Human Validation of Computer-Driven Action Segmentation

Contact: Jennifer Pomp

Human action is composed of chunks that are smoothly joined to form a continuous signal. Correspondingly, human observers as well as computers can decompose (segment) actions into characteristic phases. However, it remains unclear whether or not these phases are representative and relevant for action processing in the brain. To address this problem, we seek to quantify action phases asking whether purely data-driven action segmentation by computers is convergent with signatures of natural action segmentation as measured by human brain responses (fMRI) and behavior.
Thus, the common aim of this work is to arrive at an objective description of actions, and to biologically validate this approach to the characterization of human object manipulations.

This project gets funded by the German Research Foundation (SCHU1439-8-1)

Read more:

Pomp, J., Heins, N., Trempler, I., Kulvicius, T., Tamosiunaite, M., Mecklenbrauck, F., Wurm, M.F., Wörgötter, F., Schubotz, R.I. (2021) Touching events predict human action segmentation in brain and behavior. NeuroImage, 243 (2021) 118534, doi.org/10.1016/j.neuroimage.2021.118534

Ziaeetabar, F., Pomp, J., Pfeiffer, S., El-Sourani, N., Schubotz, R.I., Tamosiunaite, M., Wörgötter, F. (2020). Using enriched Semantic Event Chains to model human action prediction based on (minimal) spatial information. PLOS One, 15(12): e0243829, doi.org/10.1371/journal.pone.0243829