From bottom-up to top-down: a computational exploration of plasticity in neural networks of the brain
File(s)
Author(s)
Kaleb, Klara
Type
Thesis
Abstract
Neural plasticity, the brain’s remarkable capacity for adaptation and learning, is essential for acquiring new skills and knowledge. Computational modelling plays a crucial role in under- standing the intricate mechanisms of neural plasticity, bridging the gap between experimental findings and theoretical insights. This thesis explores three computational approaches to ad- dress challenges in understanding biological learning:
Network Homeostasis. The first part of this thesis proposes a novel form of inhibitory plas- ticity that enables global control of network activity, going beyond neuron-centered homeosta- sis. The model proposed can explain how neural representations can remap while maintaining network stability, both in the hippocampus and the cortex.
Temporal Credit Assignment. The second part of this thesis investigates the mechanisms of a biologically plausible model for rapid motor adaptation. This work demonstrates that feed- back control allows accurate and efficient learning in recurrent neural networks, even without a computationally expensive separate relaxation phase.
Spatial Credit Assignment. Using a meta-learning framework, the third part of this thesis explores the space of effective, biologically plausible feedback learning rules in cortical multi- layer networks. A teacher-student network setup is introduced to evaluate credit assignment, leading to the discovery of a novel feedback learning rule that demonstrates successful, biolog- ically plausible learning.
By tackling these challenges, this thesis contributes to a deeper understanding of learning in the brain, paving the way for more effective computational models.
Network Homeostasis. The first part of this thesis proposes a novel form of inhibitory plas- ticity that enables global control of network activity, going beyond neuron-centered homeosta- sis. The model proposed can explain how neural representations can remap while maintaining network stability, both in the hippocampus and the cortex.
Temporal Credit Assignment. The second part of this thesis investigates the mechanisms of a biologically plausible model for rapid motor adaptation. This work demonstrates that feed- back control allows accurate and efficient learning in recurrent neural networks, even without a computationally expensive separate relaxation phase.
Spatial Credit Assignment. Using a meta-learning framework, the third part of this thesis explores the space of effective, biologically plausible feedback learning rules in cortical multi- layer networks. A teacher-student network setup is introduced to evaluate credit assignment, leading to the discovery of a novel feedback learning rule that demonstrates successful, biolog- ically plausible learning.
By tackling these challenges, this thesis contributes to a deeper understanding of learning in the brain, paving the way for more effective computational models.
Version
Open Access
Date Issued
2024-05
Date Awarded
2024-10
Copyright Statement
Creative Commons Attribution NonCommercial Licence
License URL
Advisor
Clopath, Claudia
Sponsor
Wellcome Trust (London, England)
Grant Number
219995/Z/19/Z
Publisher Department
Bioengineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)