Joaquin Vanschoren

Auto-continual learning

Automated Machine Learning (AutoML) and Continual Learning (CL) have emerged as pivotal research domains in advancing artificial intelligence. AutoML seeks to automate the design of neural architectures and machine learning pipelines, yet this process is often computationally expensive. Conversely, CL aims to enable models to learn incrementally from new data streams without forgetting what we learned before, yet often relies on fixed predefined architectures. A synergistic integration of AutoML and CL offers a compelling pathway to surmount both limitations, resulting in AI systems capable of efficient lifelong adaptation by dynamically modifying not only their model weights but also their core architecture as needed. Such architectural plasticity is crucial for model to genuinely evolve in response to novel tasks and changing data distributions. This lecture will survey the state-of-the-art in both foundational fields, spanning large-scale and resource-constrained models. Furthermore, we will present recent breakthroughs at their intersection, demonstrating how this synthesis helps us realize more autonomous AI, democratizing its application, and accelerating scientific discovery.

Biography

I am an associate professor at TU Eindhoven and head of the Artificial Minds research group. We aim to scientifically understand and build AI systems with advanced capabilities, and make AI accessible to benefit all of humanity.

I founded OpenML, a useful open science platform for machine learning, to streamline and accelerate reproducible AI research. I was the inaugural chair of the NeurIPS Datasets and Benchmarks track and am editor-in-chief of the DMLR journal, to incentivize and reward good data and evaluation practices in AI. I also co-chair the MLCommons AI Risk & Reliability working group to help make AI models safer through science, am one of the founders of the Croissant standard for sharing AI resources, and contribute to data-centric AI.

I am a founding member of the European AI societies ELLIS and CAIRNE, authored the first book on AutoML, gave tutorials at NeurIPS and AAAI, won several awards (including the Dutch Data Prize and Amazon Research Award), and I’ve been interviewed for news articles in Nature, Science, and podcasts. I am always eager to collaborate with new people. Do reach out!