Abstract. 
Large, strongly coupled neural networks tend to produce chaotic
spontaneous activity.  This might appear to make them unsuitable for
generating reliable sensory responses or repeatable motor patterns.
However, this is not the case.  Inputs can induce a phase transition,
leading to responses uncontaminated by chaotic "noise".  Likewise,
appropriately trained feedback units can control the chaos, resulting in a
wide variety of repeatable output patterns.  Methods will be discussed for
quantifying the capacity of a network for generating complex dynamics and
for transferring information from one neuron to another.