Research

The Amsterdam Machine Learning Lab (AMLab) conducts research in the area of large scale modelling of complex data sources. We develop deep generative models, methods for approximate inference, probabilistic programming, Bayesian deep learning, causal inference, reinforcement learning, graph neural networks, and geometric deep learning. Our predominant question is how to improve generalizability of machine learning algorithms. We mainly focus is on three research questions:

1. How to handle uncertainty?

In the field of machine learning, uncertainty is important because it allows us to measure how confident a model is in its predictions: where does the model perform well and where does it not? Uncertainty can be estimated using a variety of techniques of which Bayesian inference is an important focus point in our lab. By taking uncertainty into account, we can build more robust and reliable machine learning models.

2. How to handle geometry and construct physics-grounded AI systems?

Almost all of our data has some form of grounding in our physical world, and many AI systems demand that we respect such structure. For example, in physics-related tasks one can rely on the principles of physics, and in healthcare there are cases where a tumor should be considered equally malignant regardless of where and at which orientation it appears in the body. Many real-world applications demand certain geometric constraints to be satisfied. A more fundamental question is how to build AI systems that understand geometry and physics, and are capable of reasoning about it?

3. How to uncover or model the causal relations underlying observations?

The notion of causality is important in machine learning to understand the underlying relationships between different variables. This allows us to build more accurate and reliable models and it can also help us to identify potential biases or confounding factors in our data.

The results from our lab are widely applicable, particularly in the field of AI-assisted sciences. We make a concrete impact in domains such as computational material science, chemistry and physics. We have also partnered with clinicians to focus on reliable quantification of disease risk factors and the development of predictive systems for cancer treatment planning. Our research has the potential to revolutionize the way we approach many scientific fields.

Facts & figures

Our lab has produced some of the most cited papers in our field: Adam: A Method for Stochastic Optimization by Kingma and Ba (127,633 citations). Furthermore, seminal papers from our group ignited the rise or formation of AI sub-disciplines such as variational inference (with VAEs) and geometric deep learning (graph and group-equivariant NNs).

In recent years NWO Veni grants have been won by: Erik Bekkers (link), Eric Nalisnick (link) and Jamie Townsend (link). PhD student Sindy Löwe won a Google Research Fellowship (link).

Partnership & collaborations

AMLab collaborates with a number of private and public partners, such as Microsoft Research, Qualcomm, Bosch, Philips, Janssen Pharmaceutica and Amsterdam University Medical Centre. Mostly we collaborate on fundamental AI research with applications in structural biology, material science, computational physics, computer vision, and more.

Education

AMLab teaches the following courses: Machine Learning 1, Machine Learning 2, Reinforcement Learning, Deep Learning 2, Leren, Bayesian Statistics for Machine Learning

With our education we aim to provide a solid foundation in machine learning and equip our students with the skills to keep up with the state of the art in machine learning. Our students are in an optimal position to perform research in machine learning and are also well-equipped to contribute to cutting-edge technological solutions in industry. In our approach to education, we further believe it is important to keep up with technological developments and embrace the hybrid format of online and on-site teaching. Several of our courses have produced successful publicly available video lectures (see e.g. Machine Learning 1 (2020) by Erikk Bekkers and Group Equivariant Deep Learning (UvA - 2022) by Erik Bekkers), and more will follow.

Future mission

Over the years, the deep learning field has shown that for many tasks scaling up to larger models is sufficient to make more impact. However, when neural networks need to deal with the physical world, they often fail. This failure can be overcome if we equip them with the right (geometric) structure and inductive biases, which is the focus of our lab. It may be unrealistic to expect a breakthrough in the sense of achieving truly generalizable reasoning capabilities with AI, but we can certainly expect an increasing number of convincing (simulated) reasoning capabilities in a wide range of environments. That would make many machine learning applications more widely applicable in a reliable way.

AMlab positions itself in the research theme AI with clear links to Computional Science and Data Science.