Colloquium Computer Science - Dr. Nicola Strisciuglio, University of Twente

Title: Beyond Scaling: Toward Data-Efficient and Reliable Vision Models

Abstract:

The recent success of computer vision foundation models has largely been driven by aggressive scaling of data, model size, and compute. While effective, this paradigm comes at a steep cost: high energy consumption, poorly controlled training data, embedded biases, and limited accessibility. This trajectory is increasingly difficult to justify.

In this talk, I argue for a shift from brute-force learning: data efficiency and reliability as core research objectives. will present empirical findings that reveal latent compositional structures in vision–language models, alongside persistent shortcut-driven biases that affect generalization abilities of current models. These observations motivate open research challenges: moving beyond post-hoc analysis toward training strategies that explicitly promote compositional representations and bias-aware learning. Building on work on prior knowledge in vision, I envision to lift priors, compositionality and bias-awareness into self-supervised learning objectives, embedding structure and similarity at a level that remains general, data-efficient, and scalable.

This perspective challenges prevailing scaling assumptions and opens new pathways toward vision foundation models that are easier to train and curate, and more transparent in what they learn. Ultimately, sustainability in AI is not only a hardware problem, but a learning problem: rethinking how models learn is key to building vision systems that are robust, efficient, and broadly accessible.

Short Bio:

Dr. Nicola Strisciuglio is an Associate Professor in Machine Learning and Computer Vision at the University of Twente, where he leads research on robust, data-efficient, and trustworthy AI systems in the Computer Vision Lab. His work focuses on improving the generalization of deep learning models, mitigating bias, and incorporating prior knowledge into computer vision algorithms. He obtained his PhD cum laude in 2016 from the University of Groningen and the University of Salerno and has made several contributions to the fields of computer vision, machine learning, and robotics. He is the recipient of an NWO Vidi grant in 2025, serves as Associate Editor of Pattern Recognition, and covers Area Chair roles in major AI conferences, including NeurIPS.