Is there a resource that explains ANN with simplicity?

Question

I have seen many websites and blogs explain ANNs (Artificial Neural Networks), but none take into account a step by step procedure of developing a base for people who do not know much about ANNs', instead they throw in a lot of mathematics, which unfortunately for me looks very elusive. Are there any websites/blogs that, say, teach a simple ANN program that can recognize computer generated digits or something even more simple than digits?

Answer 1

This my annotated list of the best introductory (but complete) resources on ANN (aka MLP).

Course Notebook, Willamette University This resource is about 25-30 pages total (navigate among the sections using the sidebar on the upper right of this page or by clicking next or previous at the bottom of a given page). I think so highly of this resource for two reasons (i) it contains a large number of diagrams and nicely integrates them with prose; and (ii) it's comprehensive--architecture of the network, the calculus that underlies back-propagation, selecting/iterating the training parameters (momentum, learning rate), etc. The section on back-propagation (the computational and programmatic crux of ANN/MLP) is particularly good--combining prose, network diagrams, and the actual equations, it meticulously explains each step in a single epoch that comprises the training phase.

IBM developer works class on neural networks This one, by David Mertz, contains working code (python) and a non-trivial data set. With the exception of the first few paragraphs, the entire document closely follows the code and data. To me, this is crucial, because unless i can code it, i don't think i have learned it (that's a personal threshold, might not work for others). Aside from that, the focus on code and interaction of that code with a supplied data set keeps the discussion practically grounded. Finally, the authors David Mertz and Andrew Blais, obviously have a strong command of the subject matter.

Generation5 This is the most basic of the three, so perhaps the one to start with. The author is obviously knowledgeable of the minutiae of multi-layers perceptrons but doesn't require such knowledge of the reader--i.e., he explains back-propagation as a common-sense solution to a problem, without deferring to the numerical solution techniques (which is common in many references on MLP--'backprop is solved via gradient descent'). And like the first resource i mentioned, this one relies heavily on diagrams. Rather than data, (0's and 1's), the author discusses the structure and function of a MLP in the context of a simple predictive analytics scenario, using prose rather than reducing the problem to numerical data.