Jan Bauer
Jan Bauer

PhD student in Theoretical Neuroscience

    How do the numerous components of neural circuits reliably implement complex function?

    I am approaching this questions by developing models that capture simple aspects of cognition, but are still easy to understand. I believe that bridging perspectives from both artificial and biological intelligence brings great synergies, as both of these fields fundamentally have the shared goal of understanding neural systems.

    I am approaching this questions by developing models that capture simple aspects of cognition, but are still easy to understand. I believe that bridging perspectives from both artificial and biological intelligence brings great synergies, as both of these fields fundamentally have the shared goal of understanding neural systems.

    Education
    • PhD student in Theoretical Neuroscience, 2022–

      Gatsby Unit, UCL & ELSC, Hebrew University

    • MSc in Statistical Physics, 2019–2022

      Juelich Research Center, Germany

    • BSc in Physics, 2016–2019

      RWTH Aachen University, Germany

    Blog

    Culinary choices

    Suppose you are in a new city and looking to get dinner with friends. You pull out Google Maps to help with the decision. To make the process easier, you decide to look at the ratings: A nearby Cambodian restaurant boasts 4.8 stars, but the equally close Italian place is a close competitor at 4.6 stars, but has five times the number of reviews. Surely that must make a difference? You search for your scratchpad, promising to your friends that you got the situation.

    Can random actions be optimal?

    Is random behavior helpful in any situation? By definition, random actions are the most uninformed, and if any better is known should be suboptimal. Yet, the issue is more subtle. Reinforcement learning and game theory can be paradigms to reason about this.

    Do auto-regressive models bite their own tail?

    Autoregressive models use their output to arrive at predictions. In machine learning, this amounts to “training on the output”, i.e., generated data. More broadly, intelligent behavior is often accompanied by deep thought or even dreaming between actions. In both of these cases, the system is decoupled from the ground truth. Despite this apparent conundrum, there seems to be a benefit.

    What is meta in meta-learning?

    Meta-learning summarizes the concept of learning a more general framework to learn – hence the name. Yet, this concept subsumizes a range of multiple concepts, including transfer learning, few-shot learning, continual learning, and fine-tuning. We develop an abstracted framework that unifies these notions. This extends beyond parametric models.

    Mean-field decoupling via auxiliary variables

    In statistical physics, we are often dealing with systems that comprise many components. In order to calculate their statistics, high-dimensional integrals over those variables $\boldsymbol{x}\in\mathbb{R}^{N}$ with $N\gg1$ are required. A typical form is