Skip to main content

Integrating domain knowledge with machine learning in robotics and molecular science

Save to calendar

Mar 18

Date and time: 18 March, 12 pm – 1 pm
Speaker: Anastasia Varava
Title: Integrating domain knowledge with machine learning in robotics and molecular science

Meeting ID: 644 1767 7978
Password: DF2020

Watch the recorded presentation:


Picture of Anastasia VaravaAbstract: Modelling states of the system and transitions between them is fundamental for applications in various fields, including robotics and molecular science. Currently, the so-called “end-to-end” design philosophy which seeks to directly learn representations from data while avoiding explicit structure and “feature engineering” is popular in machine learning. While this approach has clear benefits, it heavily relies on the availability of large amounts of cheap data, which is often problematic in some application fields, such as robotics and natural sciences. Furthermore, in many of these fields, the structure of the problem is often well understood; when incorporated in machine learning algorithms, it can improve data efficiency, interpretability, and overall performance.

In her talk, Anastasia Varava will cover her prior work on analytical state representations in robotics and organic chemistry, as well as her current work and ideas on how these methods can be combined with machine learning. The goal here is to incorporate a reasonable amount of domain knowledge while creating models that are general enough to be suitable for a wide range of scenarios and do not rely on hand-crafted features.

Bio: Anastasia obtained her PhD in Computer Science from KTH in 2019 and is currently working as a postdoctoral researcher at the Robotics, Perception, and Learning lab, KTH. She is interested in applying concepts and tools from Geometry, Topology and Machine Learning to designing efficient representations for complex systems in various fields, ranging from Robotics to Natural Sciences.