From natural statistics to neural codes: understanding the principles of proprioceptive information processing
File(s)
Author(s)
Grogan, Max Douglas
Type
Thesis
Abstract
Proprioception, our sense of body position and movement, is fundamental for motor control yet remains one of the least understood sensory modalities. This thesis investigates the computational principles underlying proprioceptive processing across multiple scales of neural organisation through novel machine learning approaches. The research addresses three key questions: How do natural movement statistics shape proprioceptive encoding in cortical neurons? What principles govern the spatial organisation of proprioceptive representations? And how does peripheral gain modulation through the fusimotor pathways contribute to sensorimotor control?
The first investigation develops a spike-based variational autoencoder framework to examine how efficient coding principles manifest in proprioceptive neurons. Training on naturalistic human movement data, this model reproduces key response properties of somatosensory cortical neurons without direct fitting to neural recordings, including their characteristic directional tuning and systematic biases aligned with biomechanical constraints. The second study incorporates biologically-inspired lateral connectivity into the computational architecture to generate testable predictions about the fine-scale spatial organisation of proprioceptive representations beyond current experimental resolution, which are then validated against neural recordings from somatosensory cortex of non-human primates.
The final investigation develops a dual-network architecture incorporating both alpha and gamma motor systems to investigate how fusimotor modulation of sensory feedback gains shapes sensorimotor control. This model demonstrates three crucial capabilities absent from traditional frameworks: predictive sensory encoding matching the +160ms horizon observed in human muscle spindles, preparatory sensory processing preceding movement onset, and enhanced robustness to noise. Together, my findings suggest that preparing to sense may be as fundamental to motor control as preparing to move.
The first investigation develops a spike-based variational autoencoder framework to examine how efficient coding principles manifest in proprioceptive neurons. Training on naturalistic human movement data, this model reproduces key response properties of somatosensory cortical neurons without direct fitting to neural recordings, including their characteristic directional tuning and systematic biases aligned with biomechanical constraints. The second study incorporates biologically-inspired lateral connectivity into the computational architecture to generate testable predictions about the fine-scale spatial organisation of proprioceptive representations beyond current experimental resolution, which are then validated against neural recordings from somatosensory cortex of non-human primates.
The final investigation develops a dual-network architecture incorporating both alpha and gamma motor systems to investigate how fusimotor modulation of sensory feedback gains shapes sensorimotor control. This model demonstrates three crucial capabilities absent from traditional frameworks: predictive sensory encoding matching the +160ms horizon observed in human muscle spindles, preparatory sensory processing preceding movement onset, and enhanced robustness to noise. Together, my findings suggest that preparing to sense may be as fundamental to motor control as preparing to move.
Version
Open Access
Date Issued
2024-11-28
Date Awarded
2026-02-01
Copyright Statement
Attribution-Non Commercial-No Derivatives 4.0 International Licence (CC BY-NC-ND)
Advisor
Faisal, A Aldo
Sponsor
Wellcome Trust (London, England)
Publisher Department
Department of Bioengineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)