🚀 go-pugleaf

One way for statisticians to think of neural nets
is as "recursive generalized linear models".
If you are familiar with glms, then consider a
glm in which each predictor may in fact be the
result of a glm on another set of predictors.

I wouldn't claim that the above covers all neural
models, but it does cover "feedforward neural nets
with a single hidden layer", which are among the 
ones most often used on statistical problems.

To explain glm's to a connectionist, you can
say they are feedforward nets with "no hidden layer".

Art Owen