Probabilistic Modeling of Motion Data

In our era of data abundance, we need the tools to model, search, visualize and understand it. Modeling should be able to faithfully represent uncertainty in our model structure and noise, be robust and adaptive, and scale well to many dimensions and large data sets.
Probabilistic modeling using the mathematics of probability theory allows us to express all types of uncertainty, perform inference about unknowns, make predictions, learn from data and adapt our models. It is the mainstay of modern artificial intelligence research.
In this thesis you will investigate probabilistic modeling of human motion time series data. Modeling human motion is central to both biomedical and computer vision applications. The high dimensionality, non-linearity and large variability of such data, makes it challenging to design robust and efficient models.
Some components of this line of research include:
  • comparing different models (Gaussian Processes, Gaussian Mixture, Hidden Markov, etc) for predictive modeling of motion trajectories
  • representing and learning motion through meaningful primitives (via clustering, e.g.)
  • adapting the learnt model to new situations using reinforcement learning 


Comfortably navigating the basics of probability theory and linear algebra;

Preferred familiarity with machine learning and reinforcement learning (this is possible to acquire throughout the course of the thesis as well)


Anna Harutyunyan (