Evolving very small SNNs for simple computational tasks, robust to noise and damage
We are interested in using artificial evolution (that can change the topology and weights) to obtain very small spiking neural networks. Very small means 3-10 adaptive exponential or leaky integrate and fire. Simple computational tasks may include temporal pattern recognition, controlling an animat (a simulated robot), or a multiplicative operation. When such networks are evolved with noise on state variables (membrane voltage), they seem to be robust to changes of neural parameters (AdEx parameters) and weights in the network. We will be interested to see if we could test the evolved networks on neuromorphic hardware, to see if they are robust to variability inherent in the hardware.
|Wed, 25.04.2018||15:00 - 16:00||Main lecture room|
|Thu, 26.04.2018||11:00 - 12:00||Panorama|
|Fri, 27.04.2018||14:00 - 15:00||Panorama|