Page 205 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

3. Neural Networks Enter Mainstream Science 195

These developments led to the alliance between engineering and neuroscience
that spawned INNS, the IJCNN conferences, and a host of smaller neural network
conferences across the world. They also led to a rediscovery of the work of
researchers who had started in the “dark ages,” such as Grossberg, Anderson, and
Kohonen.
Yet the popularization of neural networks within academia has often been
accompanied by a restricted view of the nature and explanatory capabilities of
neural networks. Because of the influence of the PDP approach developed by
Rumelhart and McClelland [4], many neuroscientists, psychologists, and even
philosophers (e.g., Bechtel [53]), write with a view of all or most neural networks
as requiring extensive training to achieve a predetermined goal. This means that
between the sensory and effector ends of the network, the inner components start
out with little or no intrinsic structure and emerge out of the wash as “internal
representations.” Also, many authors seem to believe that the three-layer back
propagation network with input, output, and hidden layers is the standard “neural
network.”
By contrast, an overview of the neural network field [1] shows the great diversity
of neural network architectures, of which PDP networks are just a fraction. Also,
while the PDP approach emphasizes development over genetics, both innate and
learned processes are of evolutionary importance to organisms, and the richest
neural networks incorporate and integrate both types of processes.
Most importantly, different neural and psychological functions often require
different architectures, even if those functions are integrated together into a larger
network. This means that any “one size fits all” approach to neural network
modeling is bound to have limitations. For example Gaudiano and Grossberg
[54], discuss the different requirements for sensory pattern processing and for
motor control. Motor control involves comparing the present position with a target
position and inhibiting movement if the present and target positions match. In
sensory pattern processing, by contrast, the pattern recognizer is excited if present
and expected patterns match. The different requirements suggest different architec-
tures for the two subprocesses, and these two architectures can be concatenated
into a larger system architecture that generates motion in response to appropriate
sensory events.
Buoyed by early successes in modeling perception and motor control, the
neural network field has expanded in the last 30 years into processes several
synapses away from the sensory and motor ends, such as conditioning, attention,
cognitive control, and executive function. Some models of those processes have
built on and refined earlier, more abstract models of simpler processes. Others
have started from data about the complex processes and used simple equations
for neural interactions to simulate those data without reference to simpler
processes. Because of the unity of the brain and mind, it is my contention that those
models that build on the models of simpler processes are more likely to have
staying power.