replace 50 with max_depth

[ker2x]

if you replace 50 with max_depth you get a better result. even, with the simple NN.
You still need A LOT of training point however.

# function helper, don't directly call
def _m(a, max_depth):
  z = 0
  for n in range(1, max_depth):
    z = z**2 + a
    if abs(z) > 2:
      return min((n-1)/50, 1) # <- here
      # return -(math.cos((n-1)*math.pi/50)/2) + (1/2)
  return 1.0

https://youtu.be/-HbFdvVFyXU

[ker2x]

tbh, even 0 & 1 works just as well too.