Neural Classifiers - Neural Networks and Micromechanics - page 40

All the I -layer neurons connected to one A -layer neuron form the group of this
A -layer neuron. The number of neurons in one group corresponds to the number of
positive and negative connections between one neuron of the A -layer and the retina
in the LIRA_binary structure. The I -layer neurons could be ON neurons or OFF
neurons. The output of the ON neurons i is “1” when its input is higher than the
threshold y i , and otherwise it equals “0.” The OFF neuron j output is “1” when its

input is less than the threshold y j , and otherwise it equals “0.”

The ON neuron number in each group corresponds to the number of the positive
connections of one A -layer neuron in the LIRA_binary structure. The OFF neuron
number in each group corresponds to the number of the negative connections. In our
case we selected three ON neurons and five OFF neurons. The rule of connection
arrangement between the retina and one group of the I -layer is the same as the rule
of mask design for one A -layer neuron in the LIRA-binary.
The thresholds

y i and y j are selected randomly from the range [0,

·b max ], where

b max is maximal brightness of the image pixels, and is the parameter from [0, 1 ],

which is selected experimentally. The output of the A -layer neuron is “1” if all
outputs of its I -layer group are “1.” If any neuron of this group has the output “0,”
the A -layer neuron has the output “0.”

3.4 Handwritten Digit Recognition Results for Lira-binary

We carried out preliminary experiments to estimate the performance of our classi-
fiers. On the basis of these experiments, we selected the best classifiers and carried
out final experiments to obtain the maximal recognition rate. In the preliminary
experiments we changed the A -layer neuron number from 1,000 to 128,000 (Table 3.2 ).
These experiments showed that the recognition error number has been decreased
approximately by the factor 8 with the increase of the A -layer neuron number.
Disadvantages of a very large A -layer are the increasing of train and recognition
time and memory capacity.
We also changed the ratio p = w/W S = h/H S from 0.2 to 0.8. The parameter T E
was 0.1. In these experiments we did not use distortions in either training or

Table 3.2 The recognition rates of classifiers in preliminary experiments
A -layer neuron number Error number

p = 0.2 p = 0.4 p = 0.6 p = 0.8 p =1

1000 3461 1333 1297 1355 1864

2000 1705 772 772 827 1027

4000 828 452 491 532 622

8000 482 338 335 388 451

16000 330 249 247 288 337

32000 245 205 207 246 270

64000 218 186 171 190 217

128000 207 170 168 190 195