Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Semester/Master Thesis: GAN Based Data Synthesis for the Detection of Emotion Anomaly
Awareness of driver emotions facilitates intelligent automobile applications that improve driver comfort, well-being and safety. Automotive manufacturers such as Mercedes have developed tools to improve the emotional state of a driver via built-in comfort systems such as massage, fragrance, music, ionisation of the air in the cabin. Ridesharing companies (such as Uber, Didi) are improving quality of service by monitoring the mental status of their drivers. Recognition of driver emotion is gaining momentum in both academia and industries.
Keywords: Deep Learning, GAN, CAN data, Human-Computer Interaction, Affective Computing
Numerous existing works proposed methods that can effectively detect driver mental status such as stress, valence or arousal based on driving data (i.e. CAN data). However, these emotions fail to capture a driver’s fine-grained extreme negative emotions such as anger and hostile. One biggest challenge in detecting extreme negative emotions is the unbalance of data distribution.
For example, a driver may only feel angry for 5% of the time when he drives. It it vital to detect such anger moment to prevent road rage or to guarantee service of quality for ride sharing companies (e.g. Uber, Didi, Lyft etc.). However, the extreme unbalanced dataset makes machine learning models prone to overfitting.
We can frame extreme emotion detection as an anomaly detection problem, where anomaly corresponds to minority class (i.e. angry moment) in the emotion labels. The unbalance leads to the degeneration of machine learning models because there is no enough diversity in the data of minority class. Addressing this challenge, in this work we proposed a generative adversarial network (GAN) based framework to synthesise data of minority class. The GAN framework consists of two major parts: a) a generative model that learns to synthesise artificial driving data of minority class and b) a discriminative model that learns to distinguish synthesised artificial data from real data. Two parts of model are trained jointly and can therefore lead to high quality synthesised data. The synthesis of data minimise the drawback of unbalance in real data and improve the detection accuracy of emotion anomaly.
Numerous existing works proposed methods that can effectively detect driver mental status such as stress, valence or arousal based on driving data (i.e. CAN data). However, these emotions fail to capture a driver’s fine-grained extreme negative emotions such as anger and hostile. One biggest challenge in detecting extreme negative emotions is the unbalance of data distribution. For example, a driver may only feel angry for 5% of the time when he drives. It it vital to detect such anger moment to prevent road rage or to guarantee service of quality for ride sharing companies (e.g. Uber, Didi, Lyft etc.). However, the extreme unbalanced dataset makes machine learning models prone to overfitting. We can frame extreme emotion detection as an anomaly detection problem, where anomaly corresponds to minority class (i.e. angry moment) in the emotion labels. The unbalance leads to the degeneration of machine learning models because there is no enough diversity in the data of minority class. Addressing this challenge, in this work we proposed a generative adversarial network (GAN) based framework to synthesise data of minority class. The GAN framework consists of two major parts: a) a generative model that learns to synthesise artificial driving data of minority class and b) a discriminative model that learns to distinguish synthesised artificial data from real data. Two parts of model are trained jointly and can therefore lead to high quality synthesised data. The synthesis of data minimise the drawback of unbalance in real data and improve the detection accuracy of emotion anomaly.
Your task will be focused on the design of GAN.
Framework for data cleansing and feature engineering can be reused from our previous works.
Your task will be focused on the design of GAN.
Framework for data cleansing and feature engineering can be reused from our previous works.
Shu Liu (liush@ethz.ch).
Please contact us with your CV, a short statement of motivation, and your current transcripts of records (bachelor & master).
Shu Liu (liush@ethz.ch).
Please contact us with your CV, a short statement of motivation, and your current transcripts of records (bachelor & master).