Ofertes de projectes

A more detailed description of the project can be found in

https://www.cs.upc.edu/~eromero/Downloads/Retina-TFM-Project-01.pdf

The project is proposed in collaboration with Javier Zarranz Ventura
(Institut Clínic d'Oftalmologia, ICOF, Hospital Clínic de Barcelona, and
Institut d'Investigacions Biomèdiques August Pi I Sunyer, IDIBAPS),
which would provide a large annotated database to develop the project. For further information, please contact Alfredo Vellido (avellido@cs.
upc.edu) or Enrique Romero (eromero@cs.upc.edu).

 State-of-the-art models such as LLMs are too large to fit in a single compute node (GPU, NPU, CPU), both for training and inference on a device (e.g., phone, laptop, tablet) or in larger-scale data centers. There is a need to develop optimization techniques to split and place these models onto a distributed set of compute nodes so that the overall system performance is maximized. The research will be focused on optimizing the placement of AI models onto distributed systems considering training time, energy consumption, and computational resources. 

In this thesis, the student will model and simulate distributed AI workloads using both mathematical frameworks and simulators. This encompasses the modeling of network, compute, and memory components within a distributed architecture. Developing new modules to enhance the modeling process. Evaluating and optimizing various parallelization techniques to improve overall system performance.

The Universitat Politècnica de Catalunya · BarcelonaTech offers Master thesis fellowships in the field of LLM training. The research will be supported by Qualcomm and will be carried out in an environment with a strong interaction with leading experts in the field, with opportunities for doing internships in the company.

More information here: https://www.cs.upc.edu/~jordicf/priv/eda/llm_qc.html

This project offers the candidate a unique opportunity to apply artificial intelligence techniques to real-world challenges in lung cancer research and treatment. As part of this thesis, the candidate will work with datasets, including patient records, genetic data, molecular alterations, treatment outcomes, and exposome data. These datasets will serve as the foundation for developing AI models that address critical challenges in lung cancer treatment, such as predicting patient outcomes and identifying optimal treatment strategies.

The candidate will focus on the following core tasks:

Data Exploration and Preprocessing: The candidate will gain experience in handling complex medical datasets by cleaning, preparing, and structuring the data to ensure it is suitable for advanced AI analysis.

Building AI Models: Using machine learning and deep learning techniques, the candidate will develop models aimed at predicting lung cancer progression, evaluating treatment efficacy, and understanding the impact of various environmental and genetic factors.

Interpretability and Explainability: A significant emphasis will be placed on making AI models interpretable and transparent. The candidate will explore techniques to ensure that the models produced are not just accurate but also explainable, providing healthcare professionals with clear insights into the model's predictions and decisions.

Exploring Interaction Networks: The candidate will analyze interaction networks, studying relationships between patient genetics, environmental factors, and treatment responses to identify key drivers of lung cancer outcomes.

Throughout the project, the candidate will not only gain hands-on experience with cutting-edge AI tools and methodologies but also develop a deeper understanding of AI's role in healthcare. This project provides an impactful opportunity to contribute to a field where AI innovation can directly improve patient outcomes.

The Ride-Sharing Problem (RSP) seeks to efficiently match travelers with similar itineraries and schedules on short notice, yielding substantial social and environmental benefits by reducing the number of cars used for personal travel and optimizing the use of available seat capacity in urban areas.

The static version of RSP is inadequate for real-world applications, as it does not account for the inevitable delays caused by traffic, cancellations, or new services. Therefore, a dynamic version is necessary to accommodate these circumstances.

This TFM proposes the development of a Hybrid Metaheuristic to address the Dynamic RSP, combining Metaheuristic optimization with Agent-based simulation. This optimization method must include the following features:

• Adaptation to traffic conditions.
• Integration of new services en route.
• Management of route cancellations.

The future of communication is taking us beyond traditional terrestrial networks. In the next decade, 6G networks will connect more than just smartphones-they will support flying cars and underwater robots using such high-tech enablers as lasers, satellites, drones, and high-altitude platforms (HAPs, flying at 20-40 km). To make this vision a reality, we need to tackle challenges like ensuring seamless connectivity for high-speed aerial vehicles, enabling underwater data transmission, and predicting large-scale network performance in urban environments. By working on these problems, you will be at the forefront of working on groundbreaking challenges in the rapidly evolving field of 6G and Non-Terrestrial Networks.  

The student will work on one of the following research objectives (or on a sensible combination of them):

- Performance modeling and optimization of communication links with satellites (such as starlink) and high-altitude platforms

- Development of advanced AI-driven beamsteering solutions to improve data transmission in underwater laser communication

- Design of an underwater-to-satellite (or HAP) communication system

- 6G-enabled 3D weather sensing and climate monitoring

- Usage of open big real-world data (e.g., OpenStreetMaps) to predict network coverage and performance in large urban areas, helping to optimize infrastructure deployment

- Design of 6G handover strategies for reliable connections with highly mobile (up to 300 kmh) flying cars/taxis

The aim is to enhance the capabilities of small UAVs by integrating RGB AI decks and exploring novel GenAI approaches for 3D object reconstruction. The project will involve developing and refining the UAV setup, conducting experiments with different GenAI techniques, and evaluating the performance across various use cases.

· Extend UAV systems with RGB AI decks for enhanced image capture and processing.

· Explore different Generative AI approaches to enhance the quality of 3D reconstructions.

· Evaluate the performance of the system across different object types and scenarios.

· Conduct experiments to compare reconstruction quality, processing time, and adaptability of dif-

ferent GenAI techniques.

Protein design has flourished in recent years thanks to the application of deep learning algorithms, culminating in the award of the Nobel Prize in Chemistry in 2024. The possibility of designing proteins with tailor-made functions ("à la carte") has enticing applications in drug design, diagnostics, and environmental science, such as the design of catalyists or biosensors. Unfortunately, linking structure to function is not trivial and often requires understanding protein flexibility and dynamics.

The state-of-the-art tool used to generate protein sequences for a given structure is ProteinMPNN, but it has been trained on crystallographic static structures and therefore poorly accounts for protein dynamics. Until 2025, it was impossible to efficiently generate large numbers of protein conformations to account for flexibility, but recent generative methods have proven both precise and effective at low computational cost. This opens the possibility of fine-tuning ProteinMPNN with ensembles of conformations associated with each sequence, thereby encoding protein dynamics into the designed sequences.

The goal of the present TFM is to:

1. Generate ensembles of protein structures with a generative model (such as BioEmu).

2. Fine-tune ProteinMPNN using data augmentation, associating each protein sequence not with a single structure (as done so far) but with an ensemble of structures.

By linking protein dynamics to sequence, this work will represent a significant step forward in the design of de novo functional proteins.

Both BioEmu and ProteinMPNN are open-source and available on GitHub, along with training instructions. The resulting code will also be deposited in GitHub for use by the scientific community.