D. Hush, Classification with neural networks: a performance analysis

Don Hush has analyzed a very simple multi-layer perceptron (MLP) to quantify its capacity and performance in an article titled: “Classification with neural networks: a performance analysis“, IEEE International Conference on Systems Engineering, pp 277 – 280, Fairborn, OH , USA, 24 Aug 1989 – 26 Aug 1989. Some conclusions he draws are: networks with one hidden layer perform better than those with two hidden layers; the number of nodes in the hidden layer must be no smaller than d+1 and optimally about 3d where d is the dimension of the data pattern; finally, for best performance the number of training samples should be approximately 60d(d+1).

Don Hush and Bill Horne documented ” Progress in supervised neural networks” in IEEE Signal Processing Magazine,Vol.10, Issue 1, pp 8 – 39, Jan 1993. This review article describes MLP neural net processing, and more crucially, MLP training algorithms. Back in the early nineties, I used this review to specify processing for an MLP that fused the results from multiple independent classifiers. I observed two inescapable performance features.

If the training data contained near identical inputs for two apriori distinct classes then the MLP could not reliably distinguish between the classes (self-evident but with serious consequences). The other was that MLP fusion performance was dominated by the best classifier. In fact, the MLP fusion performance was alway less than that of the dominant input classifier. I concluded that either one had to use classifiers of comparable capability or it paid to reject the fusion process.

I found that the same MLP training software could be used to train a time delay neural network (TDNN) with little modification. I trained the seven node (one hidden layer)  TDNN to “match filter” a discrete representation of the chaotic “Logistic map“.

P. Cochrane, A Measure of Machine Intelligence

Peter Cochrane’s opinion piece: “A Measure of Machine Intelligence” in the Proceedings of the IEEE, Vol. 98, No. 9, pp 1543 – 1545, September 2010 brings out several important points about “robots”.  He says:

” A further important observation at this point is the fact that the sensors and actuators have largely been neglected as components of intelligence…sophisticated sensors have only recently emerged as key capability components in robotics, artificial intelligence, and control systems. “

Recently, in a discussion with an up and coming roboticist, I mentioned this fact and he agreed. Sensors and their high fidelity outputs will enable intelligent machines to do the dull, dangerous and dirty work we need them to perform. 

Robots will derive their actions from their observations of their constructed world picture. They may use top down or bottoms up approaches to machine intelligence, mobility and goal achievement. Most likely, they will use a fusion of control approaches: high level remote supervisory control for human interaction and goal oriented behaviors and lower level autonomic control for health maintenance, vehicle stability and navigation, failsafes, etc. The key to both is the right mix of task optimized sensors.

The best example of this approach to intelligence is Homo Sapiens although most other fauna and some flora on Earth do a pretty good job as well.