving a probabilistic model of vehicle motion;

training, by a processing device, a first density estimator and a second density estimator jointly based on the probabilistic model of vehicle motion, wherein the first density estimator determines a distribution of outcomes and the second density estimator estimates sample quality;

identifying by the second density estimator spurious modes in the probabilistic model of vehicle motion; and

adjusting the probabilistic model of vehicle motion to eliminate the spurious modes.

2. The method as recited in claim 1 , wherein adjusting the probabilistic model of vehicle motion further comprises:

generating an adjusted model that assigns high probability to plausible paths in a scene while ensuring that every prediction generated by the adjusted model is also likely under a true distribution of future paths.

3. The method as recited in claim 1 , wherein the probabilistic model of vehicle motion includes continuous, high-dimensional, structured data based on paths of intelligent agents.

4. The method as recited in claim 1 , further comprising:

outputting a message to display a warning on an associated display based on the probabilistic model of vehicle motion.

5. The method as recited in claim 1 , further comprising:

training the probabilistic model of vehicle motion to balance dual objectives of generating diverse samples while ensuring high quality samples.

6. The method as recited in claim 1 , further comprising:

determining potential modes of future behavior; and

determining the distribution of outcomes based on the potential modes of future behavior.

7. The method as recited in claim 1 , wherein training the first density estimator and the second density estimator further comprises:

training a network to obtain improved training efficiency, stability, and a balance of diversity and precision in generated examples, while producing an interpretable model.

8. The method as recited in claim 1 , wherein the second estimator is trained by optimizing a variational lower bound based on Fenchel duality.

9. The method as recited in claim 1 , wherein an optimal value of a model associated with the second estimator is independent of the probabilistic model of vehicle motion.

10. A computer system for training and evaluating a deep generative model with an architecture consisting of two complementary density estimators, comprising:

a processor device operatively coupled to a memory device, the processor device being configured to:

receive a probabilistic model of vehicle motion;

train a first density estimator and a second density estimator jointly based on the probabilistic model of vehicle motion, wherein the first density estimator determines a distribution of outcomes and the second density estimator estimates sample quality;

identify by the second density estimator spurious modes in the probabilistic model of vehicle motion; and

adjust the probabilistic model of vehicle motion to eliminate the spurious modes.

11. The system as recited in claim 10 , wherein, when adjusting the probabilistic model of vehicle motion, the processor device is further configured to:

generate an adjusted model that assigns high probability to plausible paths in a scene while ensuring that every prediction generated by the adjusted model is also likely under a true distribution of future paths.

12. The system as recited in claim 10 , wherein the probabilistic model of vehicle motion includes continuous, high-dimensional, structured data based on paths of intelligent agents.

13. The system as recited in claim 10 , wherein the processor device is further configured to:

output a message to display a warning on an associated display based on the probabilistic model of vehicle motion.

14. The system as recited in claim 10 , wherein the processor device is further configured to:

train the probabilistic model of vehicle motion to balance dual objectives of generating diverse samples while ensuring high quality samples.

15. The system as recited in claim 10 , wherein the processor device is further configured to:

determine potential modes of future behavior; and

determine the distribution of outcomes based on the potential modes of future behavior.

16. The system as recited in claim 10 , wherein, when training the first density estimator and the second density estimator, the processor device is further configured to:

train a network to obtain improved training efficiency, stability, and a balance of diversity and precision in generated examples, while producing an interpretable model.

17. The system as recited in claim 10 , wherein the second estimator is trained by optimizing a variational lower bound based on Fenchel duality.

18. The system as recited in claim 10 , wherein an optimal value of a model associated with the second estimator is independent of the probabilistic model of vehicle motion.

19. A computer program product for training and evaluating a deep generative model with an architecture consisting of two complementary density estimators, the computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computing device to cause the computing device to perform the method comprising:

receiving a probabilistic model of vehicle motion;

training, by a processing device, a first density estimator and a second density estimator jointly based on the probabilistic model of vehicle motion, wherein the first density estimator determines a distribution of outcomes and the second density estimator estimates sample quality;

identifying by the second density estimator spurious modes in the probabilistic model of vehicle motion; and

adjusting the probabilistic model of vehicle motion to eliminate the spurious modes.