Different methods for different tasks (was symbolic equation

Paola, 
Every ANN that has ever been designed and used maps
symbols to symbols. For examples, please look at slide 36 of
http://jfsowa.com/talks/eswc.pdf .
The table at the top of the slide
is by Andrew Ng, who is an  expert in designing and developing ANNs.  The
comments about that table are summaries of what Ng said in the video,
which I cited at the bottom of that slide.
PDM> But I am in the
position that most important and even terrific is that
they begin to train ANN with symbolic input and/or output, getting
exciting results.
For pattern recognition, the input for a typical 
ANN is a matrix of symbols (triads of numbers for Red, Green, and Blue)
that represent the colors of pixels in a photograph.  The output is a
symbol (or structure of symbols) that describes the image represented by
those pixels.
In the Alpha Go system, which beat the world champion
at Go, the ANN for the evaluation function mapped symbols that represented
stones on a Go board to symbols (numbers) that estimated the strength of a
particular Go position for one player or the other.
Although the
Alpha  Go designers gave most of the credit to the ANN, the system was
actually a hybrid.  It used many symbolic steps to play the game and
search different options.  There was only one step that used an ANN: 
evaluate a board position to estimate which player had a better
position.
Research issue:   Instead of using one or more ANNs to do
all the steps of cognition, find some way of subdividing the task into a
variety of different kinds of tasks that must be performed.   Then
determine which of those tasks could be handled better by an ANN or by
some symbolic method.
John