Master 2017 2018
Stages de la spécialité SAR
Active Learning for Movement-based Interactive Systems

Site : Trac-Active Learning for Movement-based Interactive Systems
Lieu : Laboratoire de Recherche en Informatique (LRI), Université Paris-Sud
Encadrant : Baptiste Caramiaux (CNRS LRI, Université Paris-Sud)
Rémunération :Standard
Mots-clés : Parcours ATIAM : Acoustique, Parcours ATIAM : Informatique musicale, Parcours ATIAM : Traitement du signal


General Context Body movements are a primary means for human to interact with the environment. And, through practice, humans can reach expert motor skills such as in music performance. Yet, in the context of Human-Machine Interaction, movements still remain challenging for a machine to interpret as current approaches cannot yet fully account for movement complexity such as movement variations due to expressivity and skills. This is a shortcoming for the use of movement in expressive interaction like musical interaction, or technology-misdated dance performance. One technical bottleneck relies in the data resources : movement data in HCI settings are usually very limited, preventing from the possibility to train complex and rich models. One promising approach is by considering active learning methods that have the capacity to learn incrementally through small amount of data.

Internship context, general objective In a HCI setting, an active learning based approach should 1) learn as fast as possible during the first increments, and 2) provide a feedback method to the user on the learning level. General objective : we want to investigate the behaviour of classical active learning techniques when they are trained by users, understand their pros and cons, and examine their understanding by users.

Specific aims The techniques considered in this internship are probabilistic Bayesian methods. In this internship we will focus on the ability of these techniques to incrementally learn. The specific aims are :

  1. Assess the extent to which algorithms learn with the incremental data provided,
  2. Assess the extent to which decisions made by the users while providing the system with movement data can enhance early learning, and,
  3. Find pertinent feedback mechanisms to be used in such interactive systems.

Internship tasks

  • Review literature on active learning to identify specifications of classical Bayesian active learning methods.
  • Implement one selected technique to be used in a simple gesture-based interface.
  • Collect a dataset with participants and benchmark the learning capacity of the implemented technique(s).
  • Collect a second dataset where users received feedback on the algorithm’s learning level during the study.
  • Compare algorithm accuracy and learning pathways.

Applicant skills

  • Coding skills (e.g. python, javascript, etc.) and creative skills
    - ¨ Great interest in movement-based interaction, machine learning, and cognitive science


Caramiaux, B. et al. (2017). Dynamic Bayesian networks for musical interaction. Routledge Companion on Embodied Music Interaction.

Gillies, et al. (2016). Human-Centred Machine Learning. CHI Ext. Abstract.

Tong, Simon (2001). Active Learning : Theory & Applications. PhD dissertation, Stanford.