Practical Neural Network Recipies in C++

Confidence

Measures

381

the side of being too large. This helps to smooth out the effects of

sampling error. Even more importantly, it tends to broaden the

rejection tail, resulting in more conservative confidence measures.

Examination of the above graphs shows that blurring is very signifi-

cant at σ = 0.1, so this should probably be an upper limit.

More definitive statements can be made, though. If

we

include

σ in the definition of the Gaussian window function:

-(-)

2

20-8

W(d) = e °

we can rewrite this formula to tell us the difference in activation levels

needed to reduce the influence of a sample point to a given fraction of

the maximum