Mitigating a discrete sign problem with extreme learning machines

An extreme learning machine is a neural network in which only the weights in the last layer are changed during training; for such networks training can be performed efficiently and deterministically. We use an extreme learning machine to construct a control variate that tames the sign problem in the classical Ising model at imaginary external magnetic field. Using this control variate, we directly compute the partition function at imaginary magnetic field in two and three dimensions, yielding information on the positions of Lee-Yang zeros.