<s>
ADALINE	B-Algorithm
(	O
Adaptive	B-Algorithm
Linear	I-Algorithm
Neuron	I-Algorithm
or	O
later	O
Adaptive	B-Algorithm
Linear	I-Algorithm
Element	I-Algorithm
)	O
is	O
an	O
early	O
single-layer	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
and	O
the	O
name	O
of	O
the	O
physical	O
device	O
that	O
implemented	O
this	O
network	O
.	O
</s>
<s>
The	O
network	O
uses	O
memistors	B-General_Concept
.	O
</s>
<s>
The	O
difference	O
between	O
Adaline	B-Algorithm
and	O
the	O
standard	O
(	O
McCulloch	O
–	O
Pitts	O
)	O
perceptron	O
is	O
that	O
in	O
the	O
learning	O
phase	O
,	O
the	O
weights	O
are	O
adjusted	O
according	O
to	O
the	O
weighted	O
sum	O
of	O
the	O
inputs	O
(	O
the	O
net	O
)	O
.	O
</s>
<s>
A	O
multilayer	O
network	O
of	O
ADALINE	B-Algorithm
units	O
is	O
known	O
as	O
a	O
MADALINE	O
.	O
</s>
<s>
Adaline	B-Algorithm
is	O
a	O
single	O
layer	O
neural	B-Architecture
network	I-Architecture
with	O
multiple	O
nodes	O
where	O
each	O
node	O
accepts	O
multiple	O
inputs	O
and	O
generates	O
one	O
output	O
.	O
</s>
<s>
The	O
ADALINE	B-Algorithm
converges	O
to	O
the	O
least	O
squares	O
error	O
which	O
is	O
.	O
</s>
<s>
This	O
update	O
rule	O
is	O
in	O
fact	O
the	O
stochastic	B-Algorithm
gradient	I-Algorithm
descent	I-Algorithm
update	O
for	O
linear	B-General_Concept
regression	I-General_Concept
.	O
</s>
<s>
MADALINE	O
(	O
Many	O
ADALINE	B-Algorithm
)	O
is	O
a	O
three-layer	O
(	O
input	O
,	O
hidden	O
,	O
output	O
)	O
,	O
fully	O
connected	O
,	O
feed-forward	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
architecture	O
for	O
classification	B-General_Concept
that	O
uses	O
ADALINE	B-Algorithm
units	O
in	O
its	O
hidden	O
and	O
output	O
layers	O
,	O
i.e.	O
</s>
<s>
The	O
three-layer	O
network	O
uses	O
memistors	B-General_Concept
.	O
</s>
<s>
Three	O
different	O
training	O
algorithms	O
for	O
MADALINE	O
networks	O
,	O
which	O
cannot	O
be	O
learned	O
using	O
backpropagation	B-Algorithm
because	O
the	O
sign	O
function	O
is	O
not	O
differentiable	O
,	O
have	O
been	O
suggested	O
,	O
called	O
Rule	O
I	O
,	O
Rule	O
II	O
and	O
Rule	O
III	O
.	O
</s>
<s>
finds	O
the	O
hidden	O
layer	O
unit	O
(	O
ADALINE	B-Algorithm
classifier	B-General_Concept
)	O
with	O
the	O
lowest	O
confidence	O
in	O
its	O
prediction	O
,	O
</s>
<s>
MADALINE	O
Rule	O
3	O
-	O
The	O
third	O
"	O
Rule	O
"	O
applied	O
to	O
a	O
modified	O
network	O
with	O
sigmoid	B-Algorithm
activations	O
instead	O
of	O
signum	O
;	O
it	O
was	O
later	O
found	O
to	O
be	O
equivalent	O
to	O
backpropagation	B-Algorithm
.	O
</s>
