Training Classical Neural Networks by Quantum Machine Learning

Chen-Yu Liu,1, 2, ∗ En-Jui Kuo,1, 3, † Chu-Hsuan Abraham Lin,1, 4 Sean

Chen,1, 5, 6 Jason Gemsun Young,7 Yeong-Jar Chang,7 and Min-Hsiu Hsieh1, ‡
1
Hon Hai (Foxconn) Research Institute, Taipei, Taiwan
2
Graduate Institute of Applied Physics, National Taiwan University, Taipei, Taiwan
3
Physics Division, National Center for Theoretical Sciences, Taipei, Taiwan
4
Department of Electrical and Electronic Engineering, Imperial College London, London, UK
5
Department of Physics, University of California San Diego, California, US
6
Department of Mathematics, University of California San Diego, California, US
7
Industrial Technology Research Institute, Taipei, Taiwan
(Dated: February 27, 2024)
In recent years, advanced deep neural networks have required a large number of parameters for
arXiv:2402.16465v1 [quant-ph] 26 Feb 2024

training. Therefore, finding a method to reduce the number of parameters has become crucial for
achieving efficient training. This work proposes a training scheme for classical neural networks (NNs)
that utilizes the exponentially large Hilbert space of a quantum system. By mapping a classical
NN with M parameters to a quantum neural network (QNN) with O(polylog(M )) rotational gate
angles, we can significantly reduce the number of parameters. These gate angles can be updated
to train the classical NN. Unlike existing quantum machine learning (QML) methods, the results
obtained from quantum computers using our approach can be directly used on classical computers.
Numerical results on the MNIST and Iris datasets are presented to demonstrate the effectiveness
of our approach. Additionally, we investigate the effects of deeper QNNs and the number of mea-
surement shots for the QNN, followed by the theoretical perspective of the proposed method. This
work opens a new branch of QML and offers a practical tool that can greatly enhance the influence
of QML, as the trained QML results can benefit classical computing in our daily lives.

I. INTRODUCTION cant promise as a potent tool for analyzing complex data

and has the potential to bring about a revolution in var-
Quantum machine learning (QML) is an emergent field ious fields.
that utilizes the unique computational abilities of quan- In addition to issues related to learnability and train-
tum systems to revolutionize the way neural networks are ability, the practicality of QML models represents a sig-
trained and operated. Quantum operations could be used nificant concern. In instances of pure QML devoid of
to encode and process data into the quantum neural net- classical neural networks (NNs), gate angle encoding is
works (QNNs), which could theoretically evaluate numer- a common method for processing input data. However,
ous possibilities simultaneously using quantum superpo- it becomes evident that as the input data scales, the
sition and entanglement, thus accelerating the learning width and depth of the quantum circuit proportionally
process [1–4]. Aside from the general QNN methods, increase, hindering its accuracy in the noisy intermediate-
quantum kernel method utilizes quantum operations to scale quantum (NISQ) era. While classical preprocessing
generate kernels from the inner products of quantum could mitigate this by reducing data dimension, concerns
states, which are then used on classical data transposed arise regarding potential information loss in such cases.
into the quantum domain [5]. Moreover, Grover search To be considered practical, the dimension of data taken
could also be combined with QML for classification tasks, into the QML should be comparable to that of classical
thus obtaining potential advantages [6, 7]. QML has NNs. Another challenge posed against QML is that for
exhibited diverse practical applications with significant both pure and hybrid quantum-classical machine learning
impact, encompassing advancements in drug discovery (ML) models, the practical usage of the trained model at
[8], tasks related to large-scale stellar classification [9], the inference stage mandates access to a quantum com-
natural language processing [10], recommendation sys- puter. However, quantum computers, or even cloud ac-
tems [11–14], and generative learning [15–19]. Despite cess to quantum computing resources, are extremely lim-
its potential benefits, QML remains an emerging field, ited, introducing a significant requirement for the effec-
necessitating the resolution of several challenges before tive utilization of trained QML models.
achieving widespread adoption. Key challenges include
To alleviate the practicality concern, a viable approach
addressing issues related to the learnability [20–24] and
involves the utilization of quantum algorithms for train-
trainability [25–29] of QML models. QML holds signifi-
ing classical NNs. This strategy enables the utilization of
the trained classical NN model without encountering the
challenges associated with data encoding and relying on
∗ chen-yu.liu@foxconn.com quantum computers. Prior research aligning with this
† kuoenjui@umd.edu perspective employs quantum walk as a searching pro-
‡ min-hsiu.hsieh@foxconn.com cess for classical NN parameters [30]. Nevertheless, it is
 2

noteworthy that the Grover-like method in this context concerns about the associated computational require-
demonstrates an exponential scaling of the circuit depth. ments. Classically, the computational power needed
This work aims to present a novel perspective and in- scales linearly with the parameters, posing a potential
troduce an innovative QML framework. The proposed bottleneck as models become more sophisticated. Quan-
framework involves mapping the classical NN weights to tum computing, leveraging the exponential scaling of
the Hilbert space of a quantum state of QNN. This ap- Hilbert space with the number of qubits, emerges as a
proach facilitates the tuning of the parameterized quan- promising avenue to address or alleviate this impending
tum state (represented by QNN), thereby enabling the classical bottleneck. Motivated by this perspective, our
adjustment of the classical NN weights. Significantly, aim is to discover a quantum state approximation of a
this is achieved with a reduced parameter count of classical NN. In this context, the quantum state’s out-
O(polylog(M )) compared to the classical NN with M put encapsulates information about the weights, where
parameters. This novel approach addresses the data en- the quantum state’s parameters are exponentially fewer
coding issue of the QNN by training a classical NN with than those required to train a classical NN. To facilitate
classical input and output. Moreover, once the model tunability, we express the quantum state through a QNN.
is trained by the QNN, inference only requires classical By adjusting the QNN parameters, we concurrently tune
computers, significantly lowering the requirements for us- the classical NN parameters, effectively constructing the
ing QML results and enhancing the practicality of quan- classical NN as a function of the quantum state. With an
tum computers in everyday life. In Sec. II, we eluci- N -qubit quantum state generating 2N possible measure-
date the mapping between quantum state measurement ment outcomes, we map the probability values of diverse
outcomes and classical NN weights. Subsequently, in outcomes to classical NN weights through a scheme de-
Sec. III, we delve into the numerical results and engage scribed later.
in a comprehensive discussion. Finally, in Sec. IV we In this section, we initiate our discussion by elucidat-
sum up our results and outline the prospective avenues ing the mapping from QNN measurement outcomes to
for future research. classical NN weights, with 3 settings. Then the train-
ing scheme of updating the QNN corresponding to the
classical NN will be presented.
A. Main Results

Our main results of training classical NN weights by A. Equivalent setting

QML can be succinctly summarized as follows:
Consider a classical NN with M weights
• Shallow circuit depth of QNN: In contrast to con-
ventional QML approaches which often incorpo- θ⃗ = (θ1 , θ2 , ..., θM ), (1)
rate data loading layers imposing constraints on
input data size or requiring data compression, our we specify a quantum state |ψ⟩ with N = ⌈log2 M ⌉
methodology leverages the classical data input and qubits, this guarantees the Hilbert space size 2N ≥ M .
output process. This provides a natural advantage, The measurement probabilities of such quantum state
enabling a shallower quantum circuit for QNNs. in computational basis are |⟨i|ψ⟩|2 ∈ [0, 1], where i ∈
{1, 2, ..., 2N }. To relate 2N measurement probability re-
• Parameter reduction: In our approach only sults to M classical NN weights, the following setting is
O(polylog(M )) parameters is required to train a applied:
classical NNs with M parameters. This reduction
1. 2N − M weights are picked randomly, with each of
is achieved by utilizing N = ⌈log2 M ⌉ qubits and a
them be related to the average measurement prob-
polynomial number of QNN layers.
ability of 2 different bases, such that
• Inference without quantum computers: The 1
trained model exhibits compatibility with classical θj = (|⟨i|ψ⟩|2 + |⟨k|ψ⟩|2 ), i ̸= k. (2)
2
hardware. In fact, the model inference relies only
on classical computers, contributing to heightened 2. After step 1, each of the remaining 2M −2N weights
practicality, particularly in light of the restricted is related to the measurement probability of a single
availability of quantum computers in comparison basis
to classical computers.
θj = |⟨i|ψ⟩|2 . (3)

II. CLASSICAL NEURAL NETWORKS AS A Note that even some classical NN weights are related to
FUNCTION OF QUANTUM STATES 2 different bases, each basis is only related to 1 weight.
With the above setting, one can observed that exactly
The proliferation of classical NNs equipped with mil- M weights are mapped from measurement probabilities
lions or even billions of trainable parameters has raised of 2N bases.
 3

B. Positive and negative setting parameters that is polynomial in the number of qubits.
Associate with the previous qubit number setting, the
The measurement probability of |ψ⟩ in the computa- required number of parameters is O(polylog(M )). In
tional basis is always positive. However, a weight of the Fig. 1(a), an example of N = 13 qubits EfficientSU2 with
classical NN could be positive or negative. Thus, the fol- 13 layers is shown, where the circuit is initialized with
lowing setting is applied to bring the negative numbers Hadamard gates and consists of layers of single qubit op-
from the measurement probabilities: erations spanned by SU(2) and CNOT entanglements,
the random rotational angles in the example are used for
θj → θj for even j, (4) illustration purpose. In the end of the circuit, all of the
θj → −θj for odd j. (5) qubits are measured, generating 2N possible outcomes.
Following the settings in previous sections with QNN
Such that almost half of the weights (depending on the ⃗ the resulting mappings from measurement results
|ψ(ϕ)⟩,
parity of M ) become negative and others remain positive. of the QNN to the classical NN weights θ⃗ are
h  i
⃗ 2 + |⟨k|ψ(ϕ)⟩|
θj ′ = γ tanh ±2N −2 γ |⟨i|ψ(ϕ)⟩| ⃗ 2 , (8)
C. Scaling setting
where j ′ ∈ {1, 2, ..., 2N − M }, for the randomly picked
After the above settings, the range of the classical NN indices to be related to 2 different bases. The re-indexing
weight is now become θj ∈ [−1, 1]. However, as the j → j ′ is for convenience to represent the amount of
scale of the system in consideration increases (both M indices j, since these indices are picked randomly. The
as well as N ), considering the normalization condition of symbol ± is related to the parity of the original index j.
the quantum state And
N
 
2
X ⃗ 2 ,
θj ′ = γ tanh ±2N −1 γ|⟨i|ψ(ϕ)⟩| (9)
|⟨i|ψ⟩|2 = 1, (6)
i=1
where j ′ ∈ {2N −M +1, ..., M }, for the remaining weights
the average probability of basis |i⟩ being measured, 1/2N , that are related to the single basis.
decreases exponentially as N increases. Therefore, the
average magnitude of θj is also decreasing exponentially
since |θj | ∝ |⟨i|ψ⟩|2 . As a result, the average magnitude E. Training Classical NN by tuning QNN
should be preserved under different system sizes, when
designing a mapping between the quantum states and For now, given a classical NN with weights θ⃗ ∈ RM for
the classical NN weights. Thus, the scaling setting is some task T , with the mappings in Eq. (8) and Eq. (9),
designed as the following: ⃗ of the correspond-
it is possible to tune the parameters ϕ
ing QNN to improve the performance of classical NN
θj → γ tanh(γ2N −1 θj ). (7) on the task T . To achieve this, the loss function and
the optimization algorithm (optimizer) to tune the QNN
With scaling factor γ ≥ 0, the 2N dependency of the
parameters are required to be specified. In this work,
scaling setting preserves the resulting average magnitude
we consider the classical NN for the classification tasks,
of the classical NN weights under different system sizes.
where the data are inputted and outputted classically.
The usage of the function tanh is inspired by the tanh
The loss function to tune the QNN in order to train the
activation function in machine learning, where the output
corresponding classical NN is designed as
range is (−1, 1).
Nfail
Loss = CE + , (10)
Nd
D. Parameterized Quantum States as Quantum
Nd
Neural Networks 1 X
CE = − [yn log ŷn + (1 − yn ) log(1 − ŷn )](11)
,
Nd n=1
In previous sections, we specified the qubit size of the
quantum state |ψ⟩, and the corresponding relation of the where CE is the well-known Cross-entropy [32] with yn
measurement probabilities to the classical NN weights. as the true label and ŷn as the predicted label by the
In this section, we further describe the detail of con- classical NN model for the sample n. Nfail is the num-
structing |ψ⟩, starting from its parametrization by in- ber of failed predictions by the model, and Nd is the
troducing the rotational gate angle dependency |ψ(ϕ)⟩. ⃗ total number of training data. Thus, the second term
This parameterized quantum state could be used as the represents the rate of failed predictions. A training ses-
so-called QNN, with a specific ansatz that constructs the sion in our scheme consists of Ntrain training periods. In
quantum circuit. The EfficientSU2 circuit ansatz is used each training period, Nϕ iterations are used to tune the
in this work [31], where we can specify the number of parameters ϕ ⃗ using the COBYLA algorithm. Following
 4

that, Nγ iterations are employed to adjust the scaling Fig. 2(b), the loss, training accuracy, and test accuracy
factor γ using the Nelder-Mead algorithm, as illustrated throughout the training process are presented for QNN
in Fig. 1(b). The training scheme described above is sum- layers = 1. In this scenario, the number of QNN pa-
marized in Fig. 1(c). In contrast to ordinary QML, where rameters is 32, with a required parameter ratio of 24.4%
QNNs typically involve data loading layers that constrain compared to the original M = 131.
input data size or require data compression, potentially
leading to information loss, our approach, as shown in
Fig. 1(c), leverages the classical data input process, pro- C. Effects of QNN Layers and Measurement Shots
viding a natural advantage of a shallower quantum circuit
for QNN. With an increase in the number of QNN layers, im-
proved expressibility is expected due to the rising num-
ber of parameters. In Fig. 2(c), training results for
III. NUMERICAL RESULTS AND DISCUSSION the MNIST dataset with different QNN layers LN ∈
{1, 4, 7, 10, 13, 26} are presented. Other settings, aside
In this section, after describing the setup of numerical from QNN layers, remain consistent with the descrip-
experiments, we will present the training results on two tion in Sec. III A. Similarly, in Fig. 2(d), training re-
different datasets. Further investigations into the effects sults for the Iris dataset with different QNN layers LN ∈
of the number of QNN layers and that of measurement {1, 2, 3, 4, 5, 6} are also presented, with other settings fol-
shots are also shown. lowing those in Sec. III B. It can be observed that the
performance is indeed improved in both these figures. In
the MNIST case, we can obtain a testing accuracy of
A. Setup for MNIST dataset 74.6% with just a 5.38% required parameters ratio and a
testing accuracy of 80.21% with a 10.8% required param-
In our numerical experiment conducted on the IBM eters ratio. In the Iris dataset case, since it is a relatively
Quantum simulator [33], we initially utilize the MNIST simple problem compared to MNIST, with only 1 layer of
dataset [34] and a convolutional neural network (CNN) QNN and a 24.4% required parameters ratio, the testing
with M = 6690 parameters as our classical NN. Conse- accuracy is already above 95%. The performance slightly
quently, the required number of qubits is ⌈log2 6690⌉ = improves when the QNN layers increase to nearly 98%
13. The training dataset comprises Nd = 60000 sam- when QNN layers = 6.
ples, and the testing dataset consists of Ntest = 10000 In practice, the number of measurement shots is lim-
samples. Within each training period, Nϕ = 2000 and ited when utilizing real quantum devices. To explore
Nγ = 200. A total of Ntrain = 100 training periods are the impact of different measurement shot scenarios, we
conducted. The initial gate angle parameters ϕ ⃗ init are conducted a comprehensive investigation. The outcomes
randomly set, while the initial value of γ is established as for the MNIST dataset are illustrated in Fig. 3(a), while
γinit = 0.3. For now, the state vector simulation method those for the Iris dataset are depicted in Fig. 3(b). No-
is used, which means the number of measurement shots tably, the chosen numbers of measurement shots align
= ∞. The results for other numbers of measurement with multiples of the corresponding Hilbert space size,
shots will be discussed later. In Fig. 2(a), the loss, train- providing a convenient framework for observing accuracy
ing accuracy, and test accuracy throughout the training trends concerning the Hilbert space size, denoted as 2N .
process are depicted for QNN layers = 26. In this sce- Although the precise relationship between accuracy and
nario, the number of QNN parameters is 728, with a re- the requisite measurement shots remains unknown, refer-
quired parameter ratio of 10.8% compared to the original encing theoretical results in tomography tasks can offer
M = 6690. insights into potential scaling behaviors. For instance, in
fidelity tomography—aiming to satisfy F (|ψ⟩ , |ϕ⟩) ≥ 1−ϵ
for the output state |ϕ⟩ and unknown state |ψ⟩ with infi-
B. Setup for Iris dataset delity ϵ—prior studies indicate that a sufficient number of
N N
measurement shots is Nshot = O( 2ϵ log( 2ϵ )) [36]. As we
Secondly, we also conducted on the IBM Quantum sim- contemplate investigations in larger system sizes N , we
ulator using the state vector method with a number of propose that increasing measurement shots linearly with
measurement shots = ∞ on the Iris dataset [35], and a 2N is a prudent strategy for enhancing both training and
single hidden layer neural network with M = 131 pa- testing accuracy.
rameters is utilized. In this case, the required number of
qubits is ⌈log2 131⌉ = 8. The number of training data
Nd = 100, and the number of testing data Ntest = 50. D. Existence and Necessity of QNN
In each training period, Nϕ = 100 and Nγ = 10. A to- Approximation for classical NN
tal of Ntrain = 21 training periods are conducted. The
initial gate angle parameters ϕ ⃗ init are also set randomly, After presenting the numerical results, we seek to pro-
while the initial value of γ is established as γinit = 0.3. In vide the theoretical rationale for the existence of QNN
 5

FIG. 1. Elements and the training scheme of the proposed approach. (a) Circuit ansatz EfficientSU2, used as QNN in this
work. (b) A training period that consists of two stages, tuning the QNN parameters ϕ ⃗ by COBYLA algorithm, and tuning the
scaling factor γ by Nelder-Mead algorithm. Corresponding schematic histogram of the classical NN weights θ⃗ is shown, where
⃗
horizontal axis is the value of θ, and vertical axis is the appearing frequency of corresponding value in θ.

approximations for classical NNs. When examining a one dimension or geometrical local Hamiltonian.
classical NN, if the collective weights of the network give
rise to a wavefunction through a mapping operation, and
this wavefunction belongs to the complexity class Samp- IV. CONCLUSION AND FUTURE WORK
BQP [37–39], it logically follows that a corresponding
quantum state can be generated by employing quantum
By using the mapping from the 2⌈log2 M ⌉ measurement
circuits characterized by polynomial depth. This estab-
probabilities of QNN to the M weights of classical NN,
lishes a sufficient condition for the presence of such QNN
the proposed approach can significantly reduce the num-
structures.
ber of parameters to be trained. We demonstrated the
Crucially, this condition offers compelling evidence results on the MNIST dataset and Iris dataset as the
that the quantum approximation in question cannot be proof of concept, representing the practicality of the pro-
replicated by a classical probabilistic Turing machine posed method.
based on the quantum supremacy argument. As widely In this study, we use the EfficientSU2 QNN ansatz,
acknowledged, if SampBQP were equal to SampBPP (the which incorporates Ry and Rz rotational gates along with
complexity class defined by probability distributions that linearly distributed CNOT gates. While there’s potential
can be generated by a probabilistic Turing machine in for other QNN ansatzes to outperform EfficientSU2, par-
polynomial time), it would result in the Polynomial Hi- ticularly by considering qubit connectivity on real quan-
erarchy collapse [38–40]. Thus, this robustly supports the tum hardware and devising circuits accordingly, we leave
contention that SampBQP ̸= SampBPP. Consequently, this exploration for future work. The aim for the fu-
we assert that, barring a collapse of the Polynomial Hi- ture work is to delve into different ansatz options and
erarchy, the incorporation of a QNN is indispensable for assess their efficiency and effectiveness, potentially un-
compressing classical NNs. It’s worth noting that some covering more suitable alternatives through QNN archi-
tensor network [41, 42] techniques, such as the matrix tecture search algorithms. Here, We examine the effect of
product state, can be viewed as an efficient simulation number of QNN layers on the accuracy of the classifica-
of quantum states. However, these results are limited to tion tasks in both dataset. As anticipated, our findings
 6

FIG. 2. (a) Training results for QNN layers = 26 with the MNIST dataset. (b) Training results for QNN layers = 1 with the
Iris dataset. (c) Training results for different QNN layers with the MNIST dataset, along with the required parameters ratio
for each investigated QNN layer. (d) Similar to (c) but with the Iris dataset.

FIG. 3. Effects of different measurement shots. (a) MNIST dataset, where the investigated measurement shots are multiples
of the corresponding Hilbert space size 213 = 8192. (b) Iris dataset, where the investigated measurement shots are multiples of
the corresponding Hilbert space size 28 = 256.

reveal that deeper QNNs exhibit superior performance retical studies have hinted at analogous behavior in the
owing to the heightened expressibility inherent in deeper context of required measurement shots for quantum to-
circuits. It is pertinent to note that the current analy- mography tasks.
sis does not incorporate noise considerations. However,
in the future investigations, when the expressibility is Currently, we have adopted a specific mapping ap-
affected by circuit noise, it is foreseeable that the perfor- proach; however, it is worth noting that alternative map-
mance of deeper QNNs might diminish. pings could potentially enhance performance or stream-
line the training process. Investigating such possibilities
Furthermore, beyond employing the precise state vec- is also designated as a future research avenue, with a
tor simulation for the quantum circuit, we explored shot- focus on optimizing and exploring mappings that yield
based simulations involving a limited number of measure- superior results in translating measurement probabilities
ment shots. Our observations reveal that the attainable to classical NN weights. Additionally, the optimization
accuracy in classification tasks improves as the number and learning processes employed in this study rely on
of measurement shots increases, with a particular em- non-gradient-based optimizers. A promising direction for
phasis on multiples of the Hilbert space size 2N . This future exploration involves studying more efficient op-
outcome aligns with expectations, as certain prior theo- timization algorithms for our training task, potentially
 7

incorporating gradient calculations. Lastly, the break- ACKNOWLEDGMENTS

through potential of parameter reduction on a polylog-
arithmic scale could revolutionize the training of large
classical ML models. If quantum computing techniques EJK thanks the National Center for Theoretical Sci-
can render the training of such models more efficient ences of Taiwan for funding (112-2124-M-002-003). JGY
while allowing for inference on classical computers in our and YJC acknowledge the funding provided under the
daily applications, the significance of this work would be ITRI EOSL 2022 project“Quantum Circuit Design and
greatly enhanced. Application”.

[1] J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, Feature selection for recommender systems with quan-
N. Wiebe, and S. Lloyd, Quantum machine learning, Na- tum computing, Entropy 23, 10.3390/e23080970 (2021).
ture 549, 195 (2017). [15] P.-L. Dallaire-Demers and N. Killoran, Quantum gen-
[2] M. Schuld, I. Sinayskiy, and F. Petruccione, An intro- erative adversarial networks, Phys. Rev. A 98, 012324
duction to quantum machine learning, Contemporary (2018).
Physics 56, 172 (2015). [16] S. Lloyd and C. Weedbrook, Quantum generative adver-
[3] E. Farhi and H. Neven, Classification with quan- sarial learning, Phys. Rev. Lett. 121, 040502 (2018).
tum neural networks on near term processors (2018), [17] J. Tian, X. Sun, Y. Du, S. Zhao, Q. Liu, K. Zhang,
arXiv:1802.06002 [quant-ph]. W. Yi, W. Huang, C. Wang, X. Wu, M.-H. Hsieh,
[4] C. Ciliberto, M. Herbster, A. D. Ialongo, M. Pontil, T. Liu, W. Yang, and D. Tao, Recent advances for
A. Rocchetto, S. Severini, and L. Wossnig, Quantum ma- quantum neural networks in generative learning (2022),
chine learning: a classical perspective, Proceedings of the arXiv:2206.03066.
Royal Society A: Mathematical, Physical and Engineer- [18] Y. Du, M.-H. Hsieh, and D. Tao, Efficient online quantum
ing Sciences 474, 20170551 (2018). generative adversarial learning algorithms with applica-
[5] M. Schuld and N. Killoran, Quantum machine learning tions (2019), arXiv:1904.09602.
in feature hilbert spaces, Physical Review Letters 122, [19] H.-L. Huang, Y. Du, M. Gong, Y. Zhao, Y. Wu, C. Wang,
10.1103/physrevlett.122.040504 (2019). S. Li, F. Liang, J. Lin, Y. Xu, R. Yang, T. Liu, M.-H.
[6] L. K. Grover, A fast quantum mechanical algorithm for Hsieh, H. Deng, H. Rong, C.-Z. Peng, C.-Y. Lu, Y.-A.
database search, in Proceedings of the Twenty-Eighth An- Chen, D. Tao, X. Zhu, and J.-W. Pan, Experimental
nual ACM Symposium on Theory of Computing, STOC quantum generative adversarial networks for image gen-
’96 (Association for Computing Machinery, New York, eration, Phys. Rev. Applied 16, 024051 (2021).
NY, USA, 1996) p. 212–219. [20] Y. Du, M.-H. Hsieh, T. Liu, S. You, and D. Tao, Learn-
[7] Y. Du, M.-H. Hsieh, T. Liu, and D. Tao, A grover-search ability of quantum neural networks, PRX Quantum 2,
based quantum learning scheme for classification, New 040337 (2021).
Journal of Physics 23, 023020 (2021). [21] M. Soltanolkotabi, A. Javanmard, and J. D. Lee, Theo-
[8] Y. Cao, J. Romero, J. P. Olson, M. Degroote, P. D. John- retical insights into the optimization landscape of over-
son, M. Kieferová, I. D. Kivlichan, T. Menke, B. Per- parameterized shallow neural networks, IEEE Transac-
opadre, N. P. D. Sawaya, S. Sim, L. Veis, and A. Aspuru- tions on Information Theory 65, 742 (2019).
Guzik, Quantum chemistry in the age of quantum com- [22] Y. Du, M.-H. Hsieh, T. Liu, and D. Tao, A grover-search
puting, Chemical Reviews 119, 10856 (2019). based quantum learning scheme for classification, New
[9] K.-C. Chen, X. Xu, H. Makhanov, H.-H. Chung, and Journal of Physics 23, 023020 (2021).
C.-Y. Liu, Quantum-enhanced support vector machine [23] Y. Du, M.-H. Hsieh, T. Liu, and D. Tao, Expressive
for large-scale stellar classification with gpu acceleration power of parametrized quantum circuits, Phys. Rev. Re-
(2023), arXiv:2311.12328 [quant-ph]. search 2, 033125 (2020).
[10] K. Meichanetzidis, S. Gogioso, G. de Felice, N. Chiap- [24] C.-Y. Liu, C.-H. A. Lin, and K.-C. Chen, Learning quan-
pori, A. Toumi, and B. Coecke, Quantum natural lan- tum phase estimation by variational quantum circuits
guage processing on near-term quantum computers, Elec- (2023), arXiv:2311.04690 [quant-ph].
tronic Proceedings in Theoretical Computer Science 340, [25] K. Zhang, M.-H. Hsieh, L. Liu, and D. Tao, To-
213 (2021). ward trainability of quantum neural networks (2020),
[11] C.-Y. Liu, H.-Y. Wang, P.-Y. Liao, C.-J. Lai, and M.-H. arXiv:2011.06258.
Hsieh, Implementation of trained factorization machine [26] K. Zhang, M.-H. Hsieh, L. Liu, and D. Tao, Gaussian ini-
recommendation system on quantum annealer, arXiv tializations help deep variational quantum circuits escape
preprint arXiv:2210.12953 (2022). from the barren plateau (2022), arXiv:2203.09376.
[12] I. Kerenidis and A. Prakash, Quantum recommendation [27] K. Zhang, M.-H. Hsieh, L. Liu, and D. Tao, Quantum
systems (2016), arXiv:1603.08675. gram-schmidt processes and their application to efficient
[13] E. Tang, A quantum-inspired classical algorithm for rec- state readout for quantum algorithms, Phys. Rev. Re-
ommendation systems, in Proceedings of the 51st An- search 3, 043095 (2021).
nual ACM SIGACT Symposium on Theory of Comput- [28] Y. Du, M.-H. Hsieh, T. Liu, S. You, and D. Tao,
ing, STOC 2019 (Association for Computing Machinery, Quantum differentially private sparse regression learn-
New York, NY, USA, 2019) p. 217–228. ing, IEEE Transactions on Information Theory 68, 5217
[14] R. Nembrini, M. Ferrari Dacrema, and P. Cremonesi, (2022).
 8

[29] Y. Du, M.-H. Hsieh, T. Liu, D. Tao, and N. Liu, Quantum 1809.1936.tb02137.x.
noise protects quantum classifiers against adversaries, [36] J. Haah, A. W. Harrow, Z. Ji, X. Wu, and N. Yu, Sample-
Phys. Rev. Research 3, 023153 (2021). optimal tomography of quantum states, in Proceedings
[30] L. S. de Souza, J. H. A. de Carvalho, and T. A. E. Fer- of the forty-eighth annual ACM symposium on Theory of
reira, Classical artificial neural network training using Computing (2016) pp. 913–925.
quantum walks as a search procedure, IEEE Transactions [37] S. Aaronson and L. Chen, Complexity-theoretic founda-
on Computers 71, 378 (2022). tions of quantum supremacy experiments, arXiv preprint
[31] Qiskit circuit library, https://qiskit.org/ arXiv:1612.05903 (2016).
documentation/stubs/qiskit.circuit.library. [38] A. P. Lund, M. J. Bremner, and T. C. Ralph, Quan-
EfficientSU2.html (2023). tum sampling problems, bosonsampling and quantum
[32] I. J. Good, Rational decisions, Jour- supremacy, npj Quantum Information 3, 15 (2017).
nal of the Royal Statistical Society: Se- [39] A. Bouland, B. Fefferman, C. Nirkhe, and U. Vazirani,
ries B (Methodological) 14, 107 (1952), On the complexity and verification of quantum random
https://rss.onlinelibrary.wiley.com/doi/pdf/10.1111/j.2517- circuit sampling, Nature Physics 15, 159 (2019).
6161.1952.tb00104.x. [40] S. Aaronson and A. Arkhipov, The computational com-
[33] IBM Quantum, https://quantum-computing.ibm.com plexity of linear optics, in Proceedings of the forty-third
(n.d.). annual ACM symposium on Theory of computing (2011)
[34] L. Deng, The mnist database of handwritten digit images pp. 333–342.
for machine learning research, IEEE Signal Processing [41] S. Montangero, E. Montangero, and Evenson, Introduc-
Magazine 29, 141 (2012). tion to tensor network methods (Springer, 2018).
[35] R. A. FISHER, The use of multiple measurements in [42] G. Evenbly and G. Vidal, Tensor network states and ge-
taxonomic problems, Annals of Eugenics 7, 179 (1936), ometry, Journal of Statistical Physics 145, 891 (2011).
https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1469-