Open Projects

Discrete flow-matching models for graph generation achieve state-of-the-art generative performance. However, this increased flexibility comes at the cost of several sampling hyperparameters that must be tuned to fully exploit the model’s capacity. In practice, these hyperparameters are selected via uni-dimensional sweeps followed by ad-hoc joint combinations, which is computationally inefficient and ignores rich information available during the sampling process itself. In particular, models such as DeFoG [1] and Directo [2] exhibit structured behavior along the denoising trajectory, including meaningful uncertainty and confidence patterns that can provide direct signals about sampling quality and failure modes. In this project, we aim to develop principled methods for automatically tuning sampling hyperparameters by exploiting signals arising during training and/or generation, such as prediction error, uncertainty, entropy dynamics, or other trajectory-level statistics. The goal is to outperform naive tuning strategies and ultimately surpass classical Bayesian optimization in sample efficiency, enabling few-shot or near one-shot hyperparameter selection without exhaustive sweeps. This work seeks to improve both the efficiency and robustness of discrete graph generative models, with applications in structured graph synthesis and molecular generation.

Requirements:
Mandatory: At least one deep learning course + prior experience with PyTorch.
Plus: Familiarity with diffusion or flow-based generative models and graph deep learning.

References:
[1] Qin et al., DeFoG: Discrete Flow Matching for Graph Generation, ICML 2025.
[2] Carballo-Castro et al., Generating Directed Graphs with Dual Attention and Asymmetric Encoding, arXiv 2025.

Project Creator:

Manuel Madeira ([email protected])

Units Involved:

Laboratoire de traitement des signaux 4 (LTS4)

Tags:

Master, EDEE, EDIC, Data & Network Sciences, AI,

Memberships:

Manuel Madeira – Role: Owner,

This project explores concept-based representations in pretrained vision models, focusing on concept bottleneck models (CBMs), sparse autoencoders (SAEs), and hybrids of the two. Recent work suggests that separately trained CBMs and SAEs can recover similar concept-oriented representations under certain conditions [1]. Building on this, the project asks how these methods compare in terms of interpretability, disentanglement, information leakage, and downstream usefulness, with the broader aim of understanding how to build more interpretable and robust concept-based vision models. 

[1] Fel, Thomas, and Gianni Franchi. "A Geometric Unification of Concept Learning with Concept Cones." arXiv preprint arXiv:2512.07355 (2025)

Project Creator:

Amel Abdelraheem ([email protected])

Units Involved:

Laboratoire de traitement des signaux 4 (LTS4)

Tags:

AI,

Memberships:

Amel Abdelraheem – Role: Owner,

Description: Can we learn to read the inner experience of a person through their art? Art Brut — the raw, unmediated artistic expression produced by people with psychiatric diagnoses, outside of any communicative intent — may encode cognitive signatures specific to mental disorders. Working with digitised collections from the Collection de l'Art Brut (Lausanne), the Psychiatriemuseum Bern, and the Prinzhorn Collection (Heidelberg), the student will apply deep learning and computer vision methods to extract and cluster symbolic patterns across thousands of artworks, in dialogue with experts in art history and symbolism. This project is part of a broader transdisciplinary research programme combining AI, psychiatry, and molecular biology.

Tasks:
Build structured representations of artworks using unsupervised/self-supervised vision models (e.g. CLIP, ViTs)Extract and cluster recurrent visual and compositional patterns across diagnoses and artistsDevelop interpretable visualisations of discovered patterns in collaboration with domain experts
Profile: MSc student in Computer Science, Electrical Engineering, or Data Science. Strong background in deep learning and computer vision (PyTorch/JAX). Genuine curiosity for interdisciplinary applications. Experience with NLP or multimodal models is a plus.

Contact: [email protected] (CV + short motivation)

Project Creator:

Pascal Frossard ([email protected])

Units Involved:

Laboratoire de traitement des signaux 4 (LTS4)

Tags:

Master, Bachelor, SEL, SIN, SSC, Data & Network Sciences, Imaging & Vision, Learning & Decision Systems,

Memberships:

Pascal Frossard – Role: Supervisor, Raphaelle Luisier – Role: Owner,

Graph generative models aim to learn distributions over graphs and produce new instances. They are widely used in applications where structure matters such as molecule design, circuit and program synthesis, and social network simulation. Recently, graph generation through discrete flow matching has achieved remarkable success [1, 2].

Nonetheless, these methods limit themselves entirely to the discrete space. Recent works in the language space have shown that augmenting the discrete state can help improve performance and efficiency at both training and sampling time. Our objective is therefore to build efficient graph generative models, possibly taking inspiration from recent advances in discrete diffusion language models.

Candidates should be comfortable with theory (generative models, graph-structured data) and coding (PyTorch), as the work will involve designing and implementing the framework, optimizing its training dynamics, and evaluating its performance on selected benchmarks.

Project Creator:

Alba Carballo Castro ([email protected])

Units Involved:

Laboratoire de traitement des signaux 4 (LTS4)

Tags:

Master, Data & Network Sciences, AI,

Memberships:

Alba Carballo Castro – Role: Supervisor,

Discrete diffusion models are emerging as a promising alternative to autoregressive models for generating discrete data such as language or graphs. These models operate by employing iterative refinement strategies directly on discrete state spaces [1, 2, 3]. However, the sampling step typically collapses the rich contextual information produced by the model at each step, which is then discarded during the remainder of the generation process.

In this project, we study fast and learnable sampling strategies for discrete diffusion models, aiming to exploit signals produced during generation to design more effective and efficient samplers.

Requirements:
Mandatory: At least one deep learning course + prior experience with PyTorch.
Plus: Familiarity with diffusion models or generative modeling.

References

[1] S. S. Sahoo et al., Simple and Effective Masked Diffusion Language Models, NeurIPS 2024.
[2] Y. Schiff et al., Simple Guidance Mechanisms for Discrete Diffusion Models, ICLR 2025.
[3] S. S. Sahoo et al., The Diffusion Duality, ICML 2025.

Project Creator:

Manuel Madeira ([email protected])

Units Involved:

Laboratoire de traitement des signaux 4 (LTS4)

Tags:

Master, EDEE, EDIC, AI,

Memberships:

Adrien Jean Deschenaux – Role: Student, Manuel Madeira – Role: Owner,

Creating RNAs with desired properties such as folding in a specific structure or binding with a given protein requires an in-depth understanding of sequence-structure dependencies and how these relate to functional properties. Most frameworks for RNA design sequence and structure sequentially, and optimize for one single property at a time. Representing RNA as a graph is a promising path towards designing both RNA sequence and structure at the same time, especially to optimize objectives that require specific structure-sequence patterns. 

We will explore the suitability of graph diffusion frameworks for RNA design, using existing methods developed in the LTS4 lab or elsewhere, with special consideration for how methods scale with sequence length. 

Candidates should have strong mathematical and computational skills and should be familiar with the pytorch framework. Candidates do not necessarily have to have a biological background but should have a strong desire to directly work with experimental biologists.

Project Creator:

Vincent Jung ([email protected])

Units Involved:

Laboratoire de traitement des signaux 4 (LTS4)

Tags:

Master, AI, Saturna,

Memberships:

Vincent Jung – Role: Owner,

A growing body of work has shown that incorporating domain knowledge into diffusion models by constraining their trajectories can significantly improve sample quality across various data modalities [1,2]. While similar ideas have been explored for graph generation, where discrete diffusion-based models have achieved state-of-the-art performance [3], existing approaches still exhibit notable shortcomings. For instance, PRODIGY [4] constrains graph generation by relaxing the adjacency matrix to a continuous space, which fails to fully capture the categorical nature of graphs. ConStruct [5], in contrast, operates entirely in discrete spaces but only imposes constraints at the structural level, without enforcing joint conditions between node and edge attributes. Such joint constraints are critical in many applications—for example, in molecular generation, where validity requires consistency between node types and their incident edges (e.g., a carbon atom must have degree at most 4). In this project, we aim to develop a graph diffusion method that natively enforces joint node-edge constraints during generation, with the goal of advancing state-of-the-art performance in real-world tasks such as molecular design.

Requirements:
Mandatory: At least one deep learning course + prior experience with PyTorch.Plus: Familiarity with graph deep learning and/or diffusion models.
[1] Lou et al., Reflected Diffusion Models, ICML 2023.
[2] Fishman et al., Diffusion Models for Constrained Domains, TMLR 2023.
[3] Vignac et al., DiGress: Discrete Denoising Diffusion for Graph Generation, ICLR 2023.
[4] Sharma et al., Plug-and-Play Controllable Graph Generation with Diffusion Models, ICML 2024.
[5] Madeira et al., Generative Modelling of Structurally Constrained Graphs, NeurIPS 2024.

Project Creator:

Manuel Madeira ([email protected])

Units Involved:

Laboratoire de traitement des signaux 4 (LTS4)

Tags:

Master, EDEE, EDIC, Data & Network Sciences, AI,

Memberships:

Alba Carballo Castro – Role: Supervisor, Manuel Madeira – Role: Owner,

Language Models are inherently black-box models. While interpreting LLMs trained on natural language is an active area of research, particularly in the field of mechanistic interpretability, little work has been done in biological language models, let alone RNA Language models (RNA-LMs). Yet, having a strong understanding of what RNA-LMs capture is important for multiple reasons: a lack of interpretability is one of the main bottlenecks to widespread adoption, our lack of knowledge of what models do and do not capture is limiting our ability to improve them, and it holds the possibility of finding new biology. 

The goal of this project is to study the applicability of existing MechInt method in RNA-LMs, identify the most promising framework for our use case, train an interpretability model on existing RNA-LMs and use extensive annotation datasets to relate sparse features to known biological knowledge. 

This project can be a Master semester project or a Master's thesis. Candidates should have strong mathematical and computational skills and should be familiar with the pytorch framework. Knowledge in Mechanistic interpretability is an advantage. Candidates do not necessarily have to have a biological background but should have a strong desire to directly work with biologists, bioinformaticians and computer scientists. 

 

Project Creator:

Vincent Jung ([email protected])

Units Involved:

Laboratoire de traitement des signaux 4 (LTS4)

Tags:

Master, AI, Saturna,

Memberships:

Vincent Jung – Role: Owner,

Recent advances in diffusion and flow-based models have significantly improved graph generation performance, particularly on simple synthetic and standard molecular datasets [1][2][3]. However, these gains appear to saturate when evaluated on existing benchmarks, which may be partly attributed to the lack of sufficiently challenging and realistic evaluation settings for undirected graph generation [4].
This project aims to construct new real-world application-based datasets and benchmarks current state-of-the-art models [1,2,3]. While many domains can be considered, we focus on two: structured biological sequences, such as RNA and DNA graphs [5], and benchmarks derived from combinatorial optimization problems [6]. To better evaluate graph generation in these settings, we will later develop domain-specific metrics tailored to each task.
References:[1] Vignac et al., DiGress: Discrete Denoising diffusion for graph generation, ICLR 2023.[2] Eijkelboom et al., Variational Flow Matching for Graph Generation, NIPS 2024.[3] Qin et al., DeFoG: Discrete Flow Matching for Graph Generation, ICML 2025.[4] Bechler-Speicher et al., Position: Graph learning will lose relevance due to poor benchmarks, Arxiv preprint.[5] Booy et al., RNA secondary structure prediction with convolutional neural networks, BMC Bioinformatics 2022.[6] Cappart et al., Combinatorial optimization and reasoning with graph neural networks, IJCAI 2023.

Project Creator:

Yiming Qin ([email protected])

Units Involved:

Laboratoire de traitement des signaux 4 (LTS4)

Tags:

Master, EDEE, EDIC, Data & Network Sciences, AI,

Memberships:

Yiming Qin – Role: Owner, Manuel Madeira – Role: Supervisor,

The advent of foundation models has revolutionized the landscape of machine learning, introducing a new paradigm where practitioners can access pre-trained checkpoints, such as those available on platforms like Huggingface, tailored for specific tasks. These models are derived from the same initial checkpoints but are fine-tuned on specific tasks, such as CIFAR or MNIST. Task arithmetic techniques [1,2] merge these different models into one multi-task model, i.e., a model that showcases good performance on all involved tasks, without needing additional training.
While the task arithmetic literature has focused on merging of classification models fine-tuned on different inputs, more traditional multi-task learning settings remain unexplored. An important case, for instance, is of a model solving the tasks of semantic segmentation, instance segmentation and depth regression from a single image input [3]. The goal of this project is to leverage existing model merging techniques for the single-input multiple-output case.

References:[1] G. Ilharco, M. T. Ribeiro, M. Wortsman, L. Schmidt, H. Hajishirzi, and A. Farhadi, “Editing models with task arithmetic,” in ICLR 2023. [2] P. Yadav, D. Tam, L. Choshen, C. A. Raffel, and M. Bansal, “TIES-Merging: Resolving interference when merging models,” in NeurIPS 2023.
[3] A. Kendall, Y. Gal, and R. Cipolla, “Multi-task learning using uncertainty to weigh losses for scene geometry and semantics,” in CVPR 2018

Requirements:Applicants must have completed at least one deep learning course and have experience with PyTorch. Familiarity with multi-task learning and model merging techniques is preferred.

Contact: [email protected], [email protected]

Project Creator:

Nikolaos Dimitriadis ([email protected])

Units Involved:

Laboratoire de traitement des signaux 4 (LTS4)

Tags:

Master, EDEE, AI,

Memberships:

Ke Wang – Role: Supervisor, Nikolaos Dimitriadis – Role: Owner,

This project tackles the challenge of integrating valuable knowledge from specialized models trained on sensitive or proprietary datasets into larger, generalist models. In fields like medicine, where data privacy and regulatory constraints limit data sharing, a small specialized model might excel at diagnosing rare diseases based on limited data. This project aims to leverage such specialized models to improve the performance of broad, generalist models. By applying innovative techniques like unsupervised reverse knowledge distillation [1][2] and model merging[3][4], this approach aims to enhance generalist models’ capabilities while adhering to privacy concerns. Engaging in this project will provide hands-on experience with cutting-edge methods in machine learning and offer a chance to contribute to practical solutions in data-sensitive domains.
References:
[1] Sahar Almahfouz Nasser, Nihar Gupte, and Amit Sethi. Reverse knowledge distillation: Training a large model using a small one for retinal image matching on limited data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7778–7787, 2024.
[2] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015.
[3] G. Ilharco, M. T. Ribeiro, M. Wortsman, L. Schmidt, H. Hajishirzi, and A. Farhadi, “Editing models with task arithmetic,” in ICLR 2023.
[4] Guillermo Ortiz-Jimenez, Alessandro Favero, and Pascal Frossard. Task arithmetic in the tangent space: Improved editing of pre-trained models. Advances in Neural Information Processing Systems, 36, 2024.

Project Creator:

Ortal Yona Senouf ([email protected])

Units Involved:

Laboratoire de traitement des signaux 4 (LTS4)

Tags:

Master, Learning & Decision Systems, AI,

Memberships:

Ortal Yona Senouf – Role: Owner, Nikolaos Dimitriadis – Role: Supervisor,

Graph Neural Networks (GNNs) have been established as a very competitive modeling in a wide range of applications due to their inductive bias and expressivity. However, their extension to temporal data becomes very challenging as the spatial and temporal updates are generally entangled. In this project we try to investigate how Neural ODE formulation helps us to learn latent graph trajectories, enabling interpretabilty of the flow field and learning in scarce data setting.

Project Creator:

Jérémy Jean Philippe Baffou ([email protected])

Units Involved:

Laboratoire de traitement des signaux 4 (LTS4)

Tags:

Master, EDIC, SIN, SMA, SSC, SSIE, SSV, Data & Network Sciences, Learning & Decision Systems, Theory and Simulation, AI,

Memberships:

Jérémy Jean Philippe Baffou – Role: Owner,

Tokenization in most RNA-Language Models (RNA-LMs) is done at the nucleotide-level. Attempts to adopt methods from the NLP literature such as Byte-Pair Encoding have been shown to lead to worse performance on downstream tasks [1]. Yet, moving away from nucleotide-level tokenization can greatly increase the effective input length of transformer-based RNA-LMs as well as create better inductive biases for generalization where the model can learn to decompose new samples into previously seen semantic units that are more meaningful than sole nucleotides. 

The goal of this project is to develop a new dynamic tokenization method specifically designed with RNA in mind, which incorporates known biological priors.

This project can be a Master semester project or a Master's thesis. Candidates should have strong mathematical and computational skills and should be familiar with the pytorch framework. Candidates do not necessarily have to have a biological background but should be willing to learn some fundamentals of RNA biology and work with bioinformaticians.

[1] Morales-Pastor, Adrián, et al. "Character-level Tokenizations as Powerful Inductive Biases for RNA Foundational Models." arXiv preprint arXiv:2411.11808 (2024).

Project Creator:

Vincent Jung ([email protected])

Units Involved:

Laboratoire de traitement des signaux 4 (LTS4)

Tags:

Master, AI, Saturna,

Memberships:

Vincent Jung – Role: Owner,

Tumor micro-environment (TME) has been shown to be a major factor in lesion development, drug resistance and disease trajectory. However, the wide range of cell types and interactions taking place in the TME make its analysis challenging. Graph Neural Networks (GNN) were proposed as powerful methods to tackle the obstacles faced in TME analysis, while providing an interpretable framework and being able to encode previous knowledge in their architecture. Despite their promising capabilities, GNNs struggle to capture higher order spatial patterns and long range interaction. In this work, we are using methods from topological data analysis (TDA) to enhance GNNs in the context of TME modelling and analysis.

Project Creator:

Jérémy Jean Philippe Baffou ([email protected])

Units Involved:

Laboratoire de traitement des signaux 4 (LTS4)

Tags:

Master, Data & Network Sciences, Imaging & Vision, Learning & Decision Systems, Personalized Health, AI,

Memberships:

Jérémy Jean Philippe Baffou – Role: Owner, Pascal Frossard – Role: Supervisor,