Improving training schemes for encoding quantum states on neuromorphic hardware

In recent years neural network quantum states have been
successfully introduced as a new type of ansatz for simulating many-
body systems. While the focus has been mostly on artificial neural
networks, advances in specialized neuromorphic hardware promise to
exceed the capabilities of von Neumann computation in terms of sampling
speed and energy efficiency.
The high-fidelity simulation of entangled quantum states on the spike-
based BrainScalesS mixed-signal chips has recently been demonstrated
[1]. Here we aim to improve the training scheme used in this work and
explore applications to larger classes of states.
Based on a detailed understanding of the neuromorphic sample
distribution we optimize the mapping from quantum states to probability
distributions in order to improve learning performance and signal-to-
noise ratios. We test our algorithms on groundstates of well-known spin
Hamiltonians as well as steady states of their dynamics. We are able to
scale the approach up to system sizes beyond the previously achieved
ones [1].
[1] S. Czischek, A. Baumbach, S. Billaudelle, B. Cramer, L. Kades, J.
M. Pawlowski, M. K. Oberthaler, J. Schemmel, M. A. Petrovici, T.
Gasenzer, and M. Gärttner, Spiking neuromophic chip learns entangled
quantum states, arXiv:2008.01039 [cs.ET]