The rapid evolution of Machine Learning (ML) workloads, particularly Deep Neural Networks (DNNs) and Transformer-based models, has intensified demands on computing architectures, highlighting the limitations of traditional von Neumann systems due to the memory bottleneck. To address these challenges, this paper investigates the mapping of fundamental Machine Learning (ML) operations onto ARCANE, a Near-Memory Computing (NMC)-based architecture that integrates Vector Processing Units (VPUs) directly within the data cache. ARCANE offers a flexible ISA-extension (xmnmc) abstracting memory management, effectively reducing data movement and enhancing performance. We specifically explore the acceleration capabilities of ARCANE when executing fundamental Deep Neural Network (DNN) and Transformer-based operations. Experimental results show that, with a contained area overhead, ARCANE achieves consistent speedups, delivering up to 150 × improvement in 2D convolution, 305 × in Linear layer, and over 32 × in Fused-Weight Self-Attention (FWSA), compared to conventional CPU approaches. These findings underline ARCANE’s significant benefits in supporting efficient deployment of edge-oriented Machine Learning (ML) workloads.

Coral reefs are crucial for biodiversity and provide vital resources for humankind. But despite such a central role, they are confronted to increasing threats linked to climate change, pollution, and local stressors. To ensure effective conservation, efficient and scalable monitoring is key: this necessitates automated identification of benthic classes and their states on a large scale through semantic segmentation. However, segmentation of underwater videos is challenging, because of visual similarities between benthic classes, underwater distortions and limited available datasets, making it harder to create accurate and robust models. In this paper, we present a method for training a semantic segmentation model on a small dataset of video frames of coral scenes, by fine-tuning a large transformer model. Our approach uses transfer learning on the Segment Anything Model (SAM), incorporating specific training and prediction strategies. We benchmark our model against a CNN for semantic segmentation as a baseline. Our results demonstrate a substantial improvement in model performance, particularly for benthic classes that often appear as small objects and rarer classes, highlighting the potential of our approach in advancing coral reef mapping and monitoring.

The rapid growth of AI-based Internet-of-Things applications increased the demand for high-performance edge processing engines on a low-power budget and tight area constraints. As a consequence, vector processor architectures, traditionally designed for high-performance computing (HPC), made their way into edge devices, promising high utilization of floating-point units (FPUs) and low power consumption. However, vector processors can only exploit a single dimension of parallelism, leading to expensive accesses to the vector register file (VRF) when performing matrix computations, which are pervasive in AI workloads. To overcome these limitations while guaranteeing programmability, many researchers and companies are developing dedicated instructions for a more efficient matrix multiplication (MatMul) execution. In this context, we propose Quadrilatero, an open-source RISC-V programmable systolic array coprocessor for low-power edge applications that implements a streamlined matrix ISA extension. We evaluate the post-synthesis power, performance, and area (PPA) metrics of Quadrilatero in a mature 65-nm technology node, showing that it requires only 0.65 𝑚𝑚 2 and that it can reach up to 99.4% of FPU utilization. Compared to a state-of-the-art open-source RISC-V vector processor and a hybrid vector-matrix processor optimized for embedded applications, Quadrilatero improves area efficiency and energy efficiency by up to 77% and 15%, respectively.

Wireless sensing offers an alternative to wearables for contactless monitoring of human activity and vital signs. However, most existing systems use bistatic setups, which suffer from phase imperfections due to unsynchronized clocks. Monostatic systems overcome this issue, but are hindered by strong self-interference (SI) that requires effective cancellation. We present a monostatic Wi-Fi sensing system that uses an auxiliary transmit RF chain to achieve SI cancellation levels of 40 dB, comparable to existing solutions with custom cancellation hardware. We demonstrate that the cancellation filter weights, fine-tuned using least mean squares, can be directly repurposed for target sensing. Moreover, we achieve stable SI cancellation over 30 minutes in an office environment without fine-tuning, enabling traditional vital sign monitoring using channel estimates derived from baseband samples without the adaptation of the cancellation affecting the sensing channel – a significant limitation in prior work. Experimental results confirm the detection of small, slow-moving targets, representative for breathing chest movements, at distances up to 10 meters in non-line-of-sight conditions.

The EPFL Graph Platform is an open-source data infrastructure designed for academic institutions. It organizes educational and institutional data into a semantically interconnected knowledge graph, making it accessible through a graph-based search engine and an LLM-powered chatbot. The platform is composed of five core services: Graph Registry, Graph AI, Graph Ontology, Graph Search, and Graph Chat.

This dataset primarily supports the Graph Ontology service and plays a foundational role in the platform’s semantic capabilities. It includes:

• A concepts graph constructed from over 40,000 Wikipedia pages, covering a broad spectrum of academic and scientific topics.
• A category tree that defines hierarchical relationships between concepts and categories, enabling structured navigation and inference.
• The semantic backbone used to detect, disambiguate, and link entities and keywords across diverse data sources such as course descriptions, lecture slides, research publications, and lab affiliations.

These structures are central to the platform’s ability to perform entity recognition, enable semantic search, and power AI-based recommendations. They allow users to search for academic topics (e.g., "quantum computing") and receive an integrated view of relevant content within the institution — from courses and lectures to researchers, labs, and scholarly output.

The urban landscape is in flux. Traffic congestion, environmental concerns, and a growing desire for flexible transportation options are pushing cities to rethink mobility. Mobility-on-Demand (MoD) systems, such as Uber and Lyft, offer a promising opportunity by linking passenger requests with available vehicles through smartphone applications. However, the growth of these services raises concerns regarding congestion and emissions, necessitating advanced operational strategies. This dissertation investigates the integration of automated vehicles (AVs) within MoD systems to enhance fleet coordination and overall service efficiency. It addresses two key challenges: empty vehicle repositioning and ride-pooling route planning. The research focuses on innovative methods to capture the complex relationship between passenger demand and vehicle supply in dynamic mobility systems, proposing advanced control algorithms to optimize fleet operations.

This dissertation includes two main parts. In Part I, a novel distributed coverage control scheme is introduced to guide idle vehicles to high-demand areas. Initially formulated in continuous space and further extended to a graph-based representation which better captures the nature of urban road networks, the method treats the repositioning challenge as an area coverage problem, aligning vehicle distribution with passenger demand. Further, a hierarchical framework is developed to coordinate vehicle repositioning at both macroscopic and microscopic scales. At the higher level, a data-driven predictive control algorithm is mainly described to manage inter-regional vehicle transfers, while the lower level leverages node-level position guidance to control individual vehicle movements. This framework effectively coordinates between the actions of the high-level controllers, which manage aggregated traffic components, and the self-governance of individual vehicles at the lower level. Part II extends the study to ride-pooling systems, where vehicles serve multiple passengers via one pooled trip. A theoretical and numerical study validates the potential benefits of detouring partially-occupied vehicles, which enhances their likelihood of matching with additional passengers. The pool-match probability between passengers and one partially occupied vehicle is modeled. Building on this foundation, a Mixed Integer Linear Programming (MILP) algorithm is developed to construct optimal detour paths by evaluating candidate road segments. It can adapt to fluctuating demand and traffic conditions. To evaluate the proposed methods, an operational, agent-based Mobility-on-Demand simulator is implemented, enabling analysis under diverse experimental scenarios. Simulation results demonstrate significant improvements, including a higher request answer rate, reduced waiting times, minimized empty travel distances, and enhanced profitabilityâ achieving a win-win-win outcome for customers, service providers, and the environment.

Catalytic hydrogenation stands out as a powerful tool for the atom-economical transformation of unsaturated systems. Nitrogen heterocycles are widespread in organic chemistry, they are found in many natural products and exhibit notable biological activities. Among them, piperidine is the most commonly encountered saturated heterocycle in alkaloids and drug discovery. Its high sp3 carbon-content contributes to higher drug-success rate and enables access to broader chemical space, escaping from traditionally flat sp2-rich drug molecules. Piperidines are notoriously difficult to access scaffolds and there does not seem to be a single strategy that would enable their streamlined synthesis. In contrast, pyridines are very easy to access and represent a highly modulable core, thus hydrogenation provides a simple way to access complex piperidines from readily available pyridines. However, the hydrogenation of pyridines remains underexplored compared to other heterocycles and most reports rely on a three-step strategy involving activation, hydrogenation and deprotection to access piperidines. Our first objective focused on the development of a catalytic system that would enable pyridine hydrogenation without the need for prefunctionalization. We found that, using an air and moisture stable C,N-cyclometalated iridium complex with a strong acid enabled the hydrogenation of a broad scope of (highly) substituted pyridines. Our method tolerates a large range of functional groups delivering unique and sometimes unprecedented functional group combinations on the piperidine ring, allowing fast access to new chemical space. We then imagined a new design of chiral catalysts that showed good promises in the enantioselective hydrogenation of pyridines. A library of chiral iridium complexes was synthesized and tested to assess what substitution could improve the catalyst performance. Further ligand design will need to be investigated in order to improve the catalyst efficiency and enantio-induction. The hydrogenation of complex nitrogen-containing-heterocycles was then investigated using the previously established methodology. This allowed us to access many piperazines under mild conditions and to explore the hydrogenation of pyridazines to piperidazines, without hydrogenation of the sensitive N-N bond. Several bicyclic systems were also readily hydrogenated, giving access to tetrahydronaphthyridines, tetrahydroazaindoles or tetrahydropyrrolo[1,2-b]pyridazines.

As it reveals the increasing complexity of the contemporary world, the socio-ecological transition demands a renewal of our gaze. By observing to imagine and drawing to narrate, architects and urbanists shape space through its representations and narratives, accompanying societies in their quest to find their place in the world. In this regard, cartography is an essential project tool for spatial designers, allowing them to sketch out possibilities. However, maps have long promoted a conception of inhabitability closely linked to technical control of the environment, embodying a parametric and anthropocentric gaze designed for a single subject: the human being. Today, socio-ecological challenges highlight the presence of other subjects who, also involved in the transformation of cities and territories, prove our relationships of coinhabitations. This paradigm shift therefore calls for a re-invention of our urban and territorial imaginaries, a re-writing of our narratives of inhabiting, and a re-thinking of our cartographic practices and gazes, in order to finally read the geo-eco-sociological realities of the Earth. By proposing a new narrative based on coexistence, the research supports the urgent need to re-present the conditions of our being-in-the-world in order to understand the dual nature - social and ecological - of the contemporary crisis.

Through what processes of re-presentation can the narrative of inhabitability evolve into a narrative of co-inhabitability?

To address this question, the research proposes a series of cartographic narratives to help interpret and conceptualize territories in transition, according to their potential conditions of coinhabitability. These conceptual and methodological explorations - guided by the research hypothesis that reconsiders the concept of territory as Territory-Subject - reveal the dynamic and evolving nature of territories both in space and time. The Atlas thus emerges as an obvious methodological choice. By combining approaches and diversifying themes, it brings together heterogeneous materials in dialogue. As a practical tool, the Atlas enables us to sketch out a new territorial narrative of coinhabitability, supported by the production of hybrid cartographies. The narrative and cartographic explorations re-present two territories, Geneva and Paris, which constitute both practical terrains and supports for analysis. Integrated into the research, these exercises contribute to the development of the new gaze required by the socio-ecological transition. They also offer a critical lever for de-constructing and re-interpreting existing cartographic practices, thus paving the way for an Atlas of the new gaze.

Non-regular or irregular statistical problems are those that do not satisfy a set of standard regularity conditions that allow useful theoretical properties of inferential procedures to be proven. Non-regular problems are prevalent; a classical example is the Gaussian mixture model, while more recently, the advent of machine learning has introduced models that are highly non-regular as well as black-box. In this thesis, we make some contributions to non-regular statistical inference. Our framework is likelihood-based; we give an overview of likelihood theory in Chapter 1.

In Chapter 2, we study universal inference, a method proposed by~\citet{Wasserman16880} that can construct finite-sample level $\alpha$ tests with minimal regularity conditions. We identify three sources of the resulting loss of power in the normal case, as a trade-off to this great generality. We show that universal inference becomes catastrophically conservative as the number of nuisance parameters grows, and propose a correction factor to mitigate this conservativeness while maintaining finite-sample level $\alpha$ error control. We demonstrate the viability of the correction factor on the non-regular problem of testing for the number of components in a two-component Gaussian mixture model. We also study the $K$-fold variant of universal inference and caution against using certain splits that lead to degenerate statistics.

In Chapter 3, we apply universal inference to construct model confidence sets with finite-sample coverage guarantees, which we dub universal model confidence sets (UMCS). We study the asymptotic properties of UMCS and establish its ability to include true and correct models and exclude wrong ones. We examine the application of the quasi-reverse information projection (qRIPR) to mitigate the conservativeness of UMCS, and study some cases where the application of qRIPR maintains the e-value property of the universal inference statistics, a property central to its error-controlling feature. We test the performance of UMCS to pick out signal covariates on a high-dimensional gene example.

In Chapter 4 we study discrete profiling, an extension of profile likelihood through the introduction of discrete nuisance parameters that index different functional forms modelling uncertainty. We extend the phenomenon of an observed bias in mis-specified normal linear models to a general one using asymptotic theory, and examine the ability of the discrete profiling algorithm to asymptotically detect mis-specified and slightly mis-specified models. We derive an expanded form for the discrete profile likelihood statistic and study its asymptotic properties under different cases of mis-specification. We corroborate our theory on the task of finding the best Student's $t$ density to model a normal density.

We conclude in Chapter 5 and discuss directions for future work.

The global plastic waste crisis remains a fundamental challenge for mankind to address, as vast amounts of plastic continue to accumulate in the environment, threatening ecosystems and human health. Biobased and biodegradable plastics, particularly materials based on aliphatic polyesters, offer a promising alternative to conventional non-degradable, petroleum-based polyolefins, but typically suffer from inferior thermomechanical performance. To comply with existing industrial processing techniques, such as blow molding or film drawing, these sustainable materials must be modified to exhibit sufficiently high melt strength and melt extensibility. In this context, polymers bearing end groups capable of supramolecular aggregation are particularly promising. Recent studies on high-molar-mass oligopeptide-modified polymers have shown that these materials exhibit rubber-like behavior and strain hardening in the polymer melt when blended them with additives that boost supramolecular motif concentration. This enables melt drawing to high draw ratios and the fabrication of highly oriented films. The present thesis investigates the dependence of this behavior on the deformation rate, the supramolecular motif and its concentration, and the molecular-scale network dynamics by using 1,3,5-benzenetricarboxamide (BTA) as a supramolecular motif in view of its reliable self-assembly into nanofibrils, comparatively high transition temperatures, and industrial relevance. For a representative aliphatic polyester, we demonstrate that BTA-based polymer end groups efficiently co-assemble with an additive into nanofibrils that serve multiple functions. They are highly efficient nucleating agents for the crystallization of the polyester matrix, and form supramolecular networks that give rise to a high-melt-strength rubbery regime extending to temperatures of up to 149 °C. The melt behavior under large tensile deformations, such as during film melt drawing, is governed by the competition between polymer chain stretching and relaxation processes, which are modulated by the supramolecular network dynamics. Using nuclear magnetic resonance spectroscopy, we establish a site-specific, non-destructive method to determine molecular scale dissociation rate constants for the nanofibrils in the bulk melt. Applying this technique, we show that BTA-based networks exhibit faster molecular-scale dissociation kinetics than oligopeptide-based materials. Consequently, despite their higher dissociation temperatures, melts containing BTA-based networks fail via viscous flow at strain rates and temperatures where oligopeptide-based blends remain extensible and display pronounced strain hardening. However, blending polyesters modified with BTA-based end groups with 1 wt% of a chiral BTA analogue yields aggregates with significantly slower exchange dynamics, resulting in strain hardening and dramatic improvements in melt extensibility (up to 3200%). Films produced by melt-drawing of such blends are optically transparent and exhibit a highly oriented "shish-kebab" morphology. This work demonstrates that tuning the dynamics of supramolecular networks is a key design principle for enabling large-strain processability in high-molar-mass supramolecular polymer materials. Our strategy is applicable to other polymer types and may facilitate the fabrication of oriented materials from sustainable polymers for many important applications, including barrier films, foams, and food packaging.