A theory of inference and learning in cortex with spiking neurons and dendritic error computation
by Fabian Mikulasch
Date of Examination:2023-10-16
Date of issue:2023-11-03
Advisor:Prof. Dr. Viola Priesemann
Referee:Prof. Dr. Alexander Ecker
Referee:Prof. Dr. Siegrid Löwel
Referee:Prof. Dr. Theo Geisel
Referee:Prof. Dr. Fred Wolf
Referee:Dr. Andreas Neef
Files in this item
Name:PhD_Thesis.pdf
Size:11.8Mb
Format:PDF
Abstract
English
How the cortex performs its intricate computations, and how it adapts to the world around it, is one of the central mysteries in neuroscience. It is a longstanding belief that one of the main aims of the brain, and especially the cortex, is to infer the states of the world, such as the presence of objects, that underlie the sensory observations an animal makes. Currently the most discussed theory that formalizes this idea and proposes a biological implementation is classical hierarchical Predictive Coding (hPC), which hypothesizes the existence of dedicated ’error neurons’ in cortex that signal errors of the internal model. While this theory has inspired much research, it is not clear how one of its central elements—the proposed learning algorithm—can be implemented with spiking neurons, which questions its biological plausibility. In this thesis we propose an alternative theory of learning and inference with spiking neurons, where errors are computed in neural dendrites, and synaptic connections are learned with biologically plausible voltage-dependent plasticity rules. We first build on existing work of inference and learning in spiking neural networks, and show how dendritic error computation can overcome an unsolved problem for learning in these networks. Specifically, when neural activity in the network is correlated, previously assumed Hebbian-like learning leads to pathological network activity, which learning with dendritic errors prevents. We then combine this model with other theories of learning in cortex to a theory of hierarchical inference with spiking neurons, and show that this theory is isomorphic to classical hPC while overcoming its biological implausibility. Last, we employ our framework to explain how ’mismatch responses’, i.e., neural responses that signal the mismatch between an internal model and observations, emerge from inference and learning in cortex. Together, this work proposes a comprehensive theory of learning, inference and their signatures in cortex, and provides a range of readily testable predictions.
Keywords: Hierarchical Inference; Voltage-dependent plasticity; Mismatch responses; Predictive processing; Dendritic computation