It is difficult to characterize the field of ANS succinctly, because the approaches and the results are so diverse. ANS should include such systems as the single neuron models of Widrow and Hoff, Grossberg, Barto and Sutton, and Klopf; the thresholded logic networks of McCulloch and Pitts; the multilayer feedforward networks of Rosenberg; the backpropagation networks used by everybody since Werbos; and a variety of even more complex systems (many having the name Grossberg or Grossberg and Carpenter associated with them).
The predominant thrust of ANS modeling stems from the simplifying assumption that the primary phenomena of biological neural systems are related to their electrical or even electro-chemical behaviors, primarily at the synapses. This has brought the field quite a long way from the days of McCulloch and Pitts. More recently, various research efforts involving examination of quantum effects in BNS have resulted in further ANS models and refinements.
The basic unit of the BNS is the neuron. In ANS, the basic unit is called
various things: neuron, neurode, processing element (PE), and node are
popular. A neuron and a PE are not completely analogous. A neuron is a
threshold activation device, whereas many nodes employ a threshold function,
which scales the PE activation for output. In many cases, this is the sigmoid
or logistic function. A PE which uses the sigmoid function is actually
more like a collection of neurons in its behavior -- it acts like a neuronal
population where each neuron can have its own, somewhat varying, threshold
level.
