# FAU, Facial Expressions, Valence and Arousal: A Multi-task Solution

**Authors:** Didan Deng, Zhaokang Chen, Bertram E. Shi  
**arXiv:** [2002.03557](https://arxiv.org/abs/2002.03557)  
**Date:** 2020-02-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2002.03557v1/

## Key points

- The study focuses on training a unified model for three key tasks: predicting Facial Action Units (FAU, identifying seven basic facial expressions, and determining valence and arousal levels.
- Scarcity of fully-annotated datasets is a primary challenge in this endeavor.
- Authors propose an innovative algorithm for their multitask model to effectively learn from partial labels.
- Algorithm involves training a teacher model to execute all three tasks individually and utilizing its outputs as soft labels for training a student model.
- Student model outperforms the teacher model across all tasks due to exposure to complete set of labels during training.
- Ensemble modeling technique is implemented to further enhance performance on all three tasks.
- Research showcases novel methodology for addressing challenges in facial expression analysis through leveraging partial labels and ensemble modeling strategies.

## Summary

The study focuses on training a unified model capable of performing three key tasks: predicting Facial Action Units (FAU), identifying seven basic facial expressions, and determining valence and arousal levels. One of the primary challenges faced in this endeavor is the scarcity of fully-annotated datasets. Most existing datasets contain limited types of labels, making it difficult to train a comprehensive model. To address this challenge, the authors propose an innovative algorithm for their multitask model to effectively learn from partial labels. This algorithm consists of two main steps: firstly, a teacher model is trained to execute all three tasks individually. Each instance is trained using the ground truth label corresponding to its specific task. Subsequently, the outputs generated by the teacher model are utilized as soft labels. These soft labels, along with the ground truths, are then employed to train a student model. Remarkably, the results indicate that the student model surpasses the performance of the teacher model across all tasks. This improvement is attributed to the student model's exposure to a complete set of labels during training. Additionally, an ensemble modeling technique is implemented to further enhance performance on all three tasks. In conclusion, Deng et al. 's research showcases a novel methodology for addressing challenges related to multitask learning in facial expression analysis. By leveraging partial labels and employing ensemble modeling strategies, their approach demonstrates significant advancements in predicting FAUs, facial expressions, valence and arousal levels within facial imagery datasets.