Fix default noise level for NoiseAugmentation

[adrhill]

The previous std was 0.1 instead of the documented 1.0.
A value of one is more in line with the recommendation from the SmoothGrad paper, which has been added to the docstring:

For optimal results, $REF_SMILKOV_SMOOTHGRAD recommends setting `std` between 10% and 20% of the input range of every sample,
e.g. `std = 0.1 * (maximum(input) - minimum(input))`.

[adrhill]

The fixed default of std=1 should be in line with the recommendation from the SmoothGrad paper according to a simple numerical experiment, assuming normalized inputs:

julia> d = Normal(0, 1)
Normal{Float64}(μ=0.0, σ=1.0)

julia> function sample_range(n) # of normalized n-dim. input
         e = extrema(rand(d, n))
         return e[2] - e[1]
       end
sample_range (generic function with 2 methods)

julia> mean(sample_range(224*224*3) for _ in 1:1000)
8.939849898946928

julia> mean(sample_range(224*224*3) for _ in 1:1000)
8.959057153975865

julia> mean(sample_range(224*224*3) for _ in 1:1000)
8.943506279873574

scaled by 10-20%, std=1 appears to be adequate. CC @nealmckee.

[nealmckee]

While the actual noise might be similar to the recommended level over the entire dataset, the original intention of Smoothgrad as I read it was to have a constant signal to noise ratio for each image, where they determine the noise ratio by looking at the maximum amplitude across the entire image and all channels.

Whether this is the best way to get a measure of the signal amplitude and whether it makes sense to adjust the noise level per image rather than per dataset is up for discussion, but if the intention is to have a standardized implementation, this would be it if one were to base it on the text of the paper.

It definitely makes sense to adjust the noise with respect to the data in some way, and if it’s per image it would make sense to bake it into the analyzer. Per dataset it’s reasonable to let the user enter it manually, though it has the downside of having less universal defaults.

[adrhill]

the original intention of Smoothgrad as I read it was to have a constant signal to noise ratio for each image

Yes, I hope the new docstring makes that clear.

Since NoiseAugmentation is a general purpose wrapper around any analyzer and any sort of Sampleable the user defines (not just Gaussian distributions), simply specifying a "noise level" isn't possible.

I will decouple SmoothGrad from NoiseAugmentation in a follow-up PR.