<s>
The	O
Helmholtz	B-Algorithm
machine	I-Algorithm
(	O
named	O
after	O
Hermann	O
von	O
Helmholtz	O
and	O
his	O
concept	O
of	O
Helmholtz	O
free	O
energy	O
)	O
is	O
a	O
type	O
of	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
that	O
can	O
account	O
for	O
the	O
hidden	O
structure	O
of	O
a	O
set	O
of	O
data	O
by	O
being	O
trained	O
to	O
create	O
a	O
generative	O
model	O
of	O
the	O
original	O
set	O
of	O
data	O
.	O
</s>
<s>
The	O
hope	O
is	O
that	O
by	O
learning	O
economical	O
representations	B-General_Concept
of	O
the	O
data	O
,	O
the	O
underlying	O
structure	O
of	O
the	O
generative	O
model	O
should	O
reasonably	O
approximate	O
the	O
hidden	O
structure	O
of	O
the	O
data	O
set	O
.	O
</s>
<s>
A	O
Helmholtz	B-Algorithm
machine	I-Algorithm
contains	O
two	O
networks	O
,	O
a	O
bottom-up	O
recognition	O
network	O
that	O
takes	O
the	O
data	O
as	O
input	O
and	O
produces	O
a	O
distribution	O
over	O
hidden	O
variables	O
,	O
and	O
a	O
top-down	O
"	O
generative	O
"	O
network	O
that	O
generates	O
values	O
of	O
the	O
hidden	O
variables	O
and	O
the	O
data	O
itself	O
.	O
</s>
<s>
At	O
the	O
time	O
,	O
Helmholtz	B-Algorithm
machines	I-Algorithm
were	O
one	O
of	O
a	O
handful	O
of	O
learning	O
architectures	O
that	O
used	O
feedback	O
as	O
well	O
as	O
feedforward	O
to	O
ensure	O
quality	O
of	O
learned	O
models	O
.	O
</s>
<s>
Helmholtz	B-Algorithm
machines	I-Algorithm
are	O
usually	O
trained	O
using	O
an	O
unsupervised	B-General_Concept
learning	I-General_Concept
algorithm	O
,	O
such	O
as	O
the	O
wake-sleep	B-Algorithm
algorithm	I-Algorithm
.	O
</s>
<s>
They	O
are	O
a	O
precursor	O
to	O
variational	O
autoencoders	B-Algorithm
,	O
which	O
are	O
instead	O
trained	O
using	O
backpropagation	B-Algorithm
.	O
</s>
<s>
Helmholtz	B-Algorithm
machines	I-Algorithm
may	O
also	O
be	O
used	O
in	O
applications	O
requiring	O
a	O
supervised	O
learning	O
algorithm	O
(	O
e.g.	O
</s>
