Towards modelling of autonomous neuromorphic learning systems

This thesis aims to investigate physically plausible models of spiking neurons and propose a path for autonomous molecular implementation.

First, I will discuss supervised learning of multi-class stimuli in a single spiking neuron. Particularly, I will focus on the aggregate-label learning framework originally proposed by Gutig (2016). To this end, I will introduce a novel model of a spiking neuron capable of performing complex computational tasks, while remaining simple and readily interpretable. Moreover, I will demonstrate how this neuronal model can be interpreted as a chemical reaction network, and how synaptic weights can be encoded by reaction rate constants.

Next, I will investigate a minimal molecular model of a spiking neuron capable of unsupervised learning. In order to be practically useful, such molecular implementation needs to be autonomous. I will define what it means for the learning systems to be autonomous, and propose a model which implements both the neuronal functions as

well as the learning algorithm within a chemical reaction network. Through extensive simulations I will demonstrate that this model is capable of autonomous recognition of frequency biases and temporal correlations embedded into discrete spike trains.

Lastly, I will present an implementation of a spiking neuron based on DNA-strand displacement interactions. The advantage of this method is that it can realistically be synthesised in a laboratory. The DNA neuron will be shown to be capable of performing a variety of computational tasks including temporal correlation learning and novelty detection.