Deep reinforcement learning framework for sequence level prediction of high dimensional data

1. A method for performing sequence level prediction of a sequence of frames of high dimensional data, comprising:

a) providing one or more affective labels at the end of the sequence and a facial expression embedding from a trained convolutional neural network, wherein each label of the one or more affective labels pertains to the entire sequence of frames of high dimensional data;

b) taking an action with an agent controlled by a machine learning algorithm for a current frame of the sequence at a current time step, wherein an output of the action represents affective label prediction for the frame at the current time step; and

c) transforming a pool of actions taken up until the current time step including the action taken with the agent into a predicted affective history for a subsequent time step; generating a reward on predicted actions up to the current time step, by comparing the predicted actions against corresponding annotated affective labels.

2. The method of claim 1 , wherein one or more ground truth labels are provided only at the end of the sequence of frames of high dimensional data.

3. The method of claim 1 , further comprising repeating a), b), and c) for a subsequent frame of the sequence at the subsequent time step, wherein the subsequent frame becomes the current frame and the subsequent time step becomes the current time step.

4. The method of claim 1 , wherein the affective label prediction at the current time step represents whether a scene is funny or not funny.

5. The method of claim 1 , wherein a value of the reward is based on a majority vote of frame-level affective label prediction results up to the current frame at the current time step.

6. The method of claim 1 , wherein a value of the reward is based on the minimum percentage vote of frame-level results.

7. The method of claim 1 , wherein a value of the reward is zero until a frame for which annotated affective labels are available.

8. The method of claim 1 , wherein the sequence of frames of high dimensional data includes a sequence of faces appearing in a video scene.

9. The method of claim 8 , wherein the sequence of faces is a sequence of faces of an individual character in the video scene.

10. The method of claim 8 , wherein the sequence of faces is a sequence of faces of an individual character in the video scene and wherein when multiple faces are shown in the video scene the sequence of faces is a sequence of faces of located in the center of the scene.

11. A system for performing sequence level prediction of a sequence of frames of high dimensional data, comprising:

a processor;

a memory coupled to the processor; and

processor-executable instructions embodied in the memory, the instructions being configured to cause the processor to perform the following when executed:

a) providing one or more affective labels at the end of the sequence of frames of high dimensional data, wherein each label of the one or more affective labels pertains to the entire sequence of frames of high dimensional data;

b) taking an action with an agent controlled by a machine learning algorithm for a current frame of the sequence at a current time step, wherein an output of the action represents affective label prediction for the frame at the current time step; and

c) transforming a pool of actions taken up until the current time step including the action taken with the agent into a predicted affective history for a subsequent time step; generating a reward on predicted actions up to the current time step, by comparing the predicted actions against corresponding annotated affective labels.

12. The system of claim 11 , wherein the one or more ground truth labels are provided only at the end of the sequence of frames.

13. The system of claim 11 , further comprising repeating a), b), and c) for a subsequent frame of the sequence at the subsequent time step, wherein the subsequent frame becomes the current frame and the subsequent time step becomes the current time step.

14. The system of claim 11 , wherein the affective label prediction at the current time step represents whether a video scene is funny or not funny.

15. The system of claim 11 , wherein a value of the reward is based on a majority vote of frame-level affective label prediction results up to the current frame at the current time step.

16. The system of claim 11 , wherein a value of the reward is based on the minimum percentage vote of frame-level results.

17. The system of claim 11 , wherein a value of the reward is zero until a frame for which annotated scene affective labels are available.

18. The system of claim 11 , wherein the sequence of frames of high dimensional data includes a sequence of faces appearing in a video scene.

19. The system of claim 18 , wherein the sequence of faces is a sequence of faces of an individual character in the video scene.

20. The system of claim 18 , wherein the sequence of faces is a sequence of faces of an individual character in the video scene and wherein when multiple faces are shown in the video scene the sequence of faces is a sequence of faces of located in the center of the scene.

21. A non-transitory computer-readable medium having executable instructions embodied therein, the instructions being configured cause a computer to perform the following when executed:

a) providing one or more affective labels at the end of a sequence of frames of high dimensional data, wherein each affective label of the one or more affective labels pertains to the entire sequence of frames of high dimensional data;

b) taking an action with an agent controlled by a machine learning algorithm for a current frame of the sequence of frames of high dimensional data at a current time step, wherein an output of the action represents affective label prediction for the frame at the current time step; and

c) transforming a pool of actions taken up until the current time step including the action taken with the agent into a predicted affective history for a subsequent time step; generating a reward on predicted actions up to the current time step, by comparing the predicted actions against corresponding annotated affective labels.

22. The non-transitory computer-readable medium of claim 21 , wherein the one or more ground truth labels are provided only at the end of the sequence of frames of high dimensional data.

23. The non-transitory computer-readable medium of claim 21 , further comprising repeating a), b), and c) for a subsequent frame of the sequence at the subsequent time step, wherein the subsequent frame becomes the current frame and the subsequent time step becomes the current time step.

24. The non-transitory computer-readable medium of claim 21 , wherein the affective label prediction at the current time step represents whether a video scene is funny or not funny.

25. The non-transitory computer-readable medium of claim 21 , wherein a value of the reward is based on a majority vote of frame-level affective label prediction results up to the current frame at the current time step.

26. The non-transitory computer-readable medium of claim 21 , wherein a value of the reward is based on the minimum percentage vote of frame-level results.

27. The non-transitory computer-readable medium of claim 21 , wherein a value of the reward is zero until a frame for which annotated affective labels are available.

28. The non-transitory computer-readable medium of claim 21 , wherein the sequence of frames of high dimensional data includes a sequence of faces appearing in a video scene.

29. The non-transitory computer-readable medium of claim 28 , wherein the sequence of faces is a sequence of faces of an individual character in the video scene.

30. The non-transitory computer-readable medium of claim 28 , wherein the sequence of faces is a sequence of faces of an individual character in the video scene and wherein when multiple faces are shown in the video scene the sequence of faces is a sequence of faces of located in the center of the scene.