/Monostable Multivibrator Networks as Artificial Neural Network

Monostable Multivibrator Networks as Artificial Neural Network

PhD - Leuven | More than two weeks ago

Build a new generation of extremely low power neural networks using the dynamical behavior of coupled timers

The neuromorphic paradigm broadly states that we should take inspiration from nature, in particular the brain, to build hardware capable of solving real-world AI tasks in an efficient way. However, since a neuron is a living cell first and only then a computational unit, it is not at all clear that the neuromorphic approach should lead to an optimal solution. In this project, we want to reverse the question. Instead of asking how we can mimic a spiking neuron in hardware, we could ask which existing elementary electronic building blocks show interesting dynamical and computational properties and can be applied as artificial neuron, regardless of their biological plausibility. A first investigation in that direction uses the monostable multivibrator (MMV) as novel artificial neuron, see reference below.

MMVs are simple timers that can be implemented in large numbers as counters in digital hardware (FPGA/ASIC). They form a class of non-biologically inspired spiking neurons. Networking MMVs together results in interesting dynamical behavior that has not yet been explored to the fullest. Recently we have shown that MMV networks can be trained using surrogate gradient techniques (currently unpublished).

This PhD project is intended for inquisitive and creative engineering or physics student. The goal of the project is to broaden the knowledge of MMV networks and ultimately to demonstrate their use in a practical spiking event-driven neural network implementation, possibly combined with an event-based sensor such as a dynamic vision camera (DVS). The student is encouraged to follow his/her own interest within this topic, continuing the work that has been done previously. Both theoretical and more hands-on approaches are encouraged.

Required background: Artificial neural networks, nonlinear dynamics, machine learning, graph theory, statistics, renewal processes and neuroscience in general. Good programming and computing skills (Python, C/C++, Linux...) are required.

Type of work: 25% theory, 50% modelling/simulation, 25% experimental

Supervisor: Piet Wambacq

Daily advisor: Lars Keuninckx

The reference code for this position is 2023-100. Mention this reference code on your application form.

Who we are
Accept marketing-cookies to view this content.
Cookie settings
imec's cleanroom
Accept marketing-cookies to view this content.
Cookie settings

Send this job to your email