An Adaptive Neural Approach Based on Ensemble and Multitask Learning for Affect Recognition

<p> In this paper, we evaluate the effect of Multitask Learning (MTL) in an ensemble with shared representations based on convolutional networks in the task of affect recognition from facial expressions. Our convolutional architecture is divided into three levels of hierarchy regarding MTL. The first level is conditioned to learn lower-level representations, which are shared with independent convolutional branches related to different tasks on the second level. While each independent branch is fostered to learn task-specific representations, the early shared layers are fostered to learn features that are relevant to multiple tasks due to the inductive transfer mechanism from MTL. The third level consists of an ensemble of convolutional branches responsible for learning higher-level representations and allowing re-training with unlabelled expressions. Our experiments show a slight improvement in recognition performance using MTL over Single Task Learning (STL) on the AffectNet dataset, but a significant reduction in training time. Finally, we discuss the potential use of MTL and hard constraints into the inference and re-training processes of the proposed approach to improve its generalization performance. </p>

@InProceedings{Siq18, 
 	 author =  {Siqueira, Henrique},  
 	 title = {An Adaptive Neural Approach Based on Ensemble and Multitask Learning for Affect Recognition}, 
 	 booktitle = {International Ph.D. Conference on Safe and Social Robotics (SSR)},
 	 editors = {},
 	 number = {},
 	 volume = {},
 	 pages = {},
 	 year = {2018},
 	 month = {},
 	 publisher = {},
 	 doi = {}, 
 }