AMNET: Affine Multiplexing Network Toolbox

AMNET is a Python toolbox that assists in building certain kinds of neural networks, and formally verifying their behavior in-the-loop (under development).

Example usage

import numpy as np
from amnet import Variable, Linear, Mu

# a two-dimensional input variable
x = Variable(2, name='x')

# choose components
a1 = Linear(np.array([[1, 0]]), x)
a2 = Linear(np.array([[0, 1]]), x)

# find difference
a3 = Linear(np.array([[-1, 1]]), x)

# if a3 <= 0, returns a1; otherwise a2
phimax = Mu(a1, a2, a3)

# equivalently, we can also write
# phimax = amnet.atoms.max_all(x)

print phimax
print phimax.eval([1, -2]) # returns: 1

References

• I. Papusha, U. Topcu, S. Carr, N. Lauffer. "Affine Multiplexing Networks: System Analysis, Learning, and Computation," arXiv:1805.00164 [math.OC], 2018.