Sentient Networks

In this paper we consider the question whether a distributed network of sensors and data processors can form "perceptions" based on the sensory data. Because sensory data can have exponentially many explanations, the use of a central data processor to analyze the outputs from a large ensemble of sensors will in general introduce unacceptable latencies for responding to dangerous situations. A better idea is to use a distributed "Helmholtz machine" architecture in which the collective state of the network as a whole provides an explanation for the sensory data.