Identification of locust neuronal systems with recurrent neural networks

Abstract

System identification (SI) approaches have been widely used in the neurosciences for studying neuronal behaviour. In producing statistical models of neuronal data for SI, most approaches choose to use models which are easily interpretable. Interpretable modelling allows researchers to efficiently derive the implied dynamics of the neuronal system(s) under study and gain insight into its function.
The recent increase in computational power and availability higher resolution experimental techniques have allowed researchers in the neurosciences to study increasingly complex neuronal systems. This thesis suggests that some of the widely used interpretable SI models may be inadequate for modelling the growing corpus of big and complex datasets.

Recently, there has been a surge of academic interest in using Artificial Neural Networks (ANNs) for modelling complex datasets. ANNs are one of the hot focii in the machine learning literature, and are increasingly being used in other fields as a tool for modelling, control and classification. The main objective of this thesis is to use a variety of ANNs called Recurrent Neural Networks (RNNs) for modelling time-series datasets derived from neuronal systems.

In achieving this objective, this thesis investigates the efficacy of RNNs in modelling a series of neurons in the locust hind leg. First, the graded potentials of locust Fast Extensor Tibiae (FETi) motor neuron are modelled. The performance characteristics of the trained RNNs are compared with Linear-Nonlinear-Linear (LNL) and Time-Delay Neural Networks (TDNNs) previously used for the FETi modelling task. Next, RNN modelling techniques are adapted for modelling the spiking action potentials of locust proprioceptive afferents. The functionality of the afferents are investigated using the trained models and compared to results achieved by previous work using Wiener methods. Finally, RNNs are used to model the dynamics of the locust Slow Extensor Tibiae (SETi) motor neuron. The trained models are used to simulate the functionality of SETi, and the results are used to compare the efficacy of using open-loop against closed-loop setups in electrophysiological experiments.

Although the techniques developed in this thesis are used and validated in modelling simple insect neuronal systems; this thesis highlights the general issues and pitfalls associated with using RNNs as a tool for system identification. The techniques used in this thesis can be used in modelling any neuronal system, given enough data and computational capacity.

More information

Identifiers

ORCID for Philip Newland: orcid.org/0000-0003-4124-8507

Catalogue record

Export record