Topic proposals for 2022-2023 can be found below.
Personal research project in deep learning
This thesis proposal is not tied to a specific project. Instead, it welcomes students to make topic proposals on open research problems of their choosing and interest. Proposals should be centered around deep learning. Examples of projects include:
- Theoretical research in deep learning
- Improvements to existing deep learning algorithms or models
- Application of deep learning to solve a real-word problem
- Development of deep learning software Finding a research problem to work on is considered as part of this thesis subject. Students should come with a concrete and well-defined thesis topic. Proposals will be reviewed and discussed with the student before their approval (if any). [PDF]
Contact: Gilles Louppe.
Simulation-based inference in natural sciences
In many areas of science, complicated real-world phenomena are best described through computer simulations. Typically, the simulators implement a stochastic generative process in the forward mode based on a well-motivated mechanistic model with parameters \(\theta\). While the simulators can generate samples of observations \(x \sim p(x | \theta)\), they typically do not admit a tractable likelihood (or density) \(p(x | \theta)\). Probabilistic models defined only via the samples they produce are often called implicit models. Implicit models lead to intractable inverse problems, which is a barrier for statistical inference of the parameters \(\theta\) given observed data.
Depending on the student interests, the goal of this project will be to apply simulation-based inference algorithms developed within our research group to a scientific problem of her choosing. Direct collaboration opportunities include projects with climatologists, geologists, particle physicists or astronomers. Methodological improvements of those algorithms are also possible and welcome within the scope of this thesis.
Contact: Gilles Louppe.
Simulation-based inference of neural models from spikes
A fundamental question in neuroscience is how to link observed neural activity to the unobserved biophysical mechanisms that generate this activity. Therefore, there is a critical need for methods to incorporate the partial and noisy data that we observe with detailed, mechanistic models of neural activity.
In this project, we will explore how to estimate the parameters and the hidden variables of neuronal models from neuronal spike train responses. In particular, we will compare modern simulation-based inference methods to more traditional methods like particle filters. Depending on the progress, we will also investigate how to actively collect new data in closed-loop experiments to improve the inference. [PDF]
Contact: Gilles Louppe, Pierre Sacré.
Implicit neural representations for robotic grasping
Computer vision and deep learning research has recently seen impressive success in creating implicit neural representations of scenes and objects, such as Neural Radiance Fields (NeRFs) and Deep Signed Distance Functions (DeepSDFs). In this project, we will explore how these novel neural architectures can be used for learning a neural representation of the visual scene of a target object that is tailored for robotic grasping.
Contact: Gilles Louppe, Norman Marlier.
(Reproduced from Mildenhall et al, 2020.)
Using deep neural networks to understand the temporal structure of human memory
Deep learning is a powerful tool for testing theories of human cognitive functioning. In particular, recent research has highlighted the functional and architectural similarities of image classification deep convolutional networks with the hierarchy of human visual processing. Based on these works, the objective of this research project is to use a convolutional neural network to better understand the cognitive mechanisms underlying the temporal compression of events in human memory. [PDF]
Contact: Gilles Louppe, Arnaud D’Argembeau.
(Reproduced from Roseboom et al, 2019.)
Characterizing the performance of the SPHERE exoplanet imager at the Very Large Telescope using deep learning
Taking direct pictures of extrasolar planetary systems is an important, yet challenging goals of modern astronomy, which requires specialized instrumentation. The high-contrast imaging instrument SPHERE, installed since 2014 at the Very Large Telescope, has been collecting a wealth of data over the last eight years. An important aspect for the exploitation of the large SPHERE data base, the scheduling of future observations, and for the preparation of new instruments, is to understand how instrumental performance depends on environmental parameters such as the strength of atmospheric turbulence, the wind velocity, the duration of the observation, the pointing direction, etc. With this project, we propose to use deep learning techniques in order to study how these parameters drive the instrumental performance, in an approach similar to the one used by Xuan et al. 2018. This project will make use of first-hand access the large SPHERE data base through the SPHERE Data Center at IPAG/LAM (Grenoble/Marseille), and will include a stay of a few weeks to a few months at that institute to ensure direct access to the data base.
Contact: Gilles Louppe, Olivier Absil.
(Reproduced from Langlois et al, 2021.)
Deep learning for sport videos [EVS Broadcast Equipment]
In collaboration with EVS Broadcast Equipment, several master thesis topics are proposed on deep learning for sport videos. Thesis topics include: (1) tracking and re-identification of players in broadcast sport videos, (2) novel viewpoint synthesis of sport scenes sing broadcast images, (3) deblurring of sport videos using machine learning, (4) automatic captioning of broadcast videos, (5) image/sequence similarity engine for sport images and videos, (6) image quality assessment for the sport broadcast industry, and (7) video style transfer for soccer. Details can be found here.
Contact: Gilles Louppe, Oliver Barnich.
Survival analysis of glass fiber manufacturing [3B Fibreglass]
The project proposed with this MSc thesis focuses on analysing the reliability of the fiberization process at 3B Fiberglass by modelling and analysing its interruptions caused by glass filament breaks. The study will involve the probabilistic modelling of the process and an extensive survival analysis of its interruptions from historical data, taking into account all major confounding variables that may have caused them. The study will be complemented by a sensitivity analysis that will try to identify the most influencing parameters. [PDF]
Contact: Gilles Louppe, Bertrand Dechesne.
Previously supervised MSc thesis (2018-Present) can be found on Matheo.