Supervised learning

[Dave Rumelhart]

	It has seemed to me that the supervised/unsupervised distinction
is insufficiently fine-grained and is wrongly associated with learning methods.
(For example, backprop is often thought of as supervised whereas, say,
Hebbian rules are thought of as unsupervised.)  We ought to distinguish between
kinds of learning tasks on the one hand and learning rules on the other.  I have
distinguished four learning tasks: (1) PATTERN ASSOCIATION in which the goal
is to learn a series of I/O pairs.  In this case the networks learns to be the
function which maps input elements to their corresponding output elements. This
is usually what people have in mind with supervised learning.  
(2) AUTOASSOCIATION in which the goal is to store a pattern so that, 
essentially, any part of the pattern can be used to retrieve the whole pattern.
This is the connectionist implementation of content-addressible memory.  This
is difficult to classify in the supervised/unsupervised dimension.  On the one
hand it involves a single input item and hence should be unsupervised, on the 
other hand, the network is told exactly what to store, so it should be viewed
as unsupervised.  (3) REGULARITY-DETECTION in which the goal is to discover
statistical regularities (such as clusters or featural decompositions etc.)
in the input patterns.  This would seem to be the prototype unsupervised case.
(4) REINFORCEMENT LEARNING in which the goal is to learn a set of actions
which either maximize some postive reinforcement variable or minimize some 
negative reinforcement variable or both.  This to is difficult to classify on
the supervised/unsupervised dimension.  It would appear that it is
supervised because the environment is guiding its behavior.  On the other hand
it is unsupervised in as much as its free to make whatever response it wishes
so long as it maximizes reinforcement.  Perhaps partially supervised would
due for this case.

	More important than this classification, it is important to realize that
these categories are  (nearly) orthogonol to the learning rules employed.  Thus,
we have (1) CORRELATIONAL (Hebbian) LEARNING which can be employed for
pattern association tasks, autoassociation tasks, regularity detection tasks
and (perhaps) reinforcement learning tasks.  Similarly, (2) ERROR-CORRECTION
LEARNING (backprop, Widrow-Hoff, Perceptron, etc.) can be employed for each of
the classes.  Pattern association is the obvious application, but it is obvious
how it can be employed for autoassociation tasks.  Error-correction methods 
can be used to build an autoencoder and thereby be used to extract features or
principle components of the input patterns and thereby act as a regularity
detector.  Backprop, for example, can also be used in reinfoprcement learning
situations.  (3) COMPETITIVE LEARNING mechanisms can also be used for at least
regularity detection and for autoencoding and (probably) for possibly
for pattern association.  (4) BOLTZMANN LEARNING is just as versitile as
error correction learning.  

	The point is simply that a cross-classification of learning task
by learning rule is required to classify the kinds of connectionist learning
applications we see.

	der