Disputation: Constantin Venkov Bechkov

PhD candidate Constantin Venkov Bechkov at the Department of Biosciences, Faculty of Mathematics and Natural Sciences, is defending the thesis 'Neural Representations and the Topology of Cognition' for the degree of PhD.

The defense will also be live streamed using Zoom. The host of the session will moderate the technicalities while the chair of the defence will moderate the disputation.

The event opens for participation just before it starts, and closes for new participants approximately 15 minutes after it has begun.

Join the event here

Trial lecture: 'Dimension reduction of higher dimensional data'

Time and place: Sep. 10, 2024 10:15 AM, Nucleus, Bikuben, Kristine Bonnevies hus

Main research findings:

Picture of the candidate
Constantin Venkov Bechkov, Photo: Private

The question of how minds and brains interact with the external world has puzzled humanity for thousands of years. Following recent advancements in computational and neural recording technology we are now able to study the activity of populations of thousands of neurons. Despite our recent ability to illuminate this “black box”, this abundance of data has led to the need for more interpretable models.

In this thesis, tools from topology are applied to the analysis of neural networks, with the goal of understanding the way in which they transform manifolds of stimuli. This transformation can be described by the way in which manifold homology changes during propagation through the layers of a network. The thesis extends persistent homology by using geodesic distances, with which we reliably discover that rotating images induce a circular topology in the activity of primary visual cortex. We further discuss the experimental conditions for this analysis to be successful. Following this, we introduce the idea of a rank decomposition, which allows us to study the topology of networks with piecewise-linear activation functions and show that trained networks are highly topologically destructive. Finally, we study how the toroidal structure of a population of grid cells is determined by the distribution of their phases.

Illustration math
Calculations establishing the relations between neural network decompositions and the relative homology groups of neural representations. Illustration by Constantin Venkov Bechkov.

Adjudication committee

Prof. Kenneth Harris, University College London

Prof. Nils Baas, Norwegian University of Science and Technology (NTNU)

Prof. Eivind Valen, Department of Biosciences, University of Oslo

Chair of defense

Assoc. Prof. Jonas Paulsen, Department of Biosciences, University of Oslo

Supervisors 

Prof. Gaute Einevoll, Norwegian University of Life Sciences & University of Oslo

Prof. Marianne Fyhn, Department of Biosciences, University of Oslo

Prof. Anders Malthe-Sørenssen, Department of Physics, University of Oslo

Candidate contact information

LinkedIn

Published Aug. 27, 2024 11:39 AM - Last modified Sep. 9, 2024 9:57 AM