Generalized Hebbian Algoithm (GHA) stability issues

The GHA proposed by Sanger in 1989 is supposed to be very numerically stable. There are some implementations of it floating around, for example this one https://github.com/matwey/python-gha.

In that test example the implementation is working fine.

What happens is, if you send in for example MNIST data into the same code linked above and run it through a couple of data points the weight matrix quickly overflows to nan. Input size is obviously adjusted to 784 instead of 5 like in the test code.

Why does this happens, do anyone know any limitation to this algorithm. There are a lot of research using this rule in this way and I just can’t get it to work. What am I missing?