Speaker
Prof.
Joannes Schemmel
(Kirchhoff Institut fuer Physik / Electronic Vision(s))
Description
This talk presents different VLSI models of artificial neural
networks ranging from abstract ones using binary neurons to
biologically inspired pulse-coupled systems. Circuit examples
demonstrating common design principles for optimizing area usage and
network speed are shown. The usage of digital communication
protocols allows the parallelization of the analog network cores to
create large-scale artificial network systems.
A key aspect of neural network research is training. Specific
training algorithms for simple and complex electronic neuron models
have been investigated. For binary models, evolutionary- as well as
liquid-computing has proven to be successful. While these methods
operate on a global level, the pulse-coupled systems use a local
learning rule inspired by contemporary neuroscience called 'spike
time dependent plasticity' (STDP). Therefore the VLSI system allows
the investigation of important aspects of natural neural plasticity
at a speed several orders faster than biological real time.
Author
Prof.
Joannes Schemmel
(Kirchhoff Institut fuer Physik / Electronic Vision(s))
