|
Abstract:
I describe a silicon network that consists of a group of
excitatory neurons and one global inhibitory neuron that can be
used to normalize the output signal with respect to the number of
inputs. This network models the normalization property of the
wide-field direction-selective cells in the fly visual system. This
normalization property is useful in any network where we wish the
output signal to code only the strength of the inputs and not be
dependent on the number of inputs. The circuitry in each neuron is
equivalent to that in Lazzaro et al's (Lazzaro 1998)
winner-take-all circuit with one additional transistor. Results
from a fabricated chip of 20 neurons in a 1.2
m CMOS technology show the transition between the soft-max
property and the hard winner-take-all property by the tuning of one
of parameter.
|
