> Chapter 7 : with 10,000 examples, you will have a good chance of detecting an improvement of 0.1% That's not true. Accuracy based on assuming an underlying

> Chapter 7 : with 10,000 examples, you will have a good chance of detecting an improvement of 0.1% That's not true. Accuracy based on assuming an underlying binomial distribution has standard deviation sqrt( p (1 - p) / n ). Unless your accuracy is very near to 0 or 1, you're going to need more like a million samples to have a good chance of detecting a difference of 0.001 and be able to reject the hypothesis that it was not caused by statistical noise. If the assumption is not an underlying binomial, can you explain why? I was hoping for more a lot more hard maths in this text.