How about a smart model that knows what it does not know?
Selective classification for speech emotion recognition (Reject options)
-
First study in speech emotion recognition aimed at developing a reliable system via uncertainty modeling through reject options
-
Reject Options: Abstaining from classification when in doubt to improve the reliability of a emotion recognition model
-
Utilized Bayesian empirical risk minimization framework, probabilistic backpropogation technique to model uncertainty in predictions and illustrated using a reject options framework
-
Achievements: Classification - Up to 26% gains in F1-Score with 75% test coverage
-
Rejected samples showed lower inter-evaluator agreement compared to accepted samples, validating the success of this idea in speech emotion recognition
[code]