Abstract
Featured Application: The proposed Uncertainty Flow framework may benefit the facial analysis with its promised elevation in discriminability in multi-label affective classification tasks. Moreover, this framework also allows the efficient model training and between tasks knowledge transfer. The applications that rely heavily on continuous prediction on emotional valance, e.g., to monitor prisoners’ emotional stability in jail, can be directly benefited from our framework. Abstract: To lower the single-label dependency on affective facial analysis, it urges the fruition of multi-label affective learning. The impediment to practical implementation of existing multi-label algorithms pertains to scarcity of scalable multi-label training datasets. To resolve this, an inductive transfer learning based framework, i.e.,Uncertainty Flow, is put forward in this research to allow knowledge transfer from a single labelled emotion recognition task to a multi-label affective recognition task. I.e., the model uncertainty—which can be quantified in Uncertainty Flow—is distilled from a single-label learning task. The distilled model uncertainty ensures the later efficient zero-shot multi-label affective learning. On the theoretical perspective, within our proposed Uncertainty Flow framework, the feasibility of applying weakly informative priors, e.g., uniform and Cauchy prior, is fully explored in this research. More importantly, based on the derived weight uncertainty, three sets of prediction related uncertainty indexes, i.e., soft-max uncertainty, pure uncertainty and uncertainty plus are proposed to produce reliable and accurate multi-label predictions. Validated on our manual annotated evaluation dataset, i.e., the multi-label annotated FER2013, our proposed Uncertainty Flow in multi-label facial expression analysis exhibited superiority to conventional multi-label learning algorithms and multi-label compatible neural networks. The success of our proposed Uncertainty Flow provides a glimpse of future in continuous, uncertain, and multi-label affective computing.
Highlights
Contrast with conventional implementations on Bayesian neural networks, we further provide our suggestion on the usage of weakly informative priors, e.g., uniform and Cauchy prior, in perfecting the final production of model uncertainty
The article is organised as following: we introduce the proposed Uncertainty Flow framework in sketch along with the description of four core components, e.g., Bayesian neural networks (More precisely, two hierarchical Bayesian neural networks); our suggested weakly informative priors; the quantification of model uncertainty; and three prediction related uncertainty indexes, e.g., soft-max uncertainty, pure uncertainty and uncertainty plus
Past works on neural network based affective computing have focused on the segmentation of single facial expression into finer sub-components, which can be achieved via the added principal component analysis (PCA) [9] or the complex feature pre-processing engineering, e.g., the introduction of Sobel filters [10]
Summary
Affective facial analysis, which is assessed as one of most primitive functions in vivo, has yet to be successfully implemented in machine. The past effort in resolving this ambiguity has been reflected in lowering the single-label dependency in producing emotion categories [4]. Like in Bai et al [5], the pseudo soft labels can be crafted by a continuous approximation to the original labels. This relaxation trick merely provides a provisional resolution in tackling the ambiguity (cf [5]). The research on multi-label affective discrimination is in line with the finding that the decision boundaries among classes are less ostentatious in affective analysis compare to other categorisation problems, e.g., object classification [2,3]. Benefited from previous researches on multi-label classification in general, it appears straightforward to extend affective computing along this direction. There is one difficulty that hinders the success application of multi-label affective recognition: it is laborious and expensive to collect the multi-label training data [6]
Sign-in/Register to view highlights & summary Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
