Data-centric ML in Quantum Information Science

Author: Sanjaya Lohani

*Please report bugs at slohani@mlphys.com

Thanks to Brian T. Kirby, Ryan T. Glasser, Sean D. Huver and Thomas A. Searles

Preprint:

Lohani, S., Lukens, J.M., Glasser, R.T., Searles, T.A. and Kirby, B.T., 2022. Data-Centric Machine Learning in Quantum Information Science. arXiv preprint arXiv:2201.09134.

Built With

• mlphys
• tensorflow >=2.4
• qiskit

Getting Started

pip install mlphys

Usage

Simulating various distributions and measurements (including inferences)

import mlphys.deepqis.simulator.distributions as dist
import mlphys.deepqis.simulator.measurements as meas
import mlphys.deepqis.utils.Alpha_Measure as find_alpha
import mlphys.deepqis.utils.Concurrence_Measure as find_con
import mlphys.deepqis.utils.Purity_Measure as find_pm
import mlphys.deepqis.network.Inference as inference
import mlphys.deepqis.utils.Fidelity_Measure as fm
...

Sub-module

• deepQis

Tutorials

For examples (google colab), please refer to

• Generating Biased Distributions.
• Inference Examples.

Hands-on coding examples for the results

• Reducing spurious correlations:

  • Accuracy of entanglement-separability classification - Fig 2 (a)
  • network reconstruction fidelity versus the percentage of separable states added to a training set containing entangled states - Fig 2 (b)
  • Reconstruction fidelity for test states from the MA distribution for a MEMS-only trained network and after adding a small fraction of separable states into the training set - Fig 2 (c, d)

• Reconstruction fidelity versus number of trainable parameters for various training set distributions:

  • Data-centric approach (Fidelity versus trainable parameters) - Fig 3 (a)
  • The concurrence and purity of random quantum states from the Hilbert–Schmidt–Haar (HS–Haar), Zyczkowski (Z), ˙ engineered, and IBM Q distributions - Fig 3 (a) insets

• Engineered states on concurrence-purity plane:

  • The engineered and IBM Q sets - Fig 3 (b)

• Data-centric approach in the low-shot regime:

  • Reconstructing the NISQ-sampled distribution with simulated measurements performed with shots ranging from 128 to 8192 - Fig 4

• Heterogeneous state complexity:

  • Two-qubits

    • Reconstruction fidelities versus test state purity - Fig 5 (a) bottom
    • Test MA distribution - Fig 5 (a) top
    • Fidelity versus K parameter - Fig 5 (a) inset
    • Zoomed in at the crossing point - Fig 5 (a) inset

  • Three-qubits

    • Reconstruction fidelities versus test state purity - Fig 5 (b) bottom
    • Test MA distribution - Fig 5 (b) top
    • Fidelity versus K parameter - Fig 5 (b) inset
    • Zoomed in at the crossing point - Fig 5 (b) inset

  • Four-qubits

    • Reconstruction fidelities versus test state purity - Fig 5 (c) bottom
    • Test MA distribution - Fig 5 (c) top
    • Fidelity versus K parameter - Fig 5 (c) inset
    • Zoomed in at the crossing point - Fig 5 (c) inset

• Optimizing learning rate:

  • Fidelity of reconstructed density matrices versus learning rate - Extended Data Fig 1
  • The full purity distributions of the reconstructed states - Extended Data Fig 1 inset

• Engineered states:

  • Unfiltered - Extended Data Fig. 2 (a) left
  • Engineered - Extended Data Fig. 2 (a) right
  • Reconstruction fidelities versus the value of K - Fig 7 (b)

• Reconstruction fidelity of NISQ-sampled test set versus the mean purity of various MA-distributed training states when K = 4.

  • The mean purity of the training set matches the minimum and mean purity of the NISQ sampled states when D = K = 4 - Extended Data Fig. 3

• Reconstruction fidelity versus trainable parameters for various MA-distributed training sets:

  • The pairs of concentration parameter and K-value are chosen as (α, K) ∈ {(0.01, 4),(0.1, 4),(0.3, 4),(0.8, 4),(0.3394, 6)} for training sets - Extended Data Fig. 4