<s>
Unsupervised	B-General_Concept
learning	I-General_Concept
is	O
a	O
type	O
of	O
algorithm	O
that	O
learns	O
patterns	O
from	O
untagged	O
data	O
.	O
</s>
<s>
In	O
contrast	O
to	O
supervised	B-General_Concept
learning	I-General_Concept
where	O
data	O
is	O
tagged	O
by	O
an	O
expert	O
,	O
e.g.	O
</s>
<s>
The	O
other	O
levels	O
in	O
the	O
supervision	O
spectrum	O
are	O
reinforcement	O
learning	O
where	O
the	O
machine	O
is	O
given	O
only	O
a	O
numerical	O
performance	O
score	O
as	O
guidance	O
,	O
and	O
semi-supervised	B-General_Concept
learning	I-General_Concept
where	O
a	O
small	O
portion	O
of	O
the	O
data	O
is	O
tagged	O
.	O
</s>
<s>
Neural	B-Architecture
network	I-Architecture
tasks	O
are	O
often	O
categorized	O
as	O
discriminative	O
(	O
recognition	O
)	O
or	O
generative	O
(	O
imagination	O
)	O
.	O
</s>
<s>
Often	O
but	O
not	O
always	O
,	O
discriminative	O
tasks	O
use	O
supervised	O
methods	O
and	O
generative	O
tasks	O
use	O
unsupervised	O
(	O
see	O
Venn	B-Application
diagram	I-Application
)	O
;	O
however	O
,	O
the	O
separation	O
is	O
very	O
hazy	O
.	O
</s>
<s>
For	O
example	O
,	O
object	O
recognition	O
favors	O
supervised	B-General_Concept
learning	I-General_Concept
but	O
unsupervised	B-General_Concept
learning	I-General_Concept
can	O
also	O
cluster	O
objects	O
into	O
groups	O
.	O
</s>
<s>
In	O
contrast	O
to	O
supervised	O
methods	O
 '	O
dominant	O
use	O
of	O
backpropagation	B-Algorithm
,	O
unsupervised	B-General_Concept
learning	I-General_Concept
also	O
employs	O
other	O
methods	O
including	O
:	O
Hopfield	O
learning	O
rule	O
,	O
Boltzmann	O
learning	O
rule	O
,	O
Contrastive	O
Divergence	O
,	O
Wake	B-Algorithm
Sleep	I-Algorithm
,	O
Variational	O
Inference	O
,	O
Maximum	O
Likelihood	O
,	O
Maximum	O
A	O
Posteriori	O
,	O
Gibbs	O
Sampling	O
,	O
and	O
backpropagating	O
reconstruction	O
errors	O
or	O
hidden	O
state	O
reparameterizations	O
.	O
</s>
<s>
In	O
Boltzmann	B-Algorithm
machines	I-Algorithm
,	O
it	O
plays	O
the	O
role	O
of	O
the	O
Cost	O
function	O
.	O
</s>
<s>
Hence	O
,	O
early	O
neural	B-Architecture
networks	I-Architecture
bear	O
the	O
name	O
Boltzmann	B-Algorithm
Machine	I-Algorithm
.	O
</s>
<s>
thumb|Restricted	O
Boltzmann	B-Algorithm
Machine	I-Algorithm
.	O
</s>
<s>
This	O
is	O
a	O
Boltzmann	B-Algorithm
machine	I-Algorithm
where	O
lateral	O
connections	O
within	O
a	O
layer	O
are	O
prohibited	O
to	O
make	O
analysis	O
tractable	O
.	O
</s>
<s>
Helmholtz	O
Autoencoder	O
VAE	O
thumb|Instead	O
of	O
the	O
bidirectional	O
symmetric	O
connection	O
of	O
the	O
stacked	O
Boltzmann	B-Algorithm
machines	I-Algorithm
,	O
we	O
have	O
separate	O
one-way	O
connections	O
to	O
form	O
a	O
loop	O
.	O
</s>
<s>
Of	O
the	O
networks	O
bearing	O
people	O
's	O
names	O
,	O
only	O
Hopfield	O
worked	O
directly	O
with	O
neural	B-Architecture
networks	I-Architecture
.	O
</s>
<s>
Boltzmann	O
and	O
Helmholtz	O
came	O
before	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
,	O
but	O
their	O
work	O
in	O
physics	O
and	O
physiology	O
inspired	O
the	O
analytical	O
methods	O
that	O
were	O
used	O
.	O
</s>
<s>
1969	O
Perceptrons	O
by	O
Minsky	O
&	O
Papert	O
shows	O
a	O
perceptron	O
without	O
hidden	O
layers	O
fails	O
on	O
XOR	O
1970s	O
(	O
approximate	O
dates	O
)	O
First	O
AI	O
winter	O
1974	O
Ising	O
magnetic	O
model	O
proposed	O
by	O
WA	O
Little	O
for	O
cognition	O
1980	O
Fukushima	O
introduces	O
the	O
neocognitron	O
,	O
which	O
is	O
later	O
called	O
a	O
convolution	O
neural	B-Architecture
network	I-Architecture
.	O
</s>
<s>
1982	O
Ising	O
variant	O
Hopfield	B-Algorithm
net	I-Algorithm
described	O
as	O
CAMs	O
and	O
classifiers	O
by	O
John	O
Hopfield	O
.	O
</s>
<s>
1983	O
Ising	O
variant	O
Boltzmann	B-Algorithm
machine	I-Algorithm
with	O
probabilistic	O
neurons	O
described	O
by	O
Hinton	O
&	O
Sejnowski	O
following	O
Sherington	O
&	O
Kirkpatrick	O
's	O
1975	O
work	O
.	O
</s>
<s>
1995	O
Dayan	O
&	O
Hinton	O
introduces	O
Helmholtz	B-Algorithm
machine	I-Algorithm
1995-2005	O
(	O
approximate	O
dates	O
)	O
Second	O
AI	O
winter	O
2013	O
Kingma	O
,	O
Rezende	O
,	O
&	O
co	O
.	O
introduced	O
Variational	B-Algorithm
Autoencoders	I-Algorithm
as	O
Bayesian	O
graphical	O
probability	O
network	O
,	O
with	O
neural	B-Architecture
nets	I-Architecture
as	O
components	O
.	O
</s>
<s>
no	O
back	B-Algorithm
propagation	I-Algorithm
.	O
</s>
<s>
wake-sleep	O
2	O
phase	O
training	O
back	O
propagate	O
the	O
reconstruction	O
error	O
reparameterize	O
hidden	O
state	O
for	O
backprop	B-Algorithm
Strength	O
resembles	O
physical	O
systems	O
so	O
it	O
inherits	O
their	O
equations	O
←	O
same	O
.	O
</s>
<s>
hidden	O
neurons	O
act	O
as	O
internal	O
representatation	O
of	O
the	O
external	O
world	O
faster	O
more	O
practical	O
training	O
scheme	O
than	O
Boltzmann	B-Algorithm
machines	I-Algorithm
trains	O
quickly	O
.	O
</s>
<s>
The	O
classical	O
example	O
of	O
unsupervised	B-General_Concept
learning	I-General_Concept
in	O
the	O
study	O
of	O
neural	B-Architecture
networks	I-Architecture
is	O
Donald	O
Hebb	O
's	O
principle	O
,	O
that	O
is	O
,	O
neurons	O
that	O
fire	O
together	O
wire	O
together	O
.	O
</s>
<s>
Among	O
neural	B-Architecture
network	I-Architecture
models	I-Architecture
,	O
the	O
self-organizing	B-Algorithm
map	I-Algorithm
(	O
SOM	O
)	O
and	O
adaptive	B-Algorithm
resonance	I-Algorithm
theory	I-Algorithm
(	O
ART	O
)	O
are	O
commonly	O
used	O
in	O
unsupervised	B-General_Concept
learning	I-General_Concept
algorithms	O
.	O
</s>
<s>
ART	O
networks	O
are	O
used	O
for	O
many	O
pattern	O
recognition	O
tasks	O
,	O
such	O
as	O
automatic	B-Algorithm
target	I-Algorithm
recognition	I-Algorithm
and	O
seismic	O
signal	O
processing	O
.	O
</s>
<s>
Two	O
of	O
the	O
main	O
methods	O
used	O
in	O
unsupervised	B-General_Concept
learning	I-General_Concept
are	O
principal	B-Application
component	I-Application
and	O
cluster	B-Algorithm
analysis	I-Algorithm
.	O
</s>
<s>
Cluster	B-Algorithm
analysis	I-Algorithm
is	O
used	O
in	O
unsupervised	B-General_Concept
learning	I-General_Concept
to	O
group	O
,	O
or	O
segment	O
,	O
datasets	O
with	O
shared	O
attributes	O
in	O
order	O
to	O
extrapolate	O
algorithmic	O
relationships	O
.	O
</s>
<s>
Cluster	B-Algorithm
analysis	I-Algorithm
is	O
a	O
branch	O
of	O
machine	O
learning	O
that	O
groups	O
the	O
data	O
that	O
has	O
not	O
been	O
labelled	B-General_Concept
,	O
classified	O
or	O
categorized	O
.	O
</s>
<s>
Instead	O
of	O
responding	O
to	O
feedback	O
,	O
cluster	B-Algorithm
analysis	I-Algorithm
identifies	O
commonalities	O
in	O
the	O
data	O
and	O
reacts	O
based	O
on	O
the	O
presence	O
or	O
absence	O
of	O
such	O
commonalities	O
in	O
each	O
new	O
piece	O
of	O
data	O
.	O
</s>
<s>
A	O
central	O
application	O
of	O
unsupervised	B-General_Concept
learning	I-General_Concept
is	O
in	O
the	O
field	O
of	O
density	B-General_Concept
estimation	I-General_Concept
in	O
statistics	O
,	O
though	O
unsupervised	B-General_Concept
learning	I-General_Concept
encompasses	O
many	O
other	O
domains	O
involving	O
summarizing	O
and	O
explaining	O
data	O
features	O
.	O
</s>
<s>
It	O
can	O
be	O
contrasted	O
with	O
supervised	B-General_Concept
learning	I-General_Concept
by	O
saying	O
that	O
whereas	O
supervised	B-General_Concept
learning	I-General_Concept
intends	O
to	O
infer	O
a	O
conditional	O
probability	O
distribution	O
conditioned	O
on	O
the	O
label	O
of	O
input	O
data	O
;	O
unsupervised	B-General_Concept
learning	I-General_Concept
intends	O
to	O
infer	O
an	O
a	O
priori	O
probability	O
distribution	O
.	O
</s>
<s>
Some	O
of	O
the	O
most	O
common	O
algorithms	O
used	O
in	O
unsupervised	B-General_Concept
learning	I-General_Concept
include	O
:	O
(	O
1	O
)	O
Clustering	B-Algorithm
,	O
(	O
2	O
)	O
Anomaly	B-Algorithm
detection	I-Algorithm
,	O
(	O
3	O
)	O
Approaches	O
for	O
learning	O
latent	O
variable	O
models	O
.	O
</s>
<s>
One	O
of	O
the	O
statistical	O
approaches	O
for	O
unsupervised	B-General_Concept
learning	I-General_Concept
is	O
the	O
method	O
of	O
moments	O
.	O
</s>
<s>
Higher	O
order	O
moments	O
are	O
usually	O
represented	O
using	O
tensors	B-Device
which	O
are	O
the	O
generalization	O
of	O
matrices	O
to	O
higher	O
orders	O
as	O
multi-dimensional	O
arrays	O
.	O
</s>
<s>
It	O
is	O
shown	O
that	O
method	O
of	O
moments	O
(	O
tensor	B-Device
decomposition	O
techniques	O
)	O
consistently	O
recover	O
the	O
parameters	O
of	O
a	O
large	O
class	O
of	O
latent	O
variable	O
models	O
under	O
some	O
assumptions	O
.	O
</s>
<s>
The	O
Expectation	B-Algorithm
–	I-Algorithm
maximization	I-Algorithm
algorithm	I-Algorithm
(	O
EM	O
)	O
is	O
also	O
one	O
of	O
the	O
most	O
practical	O
methods	O
for	O
learning	O
latent	O
variable	O
models	O
.	O
</s>
