ELLIS PhDs students & Postdocs at EPFL

Reinforcement Learning and Imitation Learning achieved impressive empirical results in recent years. Still, little is known about the theoretical properties of commonly used algorithms. This PhD project aims to use insights from the field of game theory and optimization to develop algorithms with mathematical guarantees. The hope is to bring contributions on the practical RL side as well, proposing theory grounded algorithms easily applicable and implementable.

Primary Host: Volkan Cevher (EPFL)

Exchange Host: Gergely Neu (Universitat Pompeu Fabra)

PhD Duration: 01 September 2021 – 01 September 2025

Exchange Duration: 01 March 2023 – 01 September 2023

In recent times, research in robotics devoted a significant effort to the development of compliant manipulators capable of interacting with the surrounding environment safely and effectively. However, while many studies focused on pushing the limits of human-robot interaction with novel control, planning, and task allocation strategies, little has been done to enable an intelligent (i.e. not pre-programmed in advance) cooperation. This is a challenging problem because -for a purposeful cooperation- the robot is required to implement various capabilities, such as: i) learning to interact with a world made for humans; ii) understand and predict human actions; iii) decide on-the-fly how to favour the human activity. To reach this goal, egocentric vision seems to represent a key enabling technology. Indeed, dealing with first person videos comes with the benefit that the data source already embeds an intrinsic attention mechanism, driven by the focus of the user, and can serve as prior for human-inspired skills learning. To effectively learn how to encode human skills from egocentric videos, this thesis will investigate the feasibility to identify atomic actions from the complexity of unstructured daily living behavior. The definition of these building blocks will serve the twofold purpose of i) providing a better understanding of human behavior, helping deep neural methods to better recognize and forecast human activities and ii) enable a more efficient and accurate skill transfer as combination of atomic actions toward intelligent manipulators.

Primary Host: Giuseppe Averta (Politecnico di Torino)

Exchange Host: Pascal Frossard (EPFL)

PhD Duration: 01 November 2022 – 31 October 2025

Exchange Duration: 01 January 2025 – 30 June 2025

Understanding how neural circuits enable behavior is a critical challenge in neuroscience: it has important implications for brain-machine-interfaces (BMIs), robotics, and neuro-rehabilitation. As our ability to record large neural and behavioral data increases, there is growing interest in modeling neural dynamics during adaptive behaviors to probe neural representations (Urai et al Nature Neuroscience 2022). In particular, new non-linear methods that discover neural latent embeddings can reveal underlying correlates of behavior (Schneider et al Nature 2023), yet, we lack causal, mathematical understanding of these latents, which is required to be able to causally test their role. Moreover, we need our methods to be identifiable, and explainable. The PhD project therefore aims to bridge ideas from disentangled representation learning (Whittington et al ICLR 2023), contrastive learning (Schneider et al Nature 2023), and new works in causal component analysis (Liang et al arXiv 2023), to build new methods for causal discovery in neuroscience.

Primary Host: Mackenzie Mathis (EPFL & Harvard University)

Exchange Host: Timothy Behrens (University of Oxford)

PhD Duration: 01 February 2024 – 01 February 2029

Exchange Duration: Ongoing

Despite the extraordinary success of deep learning, its emerging weaknesses such as robustness, generalization, and bias, demand an ever closer attention. Unfortunately, many of the existing theories were developed for low capacity models and therefore do not account for the impressive scaling properties of deep learning. For example, enforcing low bias in computer vision tasks using tools developed for linear models have been shown to simply degrade the performance of neural networks. Instead, our work already shows how dedicated theories can shed light on how architectural choices impact robustness and generalization. In this PhD project, we will first investigate how common practices theoretically affect the robustness, generalization, and bias of neural networks. Second, we will propose new grounded approaches improving over existing methods. The overall goal is to improve the reliability of neural networks starting from their theoretical foundations and optimization to improving their behaviour under principled forms of distribution shifts such as covariate and label shifts.

Primary Advisor: Volkan Cevher (EPFL)

Industry Advisor: Francesco Locatello (IST Austria)

PhD Duration: 01 September 2022 – 31 August 2026

Reinforcement learning offers a solution to learning problems that require planning and has led to several breakthroughs in recent years. However, many of these breakthroughs were achieved in controlled setups. In such setups, it is common that a) one does not require a theoretical understanding of the algorithms and b) only the eventually trained policy but not performance during learning matters. This PhD project aims to provide reinforcement learning algorithms that allow for the desired mathematical guarantees. Crucially, these algorithms are supposed to provably scale to large Markov decision processes at the same time. The key idea is to view the learning problems from the viewpoint of online optimization theory.

Primary Host: Volkan Cevher (EPFL)

Exchange Host: Gergely Neu (Universitat Pompeu Fabra)

PhD Duration: 01 September 2023 – 01 September 2027

Exchange Duration:  01 September 2025 – 01 March 2026

This research project aims to enhance the understanding of the broader societal impacts and challenges associated with large language models (LLMs) such as OpenAI’s GPT-4. The study investigates the existing misalignment between LLMs and human society, and proposes efficient methods for LLMs to learn from human feedback, enabling them to acquire knowledge and capture human preferences more effectively. Additionally, the project explores the potential impacts of regular users on LLMs and develops algorithms that allow for collective correction to tackle potential ethical and legal dilemmas of LLMs. Furthermore, a key objective is to elucidate the long-term effects surrounding LLMs, particularly by examining the intricate feedback loops between humans and LLMs. In doing so, this research aims to develop methods that ensure LLMs evolve in a manner that promotes the communal welfare of humanity. By bridging the gap between human society and LLMs, this project seeks to address the ethical, legal, and social implications that may arise and strives for the responsible integration of LLMs into human environments.

Primary Host: Moritz Hardt (Max Planck Institute for Intelligent Systems)

Exchange Host: Martin Jaggi (EPFL)

PhD Duration: 01 August 2023 – 01 August 2026

Exchange Duration:  01 February 2026 – 01 August 2026

This thesis will explore the effectiveness of employing a Socratic approach to the development of chatbots with a particular focus on educational settings, with a specific focus on fostering active learning, critical thinking, and knowledge retention. Building upon existing literature on the Socratic Method in education and the potential of leveraging chatbots for learning, this thesis will first develop a Socratic chatbot. Furthermore, it will carry out a comparative analysis between traditional teaching methods, standard chatbots, and a specially-designed Socratic chatbot. Metrics for evaluation will include the depth of reasoning, knowledge retention, learner engagement, and critical thinking skills. An additional layer of this study will scrutinize the influence of power and trust dynamics on the learning process when interacting with a bot. Given the importance of emotions in learning, the research performed in this thesis will explore the challenges and value of incorporating emotional intelligence features in the chatbot to adapt to the emotional and mental states of the students. The ultimate goal is to create a more intellectually stimulating and emotionally supportive learning environment. The findings derived from this thesis aim to offer valuable insights and implications in the design of AI-based conversational agents for learning.

Primary Host: Nuria Oliver (ELLIS Alicante Unit Foundation | Institute of Humanity-centric AI)

Exchange Host: Tanja Käser (EPFL)

PhD Duration:  01 November 2023 – 01 November 2026

Exchange Duration:  01 June 2025 – 31 December 2025

Mr. Sun will explore novel and neural algorithms and representations, which in conjunction with new datasets will allow to synthesize novel motions and interactions with objects in the scene. In contrast to classical computer animation, one goal of his thesis is to simplify the user input to enable the usage by non-experts while achieving realistic looking skeletal motions. Moreover, he will also explore how a joint reasoning of the human and the scene can further benefit the tracking of a dense surface when only a single RGB video of the human and the scene is given.

Primary Host: Christian Theobalt (Max Planck Institute for Intelligent Systems)

Exchange Host: Pascal Fua (EPFL)

PhD Duration:  31 October 2022 – 31 October 2026

Exchange Duration:  31 December 2023 – 30 June 2024

There is increasing evidence that people with higher access to urban green spaces have better mental health and well-being. This project aims to examine the impact of urban green spaces on mental health, and what features play the biggest role. The combination of airborne or space-borne remote sensing images and geo-tagged social media data can enhance the measurement of the quantity and quality of urban green spaces. Remote sensing imagery can allow us to identify and classify green spaces. Social media data can represent human activity near the identified green spaces. We hope to gain insight into qualitative characteristics of urban green spaces that could be useful for city planners and has a positive impact on citizens’ mental health.

Primary Host: Xiaoxiang Zhu (Technical University of Munich)

Exchange Host: Devis Tuia (EPFL)

PhD Duration:  01 October 2019 – 30 September 2023

Exchange Duration: 01 September 2020 – 31 December 2020

Clément’s work focuses on the design of neural architectures for structured data: sets, graphs and point clouds. These problems have in common a large symmetry group, which is the invariance to all possible permutations of the points. In order to design architectures that are both computationally and data efficient, these symmetries must be accounted for. This raises several challenges, as existing functions often come with a tradeoff between computational cost and expressive power. During the first years of his PhD, Clément has developed new insights and methods to design such functions. The next steps consist in adapting these methods to the problem of graph generation and especially to molecule generation for drug discovery, which combines both interesting theoretical challenges and a large interest from the industry. This task has not been studied very extensively from a graph perspective yet, and several innovations from graph representation learning could be used to improve existing methods. The University of Amsterdam has a renowned expertise in generative models (with the development of Variational Autoencoders and important contributions to Flow-based methods) and symmetry-aware machine learning. We hope that, combined with our experience with geometric deep learning, it will result in a fruitful collaboration.

Primary Host: Pascal Frossard (EPFL)

Exchange Host: Max Welling (University of Amsterdam)

PhD Duration: 01 November 2019 – 03 May 2023

Exchange Duration:  01 September 2021 – 31 December 2021 & 01 June 2022 – 31 July 2022

Regularization lies at the core of successful training state-of-the-art deep neural networks. It allows to control overfitting and allows to obtain good generalization even with massively overparametrized models. Regularization influences the training process both implicitly – through the properties of optimizers – and explicitly – by using a regularized loss function, dropout, batch normalization, etc. The goal of this Project is to shed more light on the foundations of regularization techniques employed in deep learning and to formally ground empirical results using the insights from the regularization theory.

Primary Host: Asja Fischer (Ruhr University Bochum)

Exchange Host: Martin Jaggi (EPFL)

PhD Duration: 01 February 2021 – 31 March 2024

Exchange Duration: 01 June 2022 – 31 December 2022

Problems with a large number of interacting particles are ubiquitous in science and key to many future technologies. Recently, the successful application of deep learning methods to such problems led to remarkable progress in statistical mechanics, quantum chemistry and condensed matter physics. We will investigate variational methods and Graph Neural Networks in the context of many-body physics problems related to the identification of ground states and phases of matter. For supervised methods a crucial aspect will be the data efficiency and the combination with few-shot learning and physics inspired model architectures.

Primary Host: Sepp Hochreiter (Johannes Kepler University Linz)

Exchange Host: Guiseppe Carleo (EPFL)

PhD Duration: 02 May 2021 – 31 October 2024

Exchange Duration: 01 February 2024 – 01 August 2024

Understanding the mechanisms underlying robust learning and efficient adaptation is an open problem both in neuroscience and machine learning. While robustness and domain adaptation in ML is commonly studied with computer vision tasks, adaptation research in neuroscience has been traditionally carried out in sensorimotor paradigms. My PhD project is a collaborative project between the Mathis lab who studies motor adaptation and the Bethge lab who studies robust ML methods under domain shifts. We will build computational models of neural activity and behavior to closer study neural mechanisms of continual learning during motor adaptation using tools from machine learning. We will compare representations arising in adaptive ML models with those present in the brain. Motor adaptation tasks offer the unique opportunity to precisely control distribution shift, difficulty and learning objectives and link the biological findings to machine learning using insights from dynamical systems and network theory. 

Primary Host: Matthias Bethge (University of Tübingen)

Exchange Host: Mackenzie Mathis (EPFL & Harvard University)

PhD Duration: 01 November 2019 – 31 October 2022

Exchange Duration: 18 February 2020 – 31 March 2020 & 01 January 2021 – 31 January 2022

Years of a fierce competition towards the best results have naturally selected the fittest deep learning algorithms. However, although these models work well in practice, we still lack a proper characterization of why they do so. This poses serious questions about the robustness, trust, and fairness of modern AI systems. Understanding deep learning is, thus, fundamental for its long-term success; and Guillermo’s work is shedding new light on the fundamental mechanisms behind it. By studying the complex interaction between data, architecture, and optimization algorithm, Guillermo is exposing key insights about the way neural networks see the world; and is working towards applying these results to the design of more reliable neural networks. The University of Oxford hosts some of the leading researchers in AI robustness worldwide, so we hope that Guillermo’s exchange will result in a fruitful collaboration.

Primary Host: Pascal Frossard (EPFL)

Exchange Host: Philip H. S. Torr (University of Oxford)

PhD Duration: 01 November 2018 – 01 May 2023

Exchange Duration: 01 January 2022 – 30 June 2022

As Deep Learning gets better at visual tasks, including species identification, the learned reasoning behind its decisions gets increasingly obscure. This is in contrast with the procedures developed by taxonomists, the experts in charge of defining the hierarchy of natural species, for manual species recognition. These procedures lead users to follow an identification key, a structured set of attribute observations, to reach a final conclusion. I will be working towards incorporating this structured reasoning into Deep Learning models for species recognition such that their results become more interpretable, hopefully helping experts to spot mistakes or even yet-to-be-described species, and offering amateur users an expert explanation that can help them become experts themselves.

Primary Host: Devis Tuia (EPFL)

Exchange Host: Zeynep Akata (University of Tübingen)

PhD Duration: 01 February 2019 – 31 January 2023

Exchange Duration: 01 August 2020 – 30 October 2020

Object rigidity is still one of the most common assumptions in robotic grasping and manipulation. Nevertheless, many daily life objects like cables, plastics, and clothes present non-negligible deformations. Manipulating garments remains a challenging topic due to the high flexibility of textiles and their nearly infinite number of degrees of freedom. Due to the intricate dynamics of clothes, the existing deformation models are still far from accurate when predicting future states after forces are applied, especially over long time horizons. For this reason, robots can leverage real-time feedback to update the belief of the manipulated garment state and adapt their actions accordingly. To extract cloth properties that may be useful to understand the object’s behaviour, we propose to combine vision and proprioceptive sensory systems to capture the state of the cloth at each instant. While cameras provide global information about shapes, force and tactile sensors can provide local information about material properties like stiffness. To allow for real-world applications, the obtained cloth model should adapt to task and environment variations, be sample and computationally efficient, and deal with uncertainty.

Primary Host: Carme Torras (Universitat Politècnica de Catalunya)

Exchange Host: Amir Zamir (EPFL)

PhD Duration: 01 July 2021 – 31 May 2024

Exchange Duration: 01 January 2023 – 31 May 2023