This is a very cool task. It uses a database of 3,823 examples taken from real people's handwriting, with 1,797 examples held out for testing. As you can see by looking in the Unit Viewer, the data is pretty noisy. Nevertheless, the network can learn to recognize the digits pretty well.
The 10 output units correspond to the 10 digits. The output group has a SOFT_MAX constraint, so the output will sum to 1. You can interpret that as the probability, according to the network, that each of the digits is the intended one.
Sure the network can do well on the training set, but the question is how well it does on the testing set.
To help answer this, the output group has a MAX_CRIT criterion. This will be true if the correct digit has the highest activation (and is within the testGroupCrit of the correct value). If you run "test" on the testing set, the last thing it tells you is the percentage of testing examples it got correct. If you set the testGroupCrit to 0.5, then it will only mark an example correct if the activation of the network's choice is above 0.5.
You will probably get about 95% correct on the testing set out of the box. Try judging the examples by eye. Can you recognize digits as well as the network?
Try improving on the network's 95% performance. Maybe with a weight decay in the neighborhood of 0.001. If that doesn't help, you might try reducing the number of hidden units. It's the conventional wisdom that a smaller network leads to better generalization.
You may find, however, that neither of these things help generalization. Now trying increasing the number of hidden units to 15 or 20. Did generalization performance improve? How about combining more hidden units with weight decay? I managed to get 98% accuracy on the test set at one point, but I'm not telling how. Ok, I don't remember how. Can you beat 98%?
Sometimes you can get better generalization performance by using more hidden units and weight decay to keep the weights small. That is often true on "fuzzy" tasks like this, but probably won't be true for logical sorts of problems where you want the network to compute a discrete function.
Are the problems you're interested in more often fuzzy or logical? Fuzzy logic? No, no. That's for people who don't know how to use neural nets.