Faces of Research” series by PostGenAI@Paris

Date
21 May 2026

Teaching AI to Quantify Uncertainty for a Safer World

As part of our Faces of Research series, we highlight the researchers who are driving innovation in artificial intelligence for the benefit of society. Today, we feature Juliana Carvalho, whose PhD explores how Artificial Intelligence can become more robust, reliable, and aware of its own uncertainty.

Working at the intersection of machine learning, multimodal AI, and uncertainty estimation, Juliana’s research addresses one of the biggest challenges facing modern AI: how can we ensure that intelligent systems behave safely when confronted with unfamiliar or unexpected situations?

Advancing Robust Self-Supervised Learning

Juliana’s thesis focuses on robust self-supervised learning (SSL) for multimodal data. Self-supervised learning allows AI systems to learn meaningful representations from large amounts of unlabeled data, reducing the need for costly manual annotation.

Her work investigates advanced techniques such as contrastive learning, enabling AI models to align and process multiple data modalities — including images, text, and sensor data — in a coherent and efficient way.

A central aspect of her research is robustness: ensuring that AI systems remain reliable when dealing with ambiguous, noisy, or out-of-distribution (OOD) data that differs from what they encountered during training.

Her PhD is conducted under the supervision of:

  • Arnaud Breloy
  • Javiera Castillo Navarro
  • Clément Rambour

Why Uncertainty Matters in AI

One of the key goals of Juliana’s research is to develop methods that allow AI systems to quantify their own uncertainty.

In high-stakes domains such as medical diagnosis, autonomous driving, or Earth observation, AI models can encounter situations they were never trained for. Without reliable uncertainty estimation, these systems may produce overconfident yet incorrect predictions.

By enabling models to recognize when they do not “know,” Juliana’s work contributes to safer and more trustworthy AI systems capable of supporting critical real-world decisions.

Scaling AI for Earth Observation

Building on her expertise in multimodal learning, Juliana is currently working on integrating satellite imagery and metadata within large-scale self-supervised frameworks.

Her research now focuses on scaling these models using massive Earth Observation datasets such as fMoW (Functional Map of the World). This stage is fundamental for the next phase of her PhD, where she plans to investigate more advanced architectures, including mixtures of models designed to improve robustness and adaptability.

These developments could have important implications for environmental monitoring, climate analysis, disaster response, and many other applications relying on large-scale geospatial intelligence.

In Her Own Words 💬

“What drives me right now is the technical challenge of scaling self-supervised systems on large-scale data to ensure they are truly robust. This groundwork is essential for my next steps, where I will refine how AI handles uncertainty, ensuring it remains a reliable tool for complex real-world applications.”

— Juliana Carvalho

A Collective Research Effort

This work is part of a broader collaborative ecosystem involving leading French research and academic institutions, including: