Expert: Saurabh Kudesia - 7/28/2004

Question
I'd like to better understand how neural networks are designed, but some of the advanced mathematics involved is over my head. Maybe you could use this silly little example to explain some of the principles one uses to decide on a particular implementation:

Suppose you wanted to create a neural network that decodes MP3 files (with packets of data compressed using lossy methods and knowledge of the human ear) into WAV files (a plain list of integers). Also assume that you have some examples of encoded and decoded files, but you don't know much about the MP3 format.

I suspect that a NN decoder would be terribly inefficient compared to a small, carefully-designed algorithm or the ordinary sort. But if you did want to create such a thing, what type of NN would you use, with what specifications? Could a NN like this be created and run on a home computer?

Thank you.  

Answer
Hi JP,

It would be nice if you could be specific and elaborate your problem.

There are few main things to be consider. NN does require a lot of mathematics and other stuff, which, even if you hate, have to bear.

Secondly, You should be familiar with the coding involved in both MP3 and WAV files. You should also be familiar with the algorithms that converts them so that you can have a fair good idea about what section/part of the code are really necessary and in what manner they are contributing to the conversion process.

Thirdly, desgining a NN decoder would require inputs from the above answer as well as from different other algorithms that are generally needed to convert information into NN.

I would like to extend my answer to some other critical issues but that will only confuse. I think that you reconsider your question and let me know what exactly you are thinking so that I can come up with a more focussed and descriptive answer.

Thanks and Regards
Saurabh Kudesia