<s>
Extreme	B-Algorithm
learning	I-Algorithm
machines	I-Algorithm
are	O
feedforward	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
for	O
classification	B-General_Concept
,	O
regression	O
,	O
clustering	B-Algorithm
,	O
sparse	O
approximation	O
,	O
compression	O
and	O
feature	B-General_Concept
learning	I-General_Concept
with	O
a	O
single	O
layer	O
or	O
multiple	O
layers	O
of	O
hidden	O
nodes	O
,	O
where	O
the	O
parameters	O
of	O
hidden	O
nodes	O
(	O
not	O
just	O
the	O
weights	O
connecting	O
inputs	O
to	O
hidden	O
nodes	O
)	O
need	O
to	O
be	O
tuned	O
.	O
</s>
<s>
they	O
are	O
random	B-Architecture
projection	I-Architecture
but	O
with	O
nonlinear	O
transforms	O
)	O
,	O
or	O
can	O
be	O
inherited	O
from	O
their	O
ancestors	O
without	O
being	O
changed	O
.	O
</s>
<s>
The	O
name	O
"	O
extreme	B-Algorithm
learning	I-Algorithm
machine	I-Algorithm
"	O
(	O
ELM	O
)	O
was	O
given	O
to	O
such	O
models	O
by	O
Guang-Bin	O
Huang	O
.	O
</s>
<s>
The	O
idea	O
goes	O
back	O
to	O
Frank	O
Rosenblatt	O
,	O
who	O
not	O
only	O
published	O
a	O
single	O
layer	O
Perceptron	B-Algorithm
in	O
1958	O
,	O
but	O
also	O
introduced	O
a	O
multi	B-Algorithm
layer	I-Algorithm
perceptron	I-Algorithm
with	O
3	O
layers	O
:	O
an	O
input	O
layer	O
,	O
a	O
hidden	O
layer	O
with	O
randomized	O
weights	O
that	O
did	O
not	O
learn	O
,	O
and	O
a	O
learning	O
output	O
layer	O
.	O
</s>
<s>
According	O
to	O
some	O
researchers	O
,	O
these	O
models	O
are	O
able	O
to	O
produce	O
good	O
generalization	O
performance	O
and	O
learn	O
thousands	O
of	O
times	O
faster	O
than	O
networks	O
trained	O
using	O
backpropagation	B-Algorithm
.	O
</s>
<s>
In	O
literature	O
,	O
it	O
also	O
shows	O
that	O
these	O
models	O
can	O
outperform	O
support	B-Algorithm
vector	I-Algorithm
machines	I-Algorithm
in	O
both	O
classification	B-General_Concept
and	O
regression	O
applications	O
.	O
</s>
<s>
From	O
2001-2010	O
,	O
ELM	O
research	O
mainly	O
focused	O
on	O
the	O
unified	O
learning	O
framework	O
for	O
"	O
generalized	O
"	O
single-hidden	O
layer	O
feedforward	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
(	O
SLFNs	O
)	O
,	O
including	O
but	O
not	O
limited	O
to	O
sigmoid	O
networks	O
,	O
RBF	B-Algorithm
networks	I-Algorithm
,	O
threshold	O
networks	O
,	O
trigonometric	O
networks	O
,	O
fuzzy	O
inference	O
systems	O
,	O
Fourier	O
series	O
,	O
Laplacian	O
transform	O
,	O
wavelet	O
networks	O
,	O
etc	O
.	O
</s>
<s>
One	O
significant	O
achievement	O
made	O
in	O
those	O
years	O
is	O
to	O
successfully	O
prove	O
the	O
universal	O
approximation	O
and	O
classification	B-General_Concept
capabilities	O
of	O
ELM	O
in	O
theory	O
.	O
</s>
<s>
From	O
2010	O
to	O
2015	O
,	O
ELM	O
research	O
extended	O
to	O
the	O
unified	O
learning	O
framework	O
for	O
kernel	O
learning	O
,	O
SVM	B-Algorithm
and	O
a	O
few	O
typical	O
feature	B-General_Concept
learning	I-General_Concept
methods	O
such	O
as	O
Principal	B-Application
Component	I-Application
Analysis	I-Application
(	O
PCA	O
)	O
and	O
Non-negative	O
Matrix	O
Factorization	O
(	O
NMF	O
)	O
.	O
</s>
<s>
It	O
is	O
shown	O
that	O
SVM	B-Algorithm
actually	O
provides	O
suboptimal	O
solutions	O
compared	O
to	O
ELM	O
,	O
and	O
ELM	O
can	O
provide	O
the	O
whitebox	O
kernel	O
mapping	O
,	O
which	O
is	O
implemented	O
by	O
ELM	O
random	O
feature	O
mapping	O
,	O
instead	O
of	O
the	O
blackbox	O
kernel	O
used	O
in	O
SVM	B-Algorithm
.	O
</s>
<s>
In	O
a	O
2017	O
announcement	O
from	O
Google	B-Library
Scholar	I-Library
:	O
""	O
,	O
two	O
ELM	O
papers	O
have	O
been	O
listed	O
in	O
the	O
"	O
,	O
"	O
taking	O
positions	O
2	O
and	O
7	O
.	O
</s>
<s>
Different	O
combinations	O
of	O
,	O
,	O
and	O
can	O
be	O
used	O
and	O
result	O
in	O
different	O
learning	O
algorithms	O
for	O
regression	O
,	O
classification	B-General_Concept
,	O
sparse	O
coding	O
,	O
compression	O
,	O
feature	B-General_Concept
learning	I-General_Concept
and	O
clustering	B-Algorithm
.	O
</s>
<s>
estimate	O
by	O
least-squares	B-Algorithm
fit	I-Algorithm
to	O
a	O
matrix	O
of	O
response	O
variables	O
,	O
computed	O
using	O
the	O
pseudoinverse	O
,	O
given	O
a	O
design	B-Algorithm
matrix	I-Algorithm
:	O
</s>
<s>
In	O
most	O
cases	O
,	O
ELM	O
is	O
used	O
as	O
a	O
single	O
hidden	O
layer	O
feedforward	O
network	O
(	O
SLFN	O
)	O
including	O
but	O
not	O
limited	O
to	O
sigmoid	O
networks	O
,	O
RBF	B-Algorithm
networks	I-Algorithm
,	O
threshold	O
networks	O
,	O
fuzzy	O
inference	O
networks	O
,	O
complex	O
neural	O
networks	O
,	O
wavelet	O
networks	O
,	O
Fourier	O
transform	O
,	O
Laplacian	O
transform	O
,	O
etc	O
.	O
</s>
<s>
Due	O
to	O
its	O
different	O
learning	O
algorithm	O
implementations	O
for	O
regression	O
,	O
classification	B-General_Concept
,	O
sparse	O
coding	O
,	O
compression	O
,	O
feature	B-General_Concept
learning	I-General_Concept
and	O
clustering	B-Algorithm
,	O
multi	O
ELMs	O
have	O
been	O
used	O
to	O
form	O
multi	O
hidden	O
layer	O
networks	O
,	O
deep	B-Algorithm
learning	I-Algorithm
or	O
hierarchical	O
networks	O
.	O
</s>
<s>
Both	O
universal	O
approximation	O
and	O
classification	B-General_Concept
capabilities	O
have	O
been	O
proved	O
for	O
ELM	O
in	O
literature	O
.	O
</s>
<s>
The	O
black-box	O
character	O
of	O
neural	O
networks	O
in	O
general	O
and	O
extreme	B-Algorithm
learning	I-Algorithm
machines	I-Algorithm
(	O
ELM	O
)	O
in	O
particular	O
is	O
one	O
of	O
the	O
major	O
concerns	O
that	O
repels	O
engineers	O
from	O
application	O
in	O
unsafe	O
automation	O
tasks	O
.	O
</s>
<s>
In	O
particular	O
,	O
it	O
was	O
pointed	O
out	O
in	O
a	O
letter	O
to	O
the	O
editor	O
of	O
IEEE	O
Transactions	O
on	O
Neural	O
Networks	O
that	O
the	O
idea	O
of	O
using	O
a	O
hidden	O
layer	O
connected	O
to	O
the	O
inputs	O
by	O
random	O
untrained	O
weights	O
was	O
already	O
suggested	O
in	O
the	O
original	O
papers	O
on	O
RBF	B-Algorithm
networks	I-Algorithm
in	O
the	O
late	O
1980s	O
;	O
Guang-Bin	O
Huang	O
replied	O
by	O
pointing	O
out	O
subtle	O
differences	O
.	O
</s>
