Emotions are enmeshed in the neural networks of uncertainty
A smart Speech emotion recognition system that can tell how confident it is about its predictions
-
Selective prediction models can be most useful in real-world applications. A model with smartness to abstain from prediction when in doubt can facilitate human-in-the-loop solutions
-
Utilized Monte Carlo approximation to variational inference to implement create a smart emotion recognition system and demonstrated its benefits using reject options.
-
Achievements: Regression - relative gains in concordance correlation coefficient up to 7.34% for arousal, 13.73% for valence and 8.79% for dominance recognition
-
Modeling uncertainty resulted in prediction uncertainties being a function of emotional attribute scores where extreme emotions are predicted with higher confidence compared to neutral emotion
-
Developed sample selection strategies such as balanced and global selection criteria, using learned prediction uncertainties to further this research in areas of active learning, co-training, curriculum learning and multi-view training
[Video]
[code]