next up previous
Next: A Mixed-mode VLSI implementation Up: 1st European Workshop on Previous: Computational Principles in the

Weight vector normalization in an analog VLSI artificial neuron using a backpropagating action potential

Ph. Haefliger and M.Mahowald, Institute of Neuroinformatics, ETHZ/UNIZ, Gloriastr 32, Zurich 8006

At NIPS 96 we introduced the Modified Riccati Rule (MRR) a Hebbian-like learning algorithm that uses temporal correlation between pre-and post-synaptic spikes to determine changes in synaptic connectivity. Recent physiological experiments in young rat neocortex indicate that the relative timing of single EPSPs and APs indeed influence the synaptic efficacy. Since the backpropagating AP is shared information among all synapses, the synaptic changes could evolve in a coordinated way. Some coordinated behaviour, which in the MRR is achieved this way is weight vector normalization. We introduce a 2 tex2html_wrap_inline49 m CMOS implementation of the MRR and demonstrate its normalizing property.



Dr L S Smith (Staff)
Tue Dec 2 14:23:49 GMT 1997