PhD on AI for Safety-Critical Multi-modal Learning

Are you eager to make a difference in the advancement of AI via reliable state-of-the-art deep learning models? This position will explore efficient multi-modal models for safety-critical applications based on robustness guarantees and explainable AI built on insight into learned representations.

Information
We are seeking a highly motivated PhD student to join our research team in an ambitious project at the intersection of Safe AI, Resource Constrained, and Multi-Modal Learning. The focus of this PhD is on developing novel state-of-the-art AI models for safety-critical applications in which resilience, autonomy, and intelligence are required.

The overarching goal of this project is to develop efficient, trustworthy AI models that have a robust understanding of their environment based on various data sources. For example, the model should be able to integrate camera, radar, and other sensor modalities. Particular attention will be paid to transformer- and post-transformer architectures, and how to adapt them to learn efficiently on resource-constrained hardware, for instance with transfer learning and quantization techniques.

A key challenge of this project will be to design novel transformer-based architectures and adaptation techniques that are both efficient and reliable. The latter will require you to investigate and ultimately control the latent representations of knowledge in the model. This can be tackled from the viewpoint of various areas of machine learning, such as disentanglement of features, out-of-distribution awareness, robustness to adversarial attacks, and explainability, including the development of formal robustness guarantees that provide insights into the fundamental characteristics of the model that contribute to its safety and reliability in real-world applications. You will also assist in the integration and experimental evaluation of these techniques. Explainability will play an important role here, ensuring that models validate their explanations and provide insight into their latent representations, identifying what the model deems relevant for its predictions.

An auxiliary goal of this PhD is to make the employed models resource-efficient, ensuring they can run effectively on embedded accelerators (e.g., TPUs, NPUs, FPGAs) in constrained environments. The research will explore techniques such as model compression and quantization to optimize AI models for real-time operation on embedded platforms.

Do you recognize yourself in this profile and would you like to know more? Please contact the hiring manager Joaquin Vanschoren, Associate Professor, j.vanschoren@tue.nl.

Visit our website for more information about the application process or the conditions of employment. You can also contact Sibylle Hess, Assistant Professor, s.c.hess@tue.nl.

Are you inspired and would like to know more about working at TU/e? Please visit our career page.