Lukas Muttenthaler

Logo

PhD student in Machine Learning at the Department of Software Engineering and Theoretical Computer Science, TU Berlin, working on using human inductive biases to align and improve neural network representations, with a focus on low data regimes.

View My GitHub Profile

Research Profile

I am a (full-time) Senior ML Research Scientist at Aignostics and a (part-time) Senior Researcher in the Explainable Machine Learning Group at TU Munich, working on model merging, representation merging, and post-training approaches in Computer Vision. Prior to that, I was a Student Researcher at Google DeepMind and a PhD student in Machine Learning at TU Berlin and the Berlin Institute for the Foundations of Learning and Data (BIFOLD). Throughout most of my PhD I have also been a guest researcher in the ViCCo Group at the Max Planck Institute for Human Cognitive and Brain Sciences. I was mainly advised by Klaus-Robert Müller (TU Berlin) and co-supervised by Martin Hebart (MPI), Simon Kornblith (Anthropic), and Andrew Lampinen (Google DeepMind). During my PhD, I've been part of a one-year Research Collaboration between TU Berlin and Google Brain, where I was advised by Simon Kornblith. Previously, I was a MSc student in IT & Cognition / Computer Science of Isabelle Augenstein and Johannes Bjerva at the University of Copenhagen where I mostly worked on Question Answering and Machine Translation.

My research mainly revolves around representation learning in computer vision. In particular, I try to understand the factors that influence the degree of alignment between human mental and neural network representations and use inspiration from human cognition to improve deep learning models. My goal is to build interpretable (vision) foundation models that generalize to downstream out-of-distribution settings (similar to how the human brain does); something that we partly achieved here.

Occasionally I dabble in philosophical discussions about representational alignment and try to develop common language across research disciplines together with other people in the field. Recently, I've been thinking a lot about the transferability of representational similarities across datasets. Have a look at my Google Scholar for more information about my work. Feel free to reach out to me, if you believe our research intentions are aligned (pun intended) and you are keen to collaborate on a project.

Academic Background

Industry Roles

Service

Conference & Workshop Organization

Reviewing

News

2025

2024

2023

2022

2021

2020

2019

2018