PSD with ANN

Check it out!  Pulse Shape Discrimination with Artificial Neural Networks!  Pretty baller, I know.
The idea is to take a spectra and figure out what radiation caused it.

So we sanitize the network (a high pass filter to remove some of the noise as well as normalization), do some fitting, extract some properties, and then feed those properties to the ANN.  The ANN is essentially a matrix with different function kernels for elements, trained iteratively in a minimization of the error surface (conjugate -gradient like methods).

But here is the real question.  How can the first network train to figure 3 but not train to figure 4?  A five year old could draw that decision surface!

Figure 3

Figure 4

Below are the results, a classifications of 3%.  Not bad; the ORNL paper with the same data got 7%

.