Preise der Akademie
„New contributions in the fie/d of machine /earning"
A new frontier in machine learning research is to learn causal representations from
data that uncover underlying physical properties and natural laws. My research
aims to learn these causal representations through rigorous theoiy and scalable
algorithms. Unlike traditional machine learning techniques that only analyze cor-
relations, causal representations provide a deeper understanding of the data and
enable reasoning about the consequences of actions and events that we have not
obsei~ved. Overall, they contribute to the broader goal of developing robust, exp-
lainable, and fair machine learning models that can also inform human decision-
making.
Aspecific area ofinterest in my research is disentangled representations, which
aim to identify independent factors ofvariation in a high-dimensional dataset com-
posed of independent samples. Our award-winning work explored the theoretical
and practical challenges of this task. Against prior belief, we found that this is the-
oretically impossible without any access to supei~vision or additional assumptions,
which we also confirmed through a large-scale empirical study. As a way forward,
we advocated for incorporating the natural assumption that real-world data is ne-
ver independently sampled but is subject to temporal structure and allows for ac-
tions. These concepts are now commonly used in causal representation learning
research and are continuing to find important applications in robotics, healthcare,
physical sciences, and engineering.
383
„New contributions in the fie/d of machine /earning"
A new frontier in machine learning research is to learn causal representations from
data that uncover underlying physical properties and natural laws. My research
aims to learn these causal representations through rigorous theoiy and scalable
algorithms. Unlike traditional machine learning techniques that only analyze cor-
relations, causal representations provide a deeper understanding of the data and
enable reasoning about the consequences of actions and events that we have not
obsei~ved. Overall, they contribute to the broader goal of developing robust, exp-
lainable, and fair machine learning models that can also inform human decision-
making.
Aspecific area ofinterest in my research is disentangled representations, which
aim to identify independent factors ofvariation in a high-dimensional dataset com-
posed of independent samples. Our award-winning work explored the theoretical
and practical challenges of this task. Against prior belief, we found that this is the-
oretically impossible without any access to supei~vision or additional assumptions,
which we also confirmed through a large-scale empirical study. As a way forward,
we advocated for incorporating the natural assumption that real-world data is ne-
ver independently sampled but is subject to temporal structure and allows for ac-
tions. These concepts are now commonly used in causal representation learning
research and are continuing to find important applications in robotics, healthcare,
physical sciences, and engineering.
383