The rule for the perceptron (page 88) and the delta rule (page 93) appear similar. What is the fundamental difference?

Why will the gradient descent algorithm converge?

Why is there only one global minimum?

FAQs from demo.doc

1. What is NeuDL?
   
   NeuDL (pronounced noodle) is a Neural-Network Description Language that uses a C-like programming language interface to build, train, and run neural networks. It is currently limited to backpropagation neural networks, but it still shows the flexibility such an interface can give.

2. Where can NeuDL be downloaded from?
   
   The University of Alabama: cs.ua.edu (130.160.44.1) in the file /pub/neudl/NeuDLver021.tar This file contains all of the source code, several example NeuDL programs, makefile, user manual, and a short paper describing NeuDL.

3. What platform was NeuDL written for?
   
   It was written on an IBM RS/6000 with the xlC compiler, but it has been modified to compile with the GNU g++ compiler and it has been ported to DOS with no changes. It should compile with any good C++ compiler. If not, there should only be minor changes necessary.

4. Can a neural network trained with NeuDL be used inside another program?
   
   Yess, NeuDL is simply an interface to a backpropagation neural network object written in C++. It is very easy to embed a network in a program, and then link it with the neural network object. If fact, if a NeuDL program is translated into C++ with the automatic translate feature of the interpreter, it simply creates a C++ program calling the network object.

5. Is NeuDL an interpreter or a compiler?
   
   It is primarily an interpreter. It can execute the code directly. However, it has a feature that translate the NeuDL code into C++ which can be compiled with a C++ compiler and then executed. Translating and compiling the NeuDL code can sometime give a tremendous performance boost.