Chaotic dynamics with neuromorphic networks of spiking neurons context and motivation

The Institute of Neuroinformatics (INI) is a joint institute of the University of Zürich and the Swiss Federal Institute of Zürich (ETH Zürich). INI carries out experimental, theoretical, and applied research with the aim of discovering key principles by which the brain is built and works, and using the knowledge practically.
The Neuromorphic Cognitive Systems group at INI has been developing neuromorphic processing platforms that enable event-driven computing using algorithms and primitives inspired by biology. In this project, we aim to explore the dynamics of hardware implementations of spiking neural networks, built using analog circuits to faithfully emulate the behaviour of real synapses and neurons in cortical circuits.
The project is broadly about leveraging the nonlinear nature of the neuromorphic circuits to obtain chaotic dynamics. This represents a new usage paradigm for these types of neuromorphic processors, and is particularly interesting because chaotic dynamics are observed pervasively in nature, from pendula and weather to isolated axons through entire brains. The synchronization of chaotic systems allows generating rich patterns from the interactions of relatively simple network configurations (e.g., communities, modules, and remote interactions can arise spontaneously among neurons interlinked as a ring).

Goals
Here we propose three lines of research that could be combined in a single long (e.g., MsC thesis) project, or kept as separate shorter term (e.g., Semester) projects:
1) the design of a spiking neural network that can spread the information from a small number of high-rate spiking inputs over a large number of neurons, with each neuron firing at much lower rates (i.e., for converting the temporal information present in high frequency signals to a spatio-temporal code distributed among multiple low-frequency nodes).
2) the implementation in a neuromorphic spiking neural network the results obtained in rings of single-transistor oscillators. Namely, diverse network designs comprising one or multiple neurons, connected via excitatory and inhibitory synapses. The aim is to find a valid configuration alongside "critical" values of the control parameters, such that the symmetry of the ring is spontaneously broken, and populations of neurons firing in sync emerge spontaneously. It is expected that, based on the settings, the size and distributions of these populations should be easily controllable, reproducing observations of multiscale modularity.
3) the design of "neural resonators" built by coupling populations of neurons such that they can reproduce the properties of harmonic oscillators. The goal is to explore how such neural primitives can be configured to selectively respond to input signals that have periodic characteristics (e.g., ECG signals, motor vibrations).

All sub-projects have important implications for the design of low-power neural processors that could potentially be used for bio-signal processing (e.g. for prosthetic control or epileptic seizure detection) or in more industrial settings (e.g., for predicting the onset of faults in gear-boxes).

Requirements

Python programming, Basic understanding of spiking neural networks.

Contact

Elisa Donati (elisa (at) ini.uzh.ch); Prof. Giacomo Indiveri (giacomo (at) ini.uzh.ch), Ludovico Minati (lminati (at) ini.uzh.ch).

© 2021 Institut für Neuroinformatik