Papers
Topics
Authors
Recent
Search
2000 character limit reached

Barren Plateaus in Variational Quantum Computing

Published 1 May 2024 in quant-ph, cs.LG, and stat.ML | (2405.00781v2)

Abstract: Variational quantum computing offers a flexible computational paradigm with applications in diverse areas. However, a key obstacle to realizing their potential is the Barren Plateau (BP) phenomenon. When a model exhibits a BP, its parameter optimization landscape becomes exponentially flat and featureless as the problem size increases. Importantly, all the moving pieces of an algorithm -- choices of ansatz, initial state, observable, loss function and hardware noise -- can lead to BPs when ill-suited. Due to the significant impact of BPs on trainability, researchers have dedicated considerable effort to develop theoretical and heuristic methods to understand and mitigate their effects. As a result, the study of BPs has become a thriving area of research, influencing and cross-fertilizing other fields such as quantum optimal control, tensor networks, and learning theory. This article provides a comprehensive review of the current understanding of the BP phenomenon.

Definition Search Book Streamline Icon: https://streamlinehq.com
References (108)
  1. M.Ā Schuld, I.Ā Sinayskiy,Ā andĀ F.Ā Petruccione,Ā An introduction to quantum machine learning,Ā Contemporary PhysicsĀ 56,Ā 172 (2015).
  2. L.Ā BittelĀ andĀ M.Ā Kliesch,Ā Training variational quantum algorithms is NP-hard,Ā Phys. Rev. Lett.Ā 127,Ā 120502 (2021).
  3. E.Ā R.Ā AnschuetzĀ andĀ B.Ā T.Ā Kiani,Ā Beyond barren plateaus: Quantum variational algorithms are swamped with traps,Ā Nature CommunicationsĀ 13,Ā 7760 (2022a).
  4. E.Ā R.Ā Anschuetz,Ā Critical points in quantum generative models,Ā International Conference on Learning RepresentationsĀ  (2022).
  5. P.Ā Bermejo, B.Ā Aizpurua,Ā andĀ R.Ā OrĆŗs,Ā Improving gradient methods via coordinate transformations: Applications to quantum machine learning,Ā Physical Review ResearchĀ 6,Ā 023069 (2024).
  6. S.Ā Sim, P.Ā D.Ā Johnson,Ā andĀ A.Ā Aspuru-Guzik,Ā Expressibility and entangling capability of parameterized quantum circuits for hybrid quantum-classical algorithms,Ā Advanced Quantum TechnologiesĀ 2,Ā 1900070 (2019).
  7. Q.Ā MiaoĀ andĀ T.Ā Barthel,Ā Equivalence of cost concentration and gradient vanishing for quantum circuits: an elementary proof in the riemannian formulation,Ā arXiv preprint arXiv:2402.07883Ā  (2024).
  8. M.Ā CerezoĀ andĀ P.Ā J.Ā Coles,Ā Higher order derivatives of quantum neural networks with barren plateaus,Ā Quantum Science and TechnologyĀ 6,Ā 035006 (2021).
  9. J.Ā L.Ā CybulskiĀ andĀ T.Ā Nguyen,Ā Impact of barren plateaus countermeasures on the quantum neural network capacity to learn,Ā Quantum Information ProcessingĀ 22,Ā 442 (2023).
  10. A.Ā PĆ©rez-Salinas, H.Ā Wang,Ā andĀ X.Ā Bonet-Monroig,Ā Analyzing variational quantum landscapes with information content,Ā npj Quantum InformationĀ 10,Ā 27 (2024).
  11. S.Ā OkumuraĀ andĀ M.Ā Ohzeki,Ā Fourier coefficient of parameterized quantum circuits and barren plateau problem,Ā arXiv preprint arXiv:2309.06740Ā  (2023).
  12. N.Ā A.Ā Nemkov, E.Ā O.Ā Kiktenko,Ā andĀ A.Ā K.Ā Fedorov,Ā Fourier expansion in variational quantum algorithms,Ā Phys. Rev. AĀ 108,Ā 032406 (2023).
  13. C.Ā O.Ā Marrero, M.Ā KieferovĆ”,Ā andĀ N.Ā Wiebe,Ā Entanglement-induced barren plateaus,Ā PRX QuantumĀ 2,Ā 040316 (2021).
  14. D.Ā GarcĆ­a-MartĆ­n, M.Ā Larocca,Ā andĀ M.Ā Cerezo,Ā Deep quantum neural networks form gaussian processes,Ā arXiv preprint arXiv:2305.09957Ā  (2023).
  15. A.Ā A.Ā Mele,Ā Introduction to haar measure tools in quantum information: A beginnerā€™s tutorial,Ā arXiv preprint arXiv:2307.08956Ā  (2023).
  16. A.Ā UvarovĀ andĀ J.Ā D.Ā Biamonte,Ā On barren plateaus and cost function locality in variational quantum algorithms,Ā Journal of Physics A: Mathematical and TheoreticalĀ 54,Ā 245301 (2021).
  17. T.Ā BarthelĀ andĀ Q.Ā Miao,Ā Absence of barren plateaus and scaling of gradients in the energy optimization of isometric tensor network states,Ā arXiv preprint arXiv:2304.00161Ā  (2023).
  18. Q.Ā MiaoĀ andĀ T.Ā Barthel,Ā Isometric tensor network optimization for extensive hamiltonians is free of barren plateaus,Ā arXiv preprint arXiv:2304.14320Ā  (2023).
  19. A.Ā Letcher, S.Ā Woerner,Ā andĀ C.Ā Zoufal,Ā Tight and efficient gradient bounds for parameterized quantum circuits,Ā arXiv preprint arXiv:2309.12681Ā  (2023).
  20. H.-K.Ā Zhang, S.Ā Liu,Ā andĀ S.-X.Ā Zhang,Ā Absence of barren plateaus in finite local-depth circuits with long-range entanglement,Ā Physical Review LettersĀ 132,Ā 150603 (2024a).
  21. J.Ā Napp,Ā Quantifying the barren plateau phenomenon for a model of unstructured variational ansƤtze,Ā arXiv preprint arXiv:2203.06174Ā  (2022).
  22. C.Ā ZhaoĀ andĀ X.-S.Ā Gao,Ā Analyzing the barren plateau phenomenon in training quantum neural networks with the ZX-calculus,Ā QuantumĀ 5,Ā 466 (2021).
  23. C.-Y.Ā ParkĀ andĀ N.Ā Killoran,Ā Hamiltonian variational ansatz without barren plateaus,Ā QuantumĀ 8,Ā 1239 (2024).
  24. C.-Y.Ā Park, M.Ā Kang,Ā andĀ J.Ā Huh,Ā Hardware-efficient ansatz without barren plateaus in any depth,Ā arXiv preprint arXiv:2403.04844Ā  (2024).
  25. M.Ā Schumann, F.Ā K.Ā Wilhelm,Ā andĀ A.Ā Ciani,Ā Emergence of noise-induced barren plateaus in arbitrary layered noise models,Ā arXiv preprint arXiv:2310.08405Ā  (2023).
  26. P.Ā SingkanipaĀ andĀ D.Ā A.Ā Lidar,Ā Beyond unital noise in variational quantum algorithms: noise-induced barren plateaus and fixed points,Ā arXiv preprint arXiv:2402.08721Ā  (2024).
  27. R.Ā ZeierĀ andĀ T.Ā Schulte-HerbrĆ¼ggen,Ā Symmetry principles in quantum systems theory,Ā Journal of mathematical physicsĀ 52,Ā 113510 (2011).
  28. M.Ā J.Ā Bremner, C.Ā Mora,Ā andĀ A.Ā Winter,Ā Are random pure states useful for quantum computation?,Ā Physical review lettersĀ 102,Ā 190502 (2009).
  29. D.Ā Gross, S.Ā T.Ā Flammia,Ā andĀ J.Ā Eisert,Ā Most quantum states are too entangled to be useful as computational resources,Ā Physical review lettersĀ 102,Ā 190501 (2009).
  30. M.Ā A.Ā NielsenĀ andĀ I.Ā L.Ā Chuang,Ā Quantum Computation and Quantum InformationĀ (Cambridge University Press,Ā Cambridge,Ā 2000).
  31. J.Ā Horgan,Ā Scott aaronson answers every ridiculously big question i throw at him,Ā Scientific American. https://blogs. scientificamerican. com/cross-check/scott-aaronson-answers-every-ridiculously-big-question-i-throw-at-himĀ 21 (2016).
  32. Y.Ā Liu, S.Ā Arunachalam,Ā andĀ K.Ā Temme,Ā A rigorous and robust quantum speed-up in supervised machine learning,Ā Nature PhysicsĀ ,Ā 1 (2021).
  33. J.Ā JƤgerĀ andĀ R.Ā V.Ā Krems,Ā Universal expressiveness of variational quantum classifiers and quantum kernels for support vector machines,Ā Nature CommunicationsĀ 14,Ā 576 (2023).
  34. L.Ā FriedrichĀ andĀ J.Ā Maziero,Ā Quantum neural network cost function concentration dependency on the parametrization expressivity,Ā Scientific ReportsĀ 13,Ā 9978 (2023).
  35. T.Ā Haug, K.Ā Bharti,Ā andĀ M.Ā S.Ā Kim,Ā Capacity and Quantum Geometry of Parametrized Quantum Circuits,Ā PRX QuantumĀ 2,Ā 040309 (2021).
  36. Technically speaking, group modules are irreducible representations of the dynamical Lie group over the vector space ā„¬ā¢(ā„‹0)ā„¬subscriptā„‹0\mathcal{B}(\mathcal{H}_{0})caligraphic_B ( caligraphic_H start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT ).
  37. A.Ā Uvarov, J.Ā D.Ā Biamonte,Ā andĀ D.Ā Yudin,Ā Variational quantum eigensolver for frustrated quantum systems,Ā Physical Review BĀ 102,Ā 075104 (2020).
  38. M.Ā KashifĀ andĀ S.Ā Al-Kuwari,Ā The impact of cost function globality and locality in hybrid quantum neural networks on nisq devices,Ā Machine Learning: Science and TechnologyĀ 4,Ā 015004 (2023a).
  39. O.Ā Ogunkoya, K.Ā Morris,Ā andĀ D.Ā M.Ā KĆ¼rkĆ§Ć¼oglu,Ā Investigating parameter trainability in the snap-displacement protocol of a qudit system,Ā arXiv preprint arXiv:2309.14942Ā  (2023).
  40. B.Ā ZhangĀ andĀ Q.Ā Zhuang,Ā Energy-dependent barren plateau in bosonic variational quantum circuits,Ā arXiv preprint arXiv:2305.01799Ā  (2023).
  41. R.Ā ShaydulinĀ andĀ S.Ā M.Ā Wild,Ā Importance of kernel bandwidth in quantum machine learning,Ā Physical Review AĀ 106,Ā 042407 (2022).
  42. S.Ā Das, S.Ā Martina,Ā andĀ F.Ā Caruso,Ā The role of data embedding in equivariant quantum convolutional neural networks,Ā arXiv preprint arXiv:2312.13250Ā  (2023).
  43. D.Ā StilckĀ FranƧaĀ andĀ R.Ā Garcia-Patron,Ā Limitations of optimization algorithms on noisy quantum devices,Ā Nature PhysicsĀ 17,Ā 1221 (2021).
  44. M.Ā M.Ā Wilde,Ā Quantum information theoryĀ (Cambridge University Press,Ā 2013).
  45. D.Ā GarcĆ­a-MartĆ­n, M.Ā Larocca,Ā andĀ M.Ā Cerezo,Ā Effects of noise on the overparametrization of quantum neural networks,Ā Phys. Rev. Res.Ā 6,Ā 013295 (2024).
  46. D.Ā Wecker, M.Ā B.Ā Hastings,Ā andĀ M.Ā Troyer,Ā Progress towards practical quantum variational algorithms,Ā Physical Review AĀ 92,Ā 042303 (2015).
  47. E.Ā Farhi, J.Ā Goldstone,Ā andĀ S.Ā Gutmann,Ā A quantum approximate optimization algorithm,Ā arXiv preprint arXiv:1411.4028Ā  (2014).
  48. A.Ā BƤrtschiĀ andĀ S.Ā Eidenbenz,Ā Grover mixers for qaoa: Shifting complexity from mixer design to state preparation,Ā inĀ 2020 IEEE International Conference on Quantum Computing and Engineering (QCE)Ā (IEEE,Ā 2020)Ā pp.Ā 72ā€“82.
  49. B.Ā Zhang, A.Ā Sone,Ā andĀ Q.Ā Zhuang,Ā Quantum computational phase transition in combinatorial problems,Ā npj Quantum InformationĀ 8,Ā 1 (2022b).
  50. A.Ā G.Ā TaubeĀ andĀ R.Ā J.Ā Bartlett,Ā New perspectives on unitary coupled-cluster theory,Ā International journal of quantum chemistryĀ 106,Ā 3393 (2006).
  51. R.Ā Mao, G.Ā Tian,Ā andĀ X.Ā Sun,Ā Barren plateaus of alternated disentangled ucc ansatzs (2023).
  52. S.Ā C.Ā Marshall, C.Ā Gyurik,Ā andĀ V.Ā Dunjko,Ā High dimensional quantum machine learning with small quantum computers,Ā QuantumĀ 7,Ā 1078 (2023).
  53. T.Ā VolkoffĀ andĀ P.Ā J.Ā Coles,Ā Large gradients via correlation in random parameterized quantum circuits,Ā Quantum Science and TechnologyĀ 6,Ā 025008 (2021).
  54. S.Ā Kazi, M.Ā Larocca,Ā andĀ M.Ā Cerezo,Ā On the universality of snsubscriptš‘ š‘›s_{n}italic_s start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT-equivariant kš‘˜kitalic_k-body gates,Ā arXiv preprint arXiv:2303.00728Ā  (2023).
  55. R.Ā JozsaĀ andĀ A.Ā Miyake,Ā Matchgates and classical simulation of quantum circuits,Ā Proceedings of the Royal Society A: Mathematical, Physical and Engineering SciencesĀ 464,Ā 3089 (2008).
  56. F.Ā DeĀ Melo, P.Ā Ä†wikliński,Ā andĀ B.Ā M.Ā Terhal,Ā The power of noisy fermionic quantum computation,Ā New Journal of PhysicsĀ 15,Ā 013015 (2013).
  57. T.Ā J.Ā Volkoff,Ā Efficient trainability of linear optical modules in quantum optical neural networks,Ā Journal of Russian Laser ResearchĀ 42,Ā 250 (2021).
  58. T.Ā Volkoff, Z.Ā Holmes,Ā andĀ A.Ā Sornborger,Ā Universal compiling and (no-)free-lunch theorems for continuous-variable quantum learning,Ā PRX QuantumĀ 2,Ā 040327 (2021).
  59. M.Ā Arjovsky, A.Ā Shah,Ā andĀ Y.Ā Bengio,Ā Unitary evolution recurrent neural networks,Ā inĀ International conference on machine learningĀ (PMLR,Ā 2016)Ā pp.Ā 1120ā€“1128.
  60. A.Ā KulshresthaĀ andĀ I.Ā Safro,Ā Beinit: Avoiding barren plateaus in variational quantum algorithms,Ā inĀ 2022 IEEE International Conference on Quantum Computing and Engineering (QCE)Ā (IEEE,Ā 2022)Ā pp.Ā 197ā€“203.
  61. A.Ā Rad, A.Ā Seif,Ā andĀ N.Ā M.Ā Linke,Ā Surviving the barren plateau in variational quantum circuits with bayesian learning initialization,Ā arXiv preprint arXiv:2203.02464Ā  (2022).
  62. T.Ā HaugĀ andĀ M.Ā Kim,Ā Optimal training of variational quantum algorithms without barren plateaus,Ā arXiv preprint arXiv:2104.14543Ā  (2021).
  63. X.Ā ShiĀ andĀ Y.Ā Shang,Ā Avoiding barren plateaus via gaussian mixture model,Ā arXiv preprint arXiv:2402.13501Ā  (2024).
  64. L.Ā FriedrichĀ andĀ J.Ā Maziero,Ā Avoiding barren plateaus with classical deep neural networks,Ā Physical Review AĀ 106,Ā 042433 (2022).
  65. G.Ā Marin-Sanchez, J.Ā Gonzalez-Conde,Ā andĀ M.Ā Sanz,Ā Quantum algorithms for approximate function loading,Ā Physical Review ResearchĀ 5,Ā 033114 (2023).
  66. A.Ā Cervera-Lierta, J.Ā S.Ā Kottmann,Ā andĀ A.Ā Aspuru-Guzik,Ā The meta-variational quantum eigensolver (meta-vqe): Learning energy profiles of parameterized Hamiltonians for quantum simulation,Ā PRX QuantumĀ 2,Ā 020329 (2021).
  67. J.Ā WurtzĀ andĀ D.Ā Lykov,Ā Fixed-angle conjectures for the quantum approximate optimization algorithm on regular maxcut graphs,Ā Physical Review AĀ 104,Ā 052419 (2021).
  68. S.Ā BoulebnaneĀ andĀ A.Ā Montanaro,Ā Predicting parameters for the quantum approximate optimization algorithm for max-cut from the infinite-size limit,Ā arXiv preprint arXiv:2110.10685Ā  (2021).
  69. E.Ā Campos, A.Ā Nasrallah,Ā andĀ J.Ā Biamonte,Ā Abrupt transitions in variational quantum circuit training,Ā Physical Review AĀ 103,Ā 032607 (2021b).
  70. G.Ā Acampora, A.Ā Chiatto,Ā andĀ A.Ā Vitiello,Ā A comparison of evolutionary algorithms for training variational quantum classifiers,Ā inĀ 2023 IEEE Congress on Evolutionary Computation (CEC)Ā (IEEE,Ā 2023)Ā pp.Ā 1ā€“8.
  71. B.Ā KoczorĀ andĀ S.Ā C.Ā Benjamin,Ā Quantum natural gradient generalized to noisy and nonunitary circuits,Ā Physical Review AĀ 106,Ā 062416 (2022).
  72. K.Ā Temme, S.Ā Bravyi,Ā andĀ J.Ā M.Ā Gambetta,Ā Error mitigation for short-depth quantum circuits,Ā Physical review lettersĀ 119,Ā 180509 (2017).
  73. Y.Ā LiĀ andĀ S.Ā C.Ā Benjamin,Ā Efficient variational quantum simulator incorporating active error minimization,Ā Phys. Rev. XĀ 7,Ā 021050 (2017).
  74. K.Ā Wang, Y.-A.Ā Chen,Ā andĀ X.Ā Wang,Ā Mitigating quantum errors via truncated neumann series,Ā Science China Information SciencesĀ 66,Ā 180508 (2023b).
  75. R.Ā Takagi, H.Ā Tajima,Ā andĀ M.Ā Gu,Ā Universal sampling lower bounds for quantum error mitigation,Ā Physical Review LettersĀ 131,Ā 210602 (2023).
  76. J.-G.Ā LiuĀ andĀ L.Ā Wang,Ā Differentiable learning of quantum circuit born machines,Ā Phys. Rev. AĀ 98,Ā 062324 (2018).
  77. C.Ā Zoufal,Ā Generative quantum machine learning,Ā arXiv preprint arXiv:2111.12738Ā  (2021).
  78. S.Ā LloydĀ andĀ C.Ā Weedbrook,Ā Quantum generative adversarial learning,Ā Physical Review LettersĀ 121,Ā 040502 (2018).
  79. L.Ā CoopmansĀ andĀ M.Ā Benedetti,Ā On the sample complexity of quantum boltzmann machine learning,Ā arXiv preprint arXiv:2306.14969Ā  (2023).
  80. M.Ā Kieferova, O.Ā M.Ā Carlos,Ā andĀ N.Ā Wiebe,Ā Quantum generative training using rĆ©nyi divergences,Ā arXiv preprint arXiv:2106.09567Ā  (2021).
  81. J.Ā KĆ¼bler, S.Ā Buchholz,Ā andĀ B.Ā Schƶlkopf,Ā The inductive bias of quantum kernels,Ā Advances in Neural Information Processing SystemsĀ 34,Ā 12661 (2021).
  82. Y.Ā SuzukiĀ andĀ M.Ā Li,Ā Effect of alternating layered ansatzes on trainability of projected quantum kernel,Ā arXiv preprint arXiv:2310.00361Ā  (2023).
  83. Y.Ā Suzuki, H.Ā Kawaguchi,Ā andĀ N.Ā Yamamoto,Ā Quantum fisher kernel for mitigating the vanishing similarity issue,Ā arXiv preprint arXiv:2210.16581Ā  (2022).
  84. M.Ā SchuldĀ andĀ F.Ā Petruccione,Ā Supervised learning with quantum computers,Ā Vol.Ā 17Ā (Springer,Ā 2018).
  85. M.Ā Schuld,Ā Supervised quantum machine learning models are kernel methods,Ā arXiv preprint arXiv:2101.11020Ā  (2021).
  86. B.Ā T.Ā Kiani, S.Ā Lloyd,Ā andĀ R.Ā Maity,Ā Learning unitaries by gradient descent,Ā arXiv preprint arXiv:2001.11897Ā  (2020).
  87. X.Ā Ge, R.-B.Ā Wu,Ā andĀ H.Ā Rabitz,Ā The optimization landscape of hybrid quantumā€“classical algorithms: From quantum control to NISQ applications,Ā Annual Reviews in ControlĀ https://doi.org/10.1016/j.arcontrol.2022.06.001 (2022).
  88. L.Ā BroersĀ andĀ L.Ā Mathey,Ā Mitigated barren plateaus in the time-nonlocal optimization of analog quantum-algorithm protocols,Ā Physical Review ResearchĀ 6,Ā 013076 (2024).
  89. H.-X.Ā Tao, J.Ā Hu,Ā andĀ R.-B.Ā Wu,Ā Unleashing the expressive power of pulse-based quantum neural networks,Ā arXiv preprint arXiv:2402.02880Ā  (2024).
  90. E.Ā C.Ā MartĆ­n, K.Ā Plekhanov,Ā andĀ M.Ā Lubasch,Ā Barren plateaus in quantum tensor network optimization,Ā QuantumĀ 7,Ā 974 (2023).
  91. V.Ā Feldman,Ā Statistical query learning,Ā inĀ Encyclopedia of Algorithms,Ā edited byĀ M.-Y.Ā KaoĀ (Springer New York,Ā New York, NY,Ā 2016)Ā pp.Ā 2090ā€“2095.
  92. A.Ā Angrisani,Ā Learning unitaries with quantum statistical queries,Ā arXiv preprint arXiv:2310.02254Ā  (2023).
  93. C.Ā WadhwaĀ andĀ M.Ā Doosti,Ā Learning quantum processes with quantum statistical queries,Ā arXiv preprint arXiv:2310.02075Ā  (2023).
  94. A.Ā Nietner,Ā Unifying (quantum) statistical and parametrized (quantum) algorithms,Ā arXiv preprint arXiv:2310.17716Ā  (2023).
  95. L.Ā YangĀ andĀ N.Ā Engelhardt,Ā The complexity of learning (pseudo) random dynamics of black holes and other chaotic systems,Ā arXiv preprint arXiv:2302.11013Ā  (2023).
  96. E.Ā R.Ā AnschuetzĀ andĀ B.Ā T.Ā Kiani,Ā Quantum variational algorithms are swamped with traps,Ā Nature CommunicationsĀ 13,Ā 7760 (2022b).
  97. X.Ā YouĀ andĀ X.Ā Wu,Ā Exponentially many local minima in quantum neural networks,Ā inĀ International Conference on Machine LearningĀ (PMLR,Ā 2021)Ā pp.Ā 12144ā€“12155.
  98. A.Ā TikkuĀ andĀ I.Ā H.Ā Kim,Ā Circuit depth versus energy in topologically ordered systems,Ā arXiv preprint arXiv:2210.06796Ā  (2022).
  99. K.Ā BhartiĀ andĀ T.Ā Haug,Ā Iterative quantum-assisted eigensolver,Ā Physical Review AĀ 104,Ā L050401 (2021a).
  100. K.Ā BhartiĀ andĀ T.Ā Haug,Ā Quantum-assisted simulator,Ā Physical Review AĀ 104,Ā 042418 (2021b).
  101. C.Ā Gyurik, R.Ā Molteni,Ā andĀ V.Ā Dunjko,Ā Limitations of measure-first protocols in quantum machine learning,Ā arXiv preprint arXiv:2311.12618Ā  (2023).
  102. X.Ā GlorotĀ andĀ Y.Ā Bengio,Ā Understanding the difficulty of training deep feedforward neural networks,Ā inĀ Proceedings of the thirteenth international conference on artificial intelligence and statisticsĀ (JMLR Workshop and Conference Proceedings,Ā 2010)Ā pp.Ā 249ā€“256.
  103. S.Ā IoffeĀ andĀ C.Ā Szegedy,Ā Batch normalization: Accelerating deep network training by reducing internal covariate shift,Ā inĀ International conference on machine learningĀ (pmlr,Ā 2015)Ā pp.Ā 448ā€“456.
  104. J.Ā L.Ā Ba, J.Ā R.Ā Kiros,Ā andĀ G.Ā E.Ā Hinton,Ā Layer normalization,Ā arXiv preprint arXiv:1607.06450Ā  (2016).
  105. V.Ā NairĀ andĀ G.Ā E.Ā Hinton,Ā Rectified linear units improve restricted boltzmann machines,Ā inĀ Proceedings of the 27th international conference on machine learning (ICML-10)Ā (2010)Ā pp.Ā 807ā€“814.
  106. S.Ā Hochreiter,Ā The vanishing gradient problem during learning recurrent neural nets and problem solutions,Ā International Journal of Uncertainty, Fuzziness and Knowledge-Based SystemsĀ 6,Ā 107 (1998).
  107. S.Ā HochreiterĀ andĀ J.Ā Schmidhuber,Ā Long short-term memory,Ā Neural computationĀ 9,Ā 1735 (1997).
  108. L.Ā BairdĀ andĀ A.Ā Moore,Ā Gradient descent for general reinforcement learning,Ā Advances in neural information processing systemsĀ 11 (1998).
Citations (64)
View on Semantic Scholar

Summary

  • The paper finds that barren plateaus, arising from high-dimensional parameter spaces and circuit expressivity, result in vanishing gradients during quantum circuit training.
  • The study shows that noise and decoherence further exacerbate these effects, significantly hindering optimization in variational quantum computing.
  • The paper explores mitigation strategies such as layerwise training, smart initialization, and adaptive circuit design to enhance the scalability of VQCs.

Understanding Barren Plateaus in Variational Quantum Computing

What Are Barren Plateaus?

Barren Plateaus (BPs) in the context of variational quantum computing (VQC) refer to regions within the parameter landscape of quantum circuits where the gradient of the objective function essentially vanishes, making optimization exceedingly difficult. As quantum circuit depth increases or as the number of qubits gets larger, landscapes tend to flatten out exponentially, leading to these barren plateaus.

Why Do They Matter?

Experts in quantum computing are enthusiastic about VQCs because of their potential to solve complex problems more efficiently than classical computers, particularly in fields like chemistry and cryptography. BPs are a significant hurdle in realizing this potential because they make training quantum models challenging, resulting in slower learning and poorer performance.

How Do Barren Plateaus Occur?

Several factors contribute to the emergence of barren plateaus:

  • Increase in System Size: As the number of qubits increases, the dimensionality of the Hilbert space (quantum state space) they span grows exponentially. High-dimensional spaces often exhibit properties where randomly chosen vectors (states) are nearly orthogonal, leading to very small gradients.
  • High Expressivity of Quantum Circuits: When quantum circuits are capable of expressing a very large set of unitary transformations (making them highly expressive), they explore the Hilbert space very thoroughly. Ironically, this thorough exploration can lead to averaging effects where the gradients of the objective function average out to zero, leading to barren plateaus.
  • Noise and Decoherence: Practical quantum devices are noisy. Research indicates that certain types of noise and decoherence can exacerbate the flatness of the optimization landscape, deepening barren plateaus.

Practical Consequences

If not properly addressed, barren plateaus can severely limit the scalability of VQCs. This limitation is not just about longer training times but also concerns the feasibility of obtaining meaningful quantum computational advantages over classical approaches.

Strategies to Mitigate Barren Plateaus

Several techniques have been proposed and tested to mitigate or avoid barren plateaus:

  • Layerwise Training: Training shallow layers of the quantum circuit first and gradually increasing the depth.
  • Smart Initialization: Choosing initial parameters carefully, potentially using classical pre-training methods to find a good starting point.
  • Use of Local Costs: Designing cost functions that depend only on a subset of qubits or employ local observables.
  • Adaptive Circuit Design: Dynamically adjusting the circuit's architecture during training to navigate away from flat regions.

Future Directions

Understanding barren plateaus is an ongoing area of research. Future work includes developing more effective strategies for initialization and adaptive circuit design, exploring the role of symmetries and other inductive biases in preventing barren plateaus, and quantifying the impact of noise on training dynamics. Further theoretical insights into the structure of high-dimensional quantum landscapes will also be crucial for devising new algorithms that are inherently resistant to barren plateaus.

Conclusion

Barren plateaus represent a fundamental challenge in the field of quantum computing, particularly affecting the trainability of variational quantum algorithms. By continuing to address these challenges through innovative strategies and deeper understanding, researchers aim to unlock the full potential of quantum computing technologies. As this field evolves, overcoming barren plateaus will be critical for achieving practical quantum advantage.

Paper to Video (Beta)

No one has generated a video about this paper yet.

Sign Up to Generate All Videos Subscribe on YouTube

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Sign Up to Generate

Open Problems

We're still in the process of identifying open problems mentioned in this paper. Please check back in a few minutes.

Collections

Sign up for free to add this paper to one or more collections.

Sign Up

Tweets

Sign up for free to view the 7 tweets with 153 likes about this paper.

Sign Up for Free