The Data Mining and Machine Learning group, http://dmml.ch/, at the University of Applied Sciences in Geneva has an opening for a full-time post-doc position funded by a Swiss National Science Foundation, funding is available for up to four years. Within the project we seek to develop learning methods and models that will accurately replicate human locomotion from observations only, i.e. sequences of state-state transitions as these are captured by motion capture systems. The overall objective of the project is the modelling of pathological human gait in order to be able to explore the result of different theurapetic approaches. The project is a collaborative effort with the Kinesiology Lab of the University of Geneva [1].
Within the project we will follow two approaches to model gait. A generative modelling approach in which we will seek to exploit the presence of locomotion simulators, potentially differentiable, in order to learn to generate gait that is compliant by construction with the underlying physics, see our previous work [2,3]. An imitation learning approach, [4], in which the emphasis will be on the development of methods that learn from observations only, i.e. the agent's actions are not observed; within that approach we are particularly interested in the exploitation of inverse and forward models as means to improve the learning efficiency [5]. Additional challenges that we want to address include the high dimensionality of the state and action spaces.
We seek strongly motivated candidates dedicated to high quality research. Candidates should have (or be close to obtaining) a Phd in machine learning, ideally in the area of generative modelling, and a strong research track-record attested by high quality publications in relevant machine learning venues such as ICML, NeurIPS, ICLR, AI-STATS, UAI, etc. The selected candidate is expected to demonstrate a high degree of independence and autonomy, drive their own research and actively contribute to the scientific development of the group through their knowledge and expertise as well as by proposing and contributing to group activities such as readings, schools, workshops etc. They are expected to participate in the supervision of PhD students.
The DMML group consists of roughly a dozen of researchers at the PhD and Post-doc level, working in different areas of machine learning, such as generative models, imitation learning. The team collaborates closely with the VIPER group, https://viper.unige.ch, from the computer science department of the University of Geneva headed by Prof. Stephane Marchand-Maillet. We offer ample opportunities and support for scientific development, e.g. providing funding for conferences, schools, research visits and exchanges etc. We strive to provide a research environment in which researchers can focus on their research and allow for space and time to develop solid ideas. The group regularly publishes to some of the best conferences in ML/AI and the alumni of the group hold research positions in academia and in GAFA companies.
If interested, please send the following to alexandros.kalousis@hesge.ch (subject: Post-Doc application, Metathesis Project)
- academic CV (max 2 pages)
- pointers to their two most important publications
- one page motivation letter explaining why the candidate is suitable for the position and what they can bring to the DMML group
- 1000 word research proposal on the research topics described above
- contact details of three referees (do not send reference letters)
- copies of diplomas (PhD, MSc, BSc) and academic transcripts
Deadline for applications: 15/03/2024, starting date May 2024 or later.
In case of any further questions, please contact alexandros.kalousis@hesge.ch.
[1] Kinesiology Lab of the University of Geneva https://www.unige.ch/medecine/kinesiology/[2] N. Takeishi and A. Kalousis. Variational autoencoder with differentiable physics engine for human gait analysis and synthesis. In NeurIPS workshop on Deep Generative Models applications, 2021a.
[3] N. Takeishi and A. Kalousis. Physics-integrated variational autoencoders for robust and interpretable generative modeling. In Advances in Neural Information Processing Systems 34, 2021b.
[4] L. Blonde at all. Lipschitzness is all you need to tame off-policy generative adversarial imitation learning. Machine Learning Journal, 2022.
[5] J. Ramos et all. Sample-Efficient On-Policy Imitation Learning from Observations, 2023.