Potential NaNs due to div by zero when normalising

[beyarkay]

So these lines in the update method:

for i in 0..self.data.x {
    for j in 0..self.data.y {
        for k in 0..self.data.z {
            self.data.map[[i, j, k]] += (elem[[k]] - self.data.map[[i, j, k]]) * g[[i, j]];
        }

        let norm = norm(self.data.map.index_axis(Axis(0), i).index_axis(Axis(0), j));
        for k in 0..self.data.z {
            self.data.map[[i, j, k]] /= norm;
        }
    }
}

Were causing me issues because norm was ending up as zero, causing a divide by zero to make self.data.map[[i, j, k]] be f64::NaN and resulting in funky results later down the line (I've got NaN values in my input features)

I'm not sure what the purpose of the normalization is? I understand that it would ensure each neuron's weights sum to 1, but I can't find where this is recommended.

On my own fork I've wrapped the normalisation with a check to make sure norm >0 and that seems to have solved the issues, although I'm not sure how valid it is.