8 January 2020

NeurIPS tutorial about human behaviour modeling by Albert Ali Salah

Albert Salah

Albert Ali Salah (Utrecht University, Social and Affective Computing) and Nuria Oliver (European Laboratory for Learning and Intelligent Systems) gave the first tutorial of the 2019 NeurIPS conference in Vancouver, entitled ‘Human Behavior Modeling with Machine Learning: Opportunities and Challenges’. Over a thousand researchers attended the tutorial. Computer analysis of human behaviour is a complex and growing research area. Behaviours are shaped by many factors, they are contextual and idiosyncratic, and have temporal dynamics at several levels.

The tutorial covered principles and difficulties of behaviour analysis at different scales: individual behaviours, such as facial expression analysis and activity recognition; dyadic behaviours, where people are interacting in different scenarios; and computational social science, where the behaviour of thousands of people are analyzed at once. The last part of the tutorial dealt with ethical and privacy aspects of behaviour analysis, which cannot be neglected. Especially for large scale behaviour analysis, it is essential that personal information is removed from the data through appropriate aggregation and anonymisation procedures.

Human behaviour analysis

Human behaviour analysis finds applications in diverse domains. At Utrecht University, the Social and Affective Computing group focuses on mental healthcare scenarios (e.g. analysis of depression, trauma, bipolar disorder), on dyadic interactions (e.g. child-parent and child-therapist interactions), as well as large-scale analysis (e.g. human mobility in migration and refugee studies).


Neural Information Processing Systems (NeurIPS) is the flagship conference of the machine learning community, organized since 1987. The number of attendees grew from 1000 in 2007 to 13,200 in 2019, with increasing popularity of deep learning. This increased popularity is enabled by new hardware and algorithms that make it possible to leverage vast data repositories on the Internet for developing machine learning models with greater capabilities than ever before. Furthermore, synthetic data generation approaches are also used to feed the data-hungry models.