














• 
Instead of
getting the correct probability distribution over


rivals by
running a Markov chain until it reaches the



equilibrium
distribution, use slight corruptions of the



datavectors as
the rivals (Bartlett’s
experiments)



• 
Learn by
modifying the weights so as to lower the



energy of the
true data and raise the energy of the



slightly
corrupted data.



• 
The learning
stops when the model’s prejudices do not



corrupt the data
in any systematic way.




– 
Its
OK if unlikely datavectors are usually corrupted



into
more likely ones.




– 
This
is exactly balanced by corrupting a small



fraction
of the likely datavectors into unlikely ones.

