Mode connectivity in the QCBM loss landscape: ICCAD Special Session Paper

Quantum circuit Born machines (QCBMs) and training via variational quantum algorithms (VQAs) are key applications for near-term quantum hardware. QCBM ansätze designs are unique in that they do not require prior knowledge of a physical Hamiltonian. Many ansätze are built from fixed designs. In this work, we train and compare the performance of QCBM models built using two commonly employed parameterizations and two commonly employed entangling layer designs. In addition to comparing the overall performance of these models, we look at features and characteristics of the loss landscape –connectivity of minima in particular – to help understand the advantages and disadvantages of each design choice. We show that the rotational gate choices can improve loss landscape connectivity.

[1]  S. Lloyd,et al.  Quantum Algorithm Providing Exponential Speed Increase for Finding Eigenvalues and Eigenvectors , 1998, quant-ph/9807070.

[2]  G. Henkelman,et al.  Improved tangent estimate in the nudged elastic band method for finding minimum energy paths and saddle points , 2000 .

[3]  Oriol Vinyals,et al.  Qualitatively characterizing neural network optimization problems , 2014, ICLR.

[4]  Jimmy Ba,et al.  Adam: A Method for Stochastic Optimization , 2014, ICLR.

[5]  Yann LeCun,et al.  Eigenvalues of the Hessian in Deep Learning: Singularity and Beyond , 2016, 1611.07476.

[6]  Hao Li,et al.  Visualizing the Loss Landscape of Neural Nets , 2017, NeurIPS.

[7]  Lei Wang,et al.  Differentiable Learning of Quantum Circuit Born Machine , 2018, Physical Review A.

[8]  Edward Grant,et al.  An initialization strategy for addressing barren plateaus in parametrized quantum circuits , 2019, Quantum.

[9]  Marcello Benedetti,et al.  Parameterized quantum circuits as machine learning models , 2019, Quantum Science and Technology.

[10]  Alejandro Perdomo-Ortiz,et al.  A generative modeling approach for benchmarking and training shallow quantum circuits , 2018, npj Quantum Information.

[11]  Peter D. Johnson,et al.  Expressibility and Entangling Capability of Parameterized Quantum Circuits for Hybrid Quantum‐Classical Algorithms , 2019, Advanced Quantum Technologies.

[12]  Kathleen E. Hamilton,et al.  Generative model benchmarks for superconducting qubits , 2018, Physical Review A.

[13]  C. Gogolin,et al.  Evaluating analytic gradients on quantum hardware , 2018, Physical Review A.

[14]  Patrick Huembeli,et al.  Characterizing the loss landscape of variational quantum circuits , 2020, Quantum Science and Technology.

[15]  N. Killoran,et al.  Estimating the gradient and higher-order derivatives on quantum hardware , 2020, 2008.06517.

[16]  Kathleen E. Hamilton,et al.  Mode connectivity in the loss landscape of parameterized quantum circuits , 2021, ArXiv.

[17]  M. Cerezo,et al.  Variational quantum algorithms , 2020, Nature Reviews Physics.

[18]  Patrick J. Coles,et al.  Cost function dependent barren plateaus in shallow parametrized quantum circuits , 2021, Nature Communications.

[19]  Alejandro Perdomo-Ortiz,et al.  Robust implementation of generative modeling with parametrized quantum circuits , 2019, Quantum Machine Intelligence.

[20]  Patrick J. Coles,et al.  Connecting ansatz expressibility to gradient magnitudes and barren plateaus , 2021, PRX Quantum.