16 Self-organizing Learning

The defining property of self-organizing learning is that it operates without requiring an explicit training signal from the environment. This can be contrasted with error backpropagation, which requires target patterns to compare against the output states in order to generate an error signal. Thus, many people regard self-organizing learning as more biologically or psychologically plausible, since it is often difficult to imagine where the explicit training signals necessary for error-driven learning come from. Further, there is some evidence that neurons actually use something like a Hebbian learning rule, which is commonly used in self-organizing learning algorithms.

There are many different flavors of self-organizing learning. Indeed, one of the main differences between self-organizing algorithms and error-driven learning is that they need to make more assumptions about what good representations should be like, since they do not have explicit error signals telling them what to do. Thus, different self-organizing learning algorithms make different assumptions about the environment and how best to represent it.

One assumption that is common to many self-organizing learning algorithms is that events in the environment can be clustered together according to their "similarity." Thus, learning amounts to trying to find the right cluster in which to represent a given event. this is often done by enforcing a competition between a set of units, each of which represents a different cluster. The competitive learning algorithm (CL) of Rumelhart and Zipser, 1985 is a classic example of this form of learning, where the single unit which is most activated by the current input is chosen as the "winner" and therefore gets to adapt its weights in response to this input pattern.

The PDP++ implementation of self-organizing learning, called So, includes competitive learning and several variations of it, including "soft" competitive learning Nowlan, 1990, which replaces the "hard" competition of standard competitive learning with a more graded activation function. Also included are a couple of different types of modified Hebbian learning rules that can be used with either hard or soft activation functions.

An additional assumption that can be made about the environment is that there is some kind of topology or ordered relationship among the different clusters. This notion is captured in the self-organizing map (SOM) algorithm of Kohonen, 1989; 1990; 1995. This algorithm adds to the basic idea of competition among the units that represent a cluster the additional assumption that units which are nearby in 2-D space should represent clusters that are somehow related. This spatial-relatedness constraint is imposed by allowing nearby units to learn a little bit when one of their neighbors wins the competition. This algorithm is also implemented in the So package.

The directory `demo/so' contains two projects which demonstrate the use of both the competitive-learning style algorithms, and the self-organizing maps.