John Resig dissects the neural network javascript OCR captcha code

He says that an edge weight in the neural network means "At pixel 9x6 the letter A is 58% likely to be filling in that pixel." That isn't how neural networks work (although, it is fairly close to how bayesian belief nets work). In a neural network the edge weights of the graph don't obviously correspond to anything - they are simply chosen (most commonly through the back propagation training algorithm) to minimize the error between the desired output and actual output on the training data.

The weights in a neural network are simply the coefficients for terms in an equation that, when plotted, produces a curve that's a good fit for a series of data points. Unfortunately, that explanation is not very sexy ;-) That's one of the big problems with neural networks - they're effectively just black boxes that incidentally produce pretty good answers. They are (in most of their standard incarnations) really just a fairly simple technique for regression analysis.

Incidentally, I got to see Resig speak about jQuery yesterday, and he's clearly a very clever guy. I say this with all due respect, and fully expect that this is just a misunderstanding due to me misinterpreting the sentence I quoted in my second paragraph.