M-5 Neural Network Computer

Simulated Neural Networks can outperform traditional computers at some tasks! M-5 uses the classic Feed Forward Back Propagation algorithm to adjust connection strengths (weights). It uses floating point math to provide results equivalent to or surpassing those of desktop systems costing hundreds of dollars. It can have up to eight variables on the Input layer, has three neurons in the Hidden layer and an Output neuron.

Review 06-Feb-03 by Steven J. The performance of M-5 has been so fantastic that words to describe it escape me at the moment! Really. I am impressed. I'm blown away! In my first real test to train M-5, I had 8 parameters applied to 30 different training examples. I expected M-5 to give me 1 of 4 choices regarding each of 10 initial NNrun examples. Based on what I thought the outcomes should be in NNout, M-5 agreed with 8 of 10. The 20% "anomaly" puzzled me, so I more closely inspected to find that on one of them, my initial assessment was incorrect and M-5 was actually correct. (Oooops, my mistake.) I ran a few additional test examples with variations on the anomaly example and guess what --- M-5 was in actuality correct on the remaining "anomaly" as well. M-5, after training for around 3.5 hours, understood my question better than I, the originator of the problem to be solved, and M-5 then pointed out my errors to me. This is a bit "scary" in a very good way!