Abstract
Nervous systems may have originated in the late Ediacaran period to rapidly detect “critical states” of the first predators (or prey) being within striking distance, likely by virtue of electroreception. However as motility and these critical distances increased, the acuity of early electroreceptors heightened in kind to detect weaker remote signals, giving rise to an inference problem of determining the true state of the environment using noisy sensory data. The statistically-optimal means of doing so lies in Bayesian inference; a computationally expensive form of probabilistic reasoning for which neuronal membranes and populations appear naturally-suited. As such, this thesis focuses on the goal of studying, extending and implementing existing models of neural inference in the context of electroreception. In doing so, the findings may not only contribute to understanding how and why neurons represent information probabilistically, but also serve as a biomorphic engineering solution for improving cheap, noisy sensors.
Firstly, I constructed a simple and interactive computational model of inference based on the system-level activation caused by ion flow in a Trichoplax-like Placozoan, in the context of low-motility, pre-Cambrian predator-prey interaction. With this, I demonstrated how the direction and proximity of an electric dipole can be inferred using simple biological building blocks and diffusing chemical gradients to form a “particle filter”; a generalised form of Bayesian inference wherein particles (ions) approximate the distribution at each time step and are used as priors to create the probability distributions in the subsequent time step. Then, I translated this model into a modern, high-motility, experimentally-based generative model; using the simulated output of a cheap Hall Effect sensor to infer the location of a magnetic dipole. Finally, I extended the preceding work into a physical and functional model of neural inference. In this final model, analogue Hall voltages were translated into spikes, which were used as particles in a map of neurons to infer the true state of the magnetic dipole in one dimension, in real time, and at distances and levels of accuracy beyond the typical limits of a Hall Effect sensor.