Advanced Topics

Neural Nets - Advanced Topics

In this example, neural nets are used to determine the fermentation activity of recombinant E. coli cultures.

Data from numerous variables, such as O₂ and CO₂ concentrations, acidity, respiration rate, etc., are used to train the neural net to identify the best mix of conditions for fermentation.

Some intermediate and advanced tutorials for neural networks can be found on this extensive list. This is a good site to look at before moving on to the resources listed below.

This online book about neural networks at Statsoft is more realistic than many sites in that it also discusses bias and adjustment in neural nets and the use of neural networks as part of an ensemble of techniques to perform a given task. This is important because some neural networks must adapt to a changing environment. In that case, there's a tradeoff between adaptability and lower accuracy due to "drift" of the model.

Mathematics of neural networks:

A summary outline of the mathematics used to describe neural nets can be found here [ 1 | 2 | 3 ]. Naturally, Wolfram Research also has something to say about the subject. A more complete treatment is in this online course. For the truly ambitious, take a look at MIT's course An Introduction to Neural Networks.

Neural network software:

This site describes the general algorithms behind neural network programs.

Here's an overview of some programming environments. Jeff Heaton has put together a tremendous amount of information about programming neural nets in Java.

The programs tend to implement different varieties of neural networks such as the Hopfield Net or the Kohonen Net. A nice applet to explore how a Hopfield net works is here. A more graphically-oriented Hopfield net can give a complementary view of how it works. Back propagation networks (the kind that your local friendly website author has programmed) are explained here. Here is yet another website that explains back propagation networks.

Extensions to neural networks:

Neural networks can be combined with another computing strategy: fuzzy logic. With fuzzy logic, computational results don't need to be a binary yes or no. They can include various values of maybe. When fuzzy logic is combined with neural networks, the output of the Fuzzy Neural Network can be tuned to higher accuracy.

Neuroscience:

The Neuroscience wiki has a lot of detail about how neurons work in the brain.

For a real flood of information about neuroscience and biological neural networks, this page-o-links will open the floodgates.