Review doi:10.

1038/nature23474

Quantum machine learning

Jacob Biamonte1,2, Peter Wittek3, Nicola Pancotti4, Patrick Rebentrost5, Nathan Wiebe6 & Seth Lloyd7

Fuelled by increasing computer power and algorithmic advances, machine learning techniques have become powerful
tools for finding patterns in data. Quantum systems produce atypical patterns that classical systems are thought not
to produce efficiently, so it is reasonable to postulate that quantum computers may outperform classical computers on
machine learning tasks. The field of quantum machine learning explores how to devise and implement quantum software
that could enable machine learning that is faster than that of classical computers. Recent work has produced quantum
algorithms that could act as the building blocks of machine learning programs, but the hardware and software challenges
are still considerable.

L
ong before we possessed computers, human beings strove to find a set of instructions solving a problem, such as determining whether two
patterns in data. Ptolemy fitted observations of the motions of the graphs are isomorphic, that can be performed on a quantum computer.
stars to a geocentric model of the cosmos, with complex epicycles Quantum machine learning software makes use of quantum algorithms
to explain the retrograde motions of the planets. In the sixteenth century, as part of a larger implementation. By analysing the steps that quantum
Kepler analysed the data of Copernicus and Brahe to reveal a previously algorithms prescribe, it becomes clear that they have the potential to out-
hidden pattern: planets move in ellipses with the Sun at one focus of the perform classical algorithms for specific problems (that is, reduce the
ellipse. The analysis of astronomical data to reveal such patterns gave rise number of steps required). This potential is known as quantum speedup.
to mathematical techniques such as methods for solving linear ­equations The notion of a quantum speedup depends on whether one takes a
(Newton–Gauss), learning optima via gradient descent (Newton), formal computer science perspective—which demands mathematical
­polynomial interpolation (Lagrange), and least-squares fitting (Laplace). The proofs—or a perspective based on what can be done with realistic, finite-
nineteenth and early twentieth centuries gave rise to a broad range of mathe­ size devices—which requires solid statistical evidence of a scaling advan-
matical methods for analysing data to reveal the patterns that it contained. tage over some finite range of problem sizes. For the case of quantum
The construction of digital computers in the mid-twentieth ­century machine learning, the best possible performance of classical algorithms
allowed the automation of data analysis techniques. Over the past is not always known. This is similar to the case of Shor’s polynomial-time
half-century, the rapid progression of computer power has allowed quantum algorithm for integer factorization: no sub-exponential-time
the implementation of linear algebraic data analysis techniques such classical algorithm has been found, but the possibility is not provably
as r­ egression and principal component analysis, and has led to more ruled out.
­complex learning methods such as support vector machines. Over the Determination of a scaling advantage contrasting quantum and
same time frame, the development and rapid advance of digital c­ omputers ­classical machine learning would rely on the existence of a quantum
spawned novel machine learning methods. Artificial neural networks ­computer and is called a ‘benchmarking’ problem. Such advantages
such as p ­ erceptrons were implemented in the 1950s (ref. 1), as soon as could include improved classification accuracy and sampling of classically
computers had the power to realize them. Deep learning built on ­neural ­inaccessible systems. Accordingly, quantum speedups in machine learning
networks (such as Hopfield networks and Boltzmann machines) and are c­ urrently characterized using idealized measures from complexity
training methods (such as back propagation) were introduced and imple- ­theory: query complexity and gate complexity (see Box 1 and Box 1 Table).
mented in the 1960s to 1990s (ref. 2). In the past decade, p ­ articularly Query complexity measures the number of queries to the information
in the past five years, the combination of powerful computers and source for the classical or quantum algorithm. A quantum speedup
­special-purpose information processors capable of implementing deep results if the number of queries needed to solve a problem is lower for the
networks with billions of weights3, together with their application to very ­quantum algorithm than for the classical algorithm. To determine the gate
large datasets, has revealed that such deep learning networks are capable complexity, the number of elementary quantum operations (or gates)
of identifying complex and subtle patterns in data. required to obtain the desired result are counted.
Quantum mechanics is well known to produce atypical p ­ atterns in Query and gate complexity are idealized models that quantify the
data. Classical machine learning methods such as deep neural networks necessary resources to solve a problem class. Without knowing how to
frequently have the feature that they can both recognize statistical pat- map this idealization to reality, not much can be said about the n ­ ecessary
terns in data and produce data that possess the same statistical patterns: resource scaling in a real-world scenario. Therefore, the required
they recognize the patterns that they produce. This observation suggests resources of classical machine learning algorithms are mostly quantified
the following hope. If small quantum information processors can pro- by numerical experimentation. The resource requirements of quantum
duce ­statistical patterns that are computationally difficult for a classical machine learning algorithms are likely to be similarly difficult to quantify
­computer to produce, then perhaps they can also recognize patterns that in practice. The analysis of their practical feasibility is a central subject
are equally difficult to recognize classically. of this review.
The realization of this hope depends on whether efficient quantum As will be seen throughout the review, there are quantum algorithms
algorithms can be found for machine learning. A quantum algorithm is for machine learning that exhibit quantum speedups4–7. For example, the

1
Quantum Complexity Science Initiative, Skolkovo Institute of Science and Technology, Skoltech Building 3, Moscow 143026, Russia. 2Institute for Quantum Computing, University of Waterloo,
Waterloo, N2L 3G1 Ontario, Canada. 3ICFO—The Institute of Photonic Sciences, Castelldefels, Barcelona 08860 Spain. 4Max Planck Institute of Quantum Optics, 1 Hans-Kopfermannstrasse,
D-85748 Garching, Germany. 5Massachusetts Institute of Technology, Research Laboratory of Electronics, Cambridge, Massachusetts 02139, USA. 6Station Q Quantum Architectures and
Computation Group, Microsoft Research, Redmond, Washington 98052, USA. 7Massachusetts Institute of Technology, Department of Mechanical Engineering, Cambridge, Massachusetts 02139, USA.

1 4 s e p t e m b e r 2 0 1 7 | V O L 5 4 9 | N A T U RE | 1 9 5
© 2017 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.
 INSIGHT Review

as quantum annealers and programmable quantum optical arrays are well

Box 1 matched to deep learning architectures20–22. Although it is not clear yet to
what extent this potential can be realized, there are reasons to be optimistic
Quantum speedups that quantum computers can recognize patterns in data that classical
Quantum computers use effects such as quantum coherence computers cannot.
and entanglement to process information in ways that classical The learning machines we consider can be either classical23–32 or quan-
computers cannot. The past two decades have seen steady tum8,9,11,13,33–36. The data they analyse can be either classical or ­quantum
advances in constructing more powerful quantum computers. states produced by quantum sensing or measuring apparatus30,37. We
A quantum algorithm is a stepwise procedure performed on briefly discuss conventional machine learning—the use of classical
a quantum computer to solve a problem, such as searching a computers to find patterns in classical data. We then turn to quantum
database. Quantum machine learning software makes use of machine learning, where the data that the quantum computer analyses
quantum algorithms to process information. Quantum algorithms can be either classical data, which ends up encoded as quantum states, or
can in principle outperform the best known classical algorithms quantum data. Finally, we discuss briefly the problem of using classical
when solving certain problems. This is known as a quantum machine learning techniques to find patterns in quantum dynamics.
speedup105.
For example, quantum computers can search an unsorted Classical machine learning
database with N entries in time proportional to √​N—that is, Classical machine learning and data analysis can be divided into several
O(√​N)—where a classical computer given blackbox access to the categories. First, computers can be used to perform ‘classic’ data a­ nalysis
same database takes time proportional to N: thus the quantum methods such as least-squares regression, polynomical interpolation
computer exhibits a square-root speedup over the classical and data analysis. Machine learning protocols can be supervised or
computer. Similarly, quantum computers can perform Fourier ­unsupervised. In supervised learning, the training data are divided into
transforms over N data points, invert sparse N ×​  N matrices, and labelled categories, such as samples of handwritten digits together with
find their eigenvalues and eigenvectors in time proportional to a the actual number the handwritten digit is supposed to represent, and
polynomial in log2N, where the best known algorithms for classical the job of the machine is to learn how to assign labels to data outside the
computers take time proportional to Nlog2N: thus the quantum training set. In unsupervised learning, the training set is unlabelled, and
computer exhibits an exponential speedup over the best classical the goal of the machine is to find the natural categories into which the
computer algorithms. training data falls (for example, different types of photos on the internet)
In the Box 1 Table, speedups are taken with respect to their and then to categorize data outside the training set. Finally, there are
classical counterpart(s)—hence, O(√​N) means quadratic speedup machine learning tasks, such as playing Go, that involve combinations of
and O(log(N)) means exponential relative to their classical supervised and unsupervised learning, together with training sets that
counterpart. may be generated by the machine itself.

Box 1 Table | Speedup techniques for given quantum machine Linear-algebra-based quantum machine learning
learning subroutines A wide variety of data analysis and machine learning protocols operate
by performing matrix operations on vectors in a high-dimensional vector
Method Speedup Amplitude HHL Adiabatic qRAM
amplification space. But quantum mechanics is all about matrix operations on vectors
Bayesian O(√​N) Yes Yes No No
in high-dimensional vector spaces.
inference106,107 The key ingredient behind these methods is that the quantum state
Online O(√​N) Yes No No Optional of n quantum bits or qubits is a vector in a 2n-dimensional complex
perceptron108 ­vector space; performing a quantum logic operations or a measure-
Least-squares O(logN)*​ Yes Yes No Yes ment on qubits multiplies the corresponding state vector by 2n ×​  2n
fitting9
Classical O(√​N) Yes/No Optional/ No/Yes Optional
­matrices. By building up such matrix transformations, quantum com-
Boltzmann No puters have been shown to perform common linear algebraic operations
machine20 such as Fourier transforms38, finding eigenvectors and eigenvalues39,
Quantum O(logN)*​ Optional/No No No/Yes No and s­ olving linear sets of equations over 2n-dimensional vector spaces
Boltzmann in time that is ­polynomial in n, exponentially faster than their best
machine22,61
Quantum O(logN)*​ No Yes No Optional
known c­ lassical counterparts8. This latter is commonly referred to as
PCA11 the Harrow, Hassidim and Lloyd (HHL) algorithm8 (see Box 2). The
Quantum O(logN)*​ No Yes No Yes ­original variant assumed a well conditioned matrix that is sparse.
support vector Sparsity is unlikely in data science, but later improvements relaxed this
machine13
assumption to include low-rank matrices as well10,33,40. Going past HHL,
Quantum O(√​N) Yes No No No
reinforcement here we survey several quantum algorithms which appear as subroutines
learning30 when linear algebra techniques are employed in quantum machine
*​There exist important caveats that can limit the applicability of the learning software.
method51.
Quantum principal component analysis
For example, consider principal component analysis (PCA). Suppose that
quantum basic linear algebra subroutines (BLAS)—Fourier ­transforms, the data are presented in the form of vectors vj in a d-dimensional vector
finding eigenvectors and eigenvalues, solving linear equations— space, where d = 2n = N. For example, vj could be the vector of changes
exhibit exponential quantum speedups over their best known classical in prices of all stocks in the stock market from time tj to time tj + 1.
­counterparts8–10. This quantum BLAS (qBLAS) translates into quantum The covariance matrix of the data is C = ∑ j vjvjT , where superscript
speedups for a variety of data analysis and machine learning algorithms T denotes the transpose ­operation: the covariance matrix summarizes
including linear algebra, least-squares fitting, gradient descent, Newton’s the correlations between the different components of the data, for exam-
method, principal component analysis, linear, semidefinite and quadratic ple, correlations between changes in the prices of different stocks. In its
programming, topological analysis and support vector machines9,11–19. simplest form, principal component analysis operates by diagonalizing
At the same time, special-purpose quantum information processors such the covariance matrix: C = ∑kek ck c†k , where the ck are the eigenvectors

1 9 6 | N A T U RE | V O L 5 4 9 | 1 4 s e p t e m b e r 2 0 1 7
© 2017 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.
 Review INSIGHT

of C, and ek are the c­ orresponding eigenvalues. (Because C is symmetric,

the eigenvectors ck form an orthonormal set.) If only a few of the eigen- Box 2
values ck are large, and the remainder are small or zero, then the eigen-
vectors corresponding to those eigenvalues are called the principal HHL algorithm
components of C. Each ­principal component represents an underlying The HHL algorithm for inverting systems of equations is a
common trend or form of correlation in the data, and decomposing a fundamental and easy-to-understand subroutine, underpinning
data vector v in terms of principal components, v =​  ∑k vk ck, allows one many quantum machine learning algorithms. The algorithm seeks
both to compress the representation of the data and to predict future to solve Ax =​  b using a quantum computer. HHL quantizes the
behaviour. Classical algorithms for p ­ erforming PCA scale as O(d2) in problem by expressing the vector b ∈ C N as a quantum state b
terms of computational complexity and query complexity. (We note that over log2N qubits, and the vector x as a quantum state x . The
we make use of ‘big O’ notation to keep track of the leading term that matrix A can be assumed to be Hermitian without loss of
dominates scaling.) generality because the space can always be expanded to make this
For quantum principal component analysis of classical data11, we true. The equation A x = b can then be solved by multiplying
choose a data vector vj at random, and use a quantum random access both sides of the equation by A−1, where A−1 is the inverse of A.
memory (qRAM)41 to map that vector into a quantum state: vj → vj . The The HHL algorithm then allows one to construct the quantum state
quantum state that summarizes the vector has logd qubits, and the proportional to A−1 b . More generally, when A is not square or has
­operation of the qRAM requires O(d) operations divided over O(logd) zero eigenvalues, the algorithm can be used to find the state x
steps that can be performed in parallel. Because vj was chosen at random, that minimizes9 A x − b .
the resulting quantum state has a density matrix ρ = (1/N ) ∑ j vj vj , The algorithm works as follows. Assume b = ∑n bn En where
where N is the number of data vectors. By comparison with the covariance En is an eigenvector of A with eigenvalue λn ≥​  Λ. By applying
matrix C for the classical data, we see that the density matrix for the phase estimation under A to compute λn and by rotating an
quantum version of the data actually is the covariance matrix, up to an ancillary qubit through an angle of arcsin(Λ/λn) and then
overall factor. By repeatedly sampling the data, and using a trick called uncomputing the phase estimation we obtain:
density matrix exponentiation42 combined with the quantum phase
­estimation algorithm39, which finds eigenvectors and eigenvalues of Λ Λ2 
 
∑ bn En  λ 1 + 1 − 2 0 
matrices, we can take the quantum version of any data vector v and λ 
n  n n 
decompose it into the principal components ck , revealing the eigenvalue
of C at the same time: v → ∑k vk ck ek . The properties of the principal If the ancillary qubit is measured and if 1 is observed then each
components of C can then be probed by making measurements on the eigenstate is divided through by λn, which affects the inverse. The
quantum representation of the eigenvectors of C. The quantum algorithm number of times that the state preparation circuit needs to be
scales as O[(logN)2] in both computational complexity and query applied to succeed, after applying amplitude amplification, is
­complexity. That is, quantum PCA is exponentially more efficient than O ( A /Λ ), which is the condition number for the matrix.
classical PCA. The HHL algorithm takes O[(logN)2] quantum steps to output
x , compared with the O(NlogN) steps required to find x using the
Quantum support vector machines and kernel methods best known method on a classical computer.
The simplest examples of supervised machine learning algorithms are There are several important caveats to the HHL algorithm. First,
linear support vector machines and perceptrons. These methods seek to finding the full answer x from the quantum state x requires O(N)
find an optimal separating hyperplane between two classes of data in a repetitions to reconstruct the N components of x. Generalizations to
dataset such that, with high probability, all training examples of one class HHL, such as least-squares fitting, sidestep this problem by allowing
are found only on one side of the hyperplane. The most robust classifier the output to have many fewer dimensions than the input. In
for the data is given when the margin between the hyperplane and the general, however, HHL can provide only features of the data such as
data are maximized. Here the ‘weights’ learned in the training are the moments of the solution vector or its expectation value x †Bx over
parameters of the hyperplane. One of the greatest powers of the support other sparse matrices B. The second caveat is that the input vector
vector machine lies in its generalization to nonlinear hypersurfaces via b needs to be prepared, either on a quantum computer or using
kernel functions43. Such classifiers have found great success in image qRAM, which may be expensive. The third caveat is that the matrix
segmentation as well as in the biological sciences. must be well conditioned and it must be possible to simulate e−iA
Like its classical counterpart, the quantum support vector machine is efficiently. Finally, although the HHL algorithms scales as O[(logN)2],
a paradigmatic example of a quantum machine learning algorithm13. A current estimates of the cost of the algorithm for practical problems
first quantum support vector machine was discussed in the early 2000s44, are prohibitive109, which underlines the importance of investigating
using a variant of Grover’s search for function minimization45. Finding s further improvements10. In general, the promise of exponential
­support vectors out of N vectors consequently takes N /s iterations. speedups for linear systems should be tempered with the
Recently, a least-squares quantum support vector machine was developed realization that they apply only to certain problems.
that harnesses the full power of the qBLAS subroutines. The data input
can come from various sources, such as from qRAM accessing classical
data or from a quantum subroutine preparing quantum states. Once the qBLAS-based optimization
data are made available to the quantum computing device, they are Many data analysis and machine learning techniques involve o­ ptimization.
­processed with quantum phase estimation and matrix inversion (the HHL Of increasing interest is the use of D-Wave processors to solve combi-
algorithm). All the operations required to construct the optimal separating natorial optimization problems by means of quantum annealing. Some
hyperplane and to test whether a vector lies on one side or the other can optimization problems can also be formulated as a single-shot ­solution
in principle be performed in time that is polynomial in logN, where N is of a linear system, such as the optimization of a quadratic function
the dimension of the matrix required to prepare a quantum version of the ­subject to equality constraints, a subset of quadratic programming
hyperplane vector. Polynomial13 and radial basis function kernels46 are ­problems. If the matrices involved are sparse or low rank, such ­problems
discussed, as well as another kernel-based method called Gaussian can be solved in time that is polynomial in logd, where d is the system
­process regression47. This approach to quantum support machines has dimension via the HHL matrix inversion algorithm, yielding an expo-
been experimentally demonstrated in a nuclear magnetic resonance nential speedup over classical algorithms, which run in time that is
­testbed for a handwritten digit recognition task48. polynomial in d.

1 4 s e p t e m b e r 2 0 1 7 | V O L 5 4 9 | N A T U RE | 1 9 7
© 2017 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.
 INSIGHT Review

a a serious bottleneck for certain algorithms. Likewise, the ‘output problem’

is faced when reading out data after being processed on a quantum device.
Thermal annealing Like the input problem, the output problem often causes a noticeable
operational slowdown.
In particular, if we wish to apply HHL, least-squares fitting, quantum
principal component analysis, quantum support vector machines, and
related approaches to classical data, the procedure begins by first loading
Quantum  considerable amounts of data into a quantum system, which can require
annealing exponential time51. This can be addressed in principle using qRAM but
the cost of doing so may be prohibitive for big data problems52. Apart
from combinatorial-optimization-based approaches, the only known
linear-algebra-based quantum machine learning algorithm that does
not rely on large-scale qRAM is the quantum algorithm for performing
topological analysis of data (persistent homology)14. With the notable
exceptions of least-squares fitting and quantum support vector machines,
linear-algebra-based algorithms can also suffer from the output problem
because desirable classical quantities such as the solution vector for HHL
or the principal components for PCA are exponentially hard to estimate.
Despite the potential for exponential quantum speedups, without much
effort put into optimization, the circuit size and circuit depth overhead
b can ­balloon (to around 1025 quantum gates in one proposed realization
of HHL53). Ongoing work is needed to optimize such algorithms, provide
better cost estimates and ultimately to understand the sort of quantum
computer that we would need to provide useful quantum alternatives to
classical machine learning.

Deep quantum learning

Classical deep neural networks are highly effective tools for machine
learning and are well suited to inspire the development of deep ­quantum
learning methods. Special-purpose quantum information processors such
as quantum annealers and programmable photonic circuits are well suited
for constructing deep quantum learning networks21,54,55. The simplest
deep neural network to quantize is the Boltzmann machine (see Box 3
Thermal state
and Box 3 Figure). The ­classical Boltzmann machine consists of bits
with tunable interactions: the Boltzmann machine is trained by adjusting
those interactions so that the thermal statistics of the bits, described by a
Boltzmann–Gibbs distribution (see Fig. 1b), reproduces the statistics of
the data. To quantize the Boltzmann machine one simply takes the neural
network and expresses it as a set of interacting quantum spins, corre-
sponding to a tunable Ising model. Then by initializing the input neurons
Figure 1 | Quantum tunnelling versus thermalization. A quantum state in the Boltzmann machines into a fixed state and allowing the system to
tunnels when approaching a resonance point before decoherence induces thermalize, we can read out the output qubits to obtain an answer.
thermalization. Shades of blue illustrate occupation of energy levels (black An essential feature of deep quantum learning is that it does not require
dashes). a, A quantum state must traverse a local minimum in thermal a large, general-purpose quantum computer. Quantum ­annealers are
annealing, whereas a coherent quantum state can tunnel when brought ­special-purpose quantum information processors that are much easier
close to resonance. b, Coherent effects decay through interaction with an to construct and to scale up than are general-purpose quantum c­ omputers
environment, causing the probability distribution of the occupancy of a (see Fig. 1a). Quantum annealers are well suited for implementing deep
system’s energy levels to follow a Gibbs distribution. quantum learners, and are commercially available. The D-Wave quantum
annealer is a tunable transverse Ising model that can be programmed to
Most methods in machine learning require iterative optimization of yield the thermal states of classical systems, and certain quantum spin
their performance. As an example, inequality constraints are often han- systems. The D-Wave device has been used to perform deep quantum
dled via penalty functions49 and variations of gradient descent or Newton’s learning protocols on more than a thousand spins56. Quantum Boltzmann
method. A modification of the quantum PCA method implements itera- machines22 with more general tunable couplings, capable of implementing
tive gradient descent and Newton’s methods for polynomial optimization, universal quantum logic, are currently at the design stage57. On-chip
and can again provide an exponential speedup over classical methods19. ­silicon waveguides have been used to construct linear optical arrays with
Multiple copies of the present solution, encoded in a quantum state, are ­hundreds of tunable interferometers, and special-purpose superconducting
used to improve that solution at each step. Brandao and Svore provide quantum information processors could be used to implement the
a quantum version of semi-definite programming that holds out the ­quantum approximate optimization algorithm.
possibility of super-polynomial speedups18. The quantum approxi- There are several ways that quantum computers can provide advantages
mate optimization algorithm (the QAO algorithm)50 provides a unique here. First, quantum methods can make the system thermalize quadra­
approach to optimization based on alternating qubit rotations with the tically faster than its classical counterpart20,58–60. This can make ­accurate
application of the problem’s penalty function. training of fully connected Boltzmann machines practical. Second,
­quantum computers can accelerate Boltzmann training by providing
Reading classical data into quantum machines improved ways of sampling. Because the neuron activation pattern in
Classical data must be input before being processed on a quantum the Boltzmann machine is stochastic, many repetitions are needed to
­computer. This ‘input problem’ often has little overhead but can present determine success probabilities, and in turn, to discover the effect that

1 9 8 | N A T U RE | V O L 5 4 9 | 1 4 s e p t e m b e r 2 0 1 7
© 2017 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.
 Review INSIGHT

changing a weight in the neural network has on the performance of the

Box 3 deep network. When training a quantum Boltzmann machine, in ­contrast,
Training quantum Boltzmann quantum coherence can quadratically reduce the number of samples
needed to learn the desired task. Furthermore, quantum access to the
machines training data (that is, qRAM or a quantum blackbox subroutine) allows
the machine to be trained using quadratically fewer access requests to the
In quantum Boltzmann machine training we wish training data than are required by classical methods: a quantum algorithm
to learn a set of Hamiltonian parameters (wj) such that for can train a deep neural network on a large training dataset while reading
a fixed set of Hj we have that our input state ρtrain is well only a minuscule number of training vectors20.
approximated22,61 by σ = e− ∑ j wjHj /Tr(e− ∑ j wjHj ). For all visible Quantum information processing provides new, fundamentally
Boltzmann machines, the quantum relative entropy ­quantum, models for deep learning. For example, adding a transverse
S(ρ train σ ) = Tr[ρ train log(ρ train) − ρ train log(σ )] is the most field to the Ising model quantum Boltzmann machine can induce a vari-
logical way to measure the quality of the approximation. It is ety of quantum effects such as tunnelling22,61. Adding further quantum
easy to see (assuming that the kernels of ρ and σ coincide) that couplings transforms the quantum Boltzmann machine into a variety of
the quantum relative entropy provides an upper bound for the quantum systems57,62. Adding a tunable transverse interaction to a ­tunable
distance between the two states. Thus, minimizing it minimizes Ising model is known to be universal for full quantum c­ omputing57: with
the error in approximating the state. the proper weight assignments this model can execute any algorithm
Although the relative entropy is an excellent measure of that a general-purpose quantum computer can perform. Such universal
the distance between two states, it can be difficult to discover deep quantum learners may recognize and classify patterns that classical
experimentally. However, the gradient (that is, the direction of ­computers cannot.
greatest change) of the relative entropy is easy to estimate61: Unlike classical Boltzmann machines, quantum Boltzmann machines
output a quantum state. Thus deep quantum networks can learn to
∂w jS (ρ train σ ) = Tr(σHj ) Tr(ρHj ) ­generate quantum states representative of a wide variety of systems,
allowing the network to act as a form of quantum associative memory63.
Given an experimental dataset of expectation values for ρtrain and a This ­ability to generate quantum states is absent from classical machine
quantum simulator for Tr(σHj) we can find the direction of greatest learning. Thus quantum Boltzmann training has applications beyond
improvement in the quantum relative entropy. Gradient descent classifying quantum states and providing richer models for classical data.
then is used to update w via w → w − η∇S (ρ train σ ) for η >​  0.
Stoquastic (quantum stochastic) Hamiltonians have the property Quantum machine learning for quantum data
that all off-diagonal matrix elements in the standard basis are real Perhaps the most immediate application of quantum machine learning
and non-positive (equivalently non-negative). No efficient classical is to quantum data—the actual states generated by quantum systems
analogue of this method is known in general for non-stoquastic H and processes. As described above, many quantum machine l­earning
(see ref. 57). ­algorithms find patterns in classical data by mapping the data to q­ uantum
We show this protocol below for learning a random state mechanical states, and then manipulating those states using basic q­ uantum
formed from a uniform mixture of four random states—random linear algebra subroutines. These quantum machine ­learning algorithms
with respect to the unique and unitarily invariant Haar measure. can be applied directly to the quantum states of light and of matter to reveal
Fewer than ten gradient steps (epochs) are needed to train it to their underlying features and patterns. The ­resulting quantum modes of
approximately generate ρtrain using a complete set of Hamiltonian analysis are frequently much more efficient and more illuminating than
terms. the classical analysis of data taken from ­quantum s­ ystems. For example,
given multiple copies of a system described by an N ×​ N density matrix,
­quantum principal component analysis can be used to find its eigenvalues
and to reveal the corresponding eigenvectors in time O[(logN)2], compared
with the O(N2) measurements needed for a classical device to p ­ erform
tomography on a density matrix, and the O(N2) operations needed to
­perform the classical PCA. Such quantum analysis of quantum data could
profitably be performed on the relatively small quantum computers that
are likely to be available over the next several years.
A particularly powerful quantum data analysis technique is the use
of quantum simulators to probe quantum dynamics. Quantum simu-
lators are ‘quantum analogue computers’—quantum systems whose
dynamics can be programmed to match the dynamics of some desired
quantum system. A quantum simulator can either be a special-purpose
device constructed to simulate a particular class of quantum systems,
or a general-purpose quantum computer. By connecting a trusted
quantum simulator to an unknown system and tuning the model of the
simulator to counteract the unknown dynamics, the dynamics of the
unknown system can be efficiently learned using approximate Bayesian
inference64–66. This exponentially reduces the number of measurements
needed to perform the simulation. Similarly, the universal quantum
emulator algorithm67 allows one to reconstruct quantum dynamics and
the quantum Boltzmann training algorithm of ref. 61 allows states to
be reconstructed, in time logarithmic in the dimension of the Hilbert
space, which is exponentially faster than reconstructing the dynamics
Box 3 Figure | Learning a random state using a quantum Boltzmann via classical tomography.
machine. To use a quantum computer to help characterize a quantum system65,66
or to accept input states for use in a quantum PCA algorithm, we must

1 4 s e p t e m b e r 2 0 1 7 | V O L 5 4 9 | N A T U RE | 1 9 9
© 2017 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.
 INSIGHT Review

face the substantial technical challenge of loading coherent input states. for small quantum computers23–31,96–98 complemented and enhanced by
Nonetheless, because such applications do not require qRAM and offer special-purpose quantum information processors21,22, digital quantum
the potential for exponential speedups for device characterization22,61,65,66 processors70,73,78,99,100 and sensors76,77,101.
they remain among the promising possibilities for near-term application In particular, quantum annealers with around 2,000 qubits have been
of quantum machine learning. built and operated, using integrated superconducting circuits that are,
in principle, scalable. The biggest challenges for quantum annealers to
Designing and controlling quantum systems implement quantum machine learning algorithms include improving
A major challenge in the development of quantum computation and connectivity and implementing more general tunable couplings between
information science involves tuning quantum gates to match the ­exacting qubits. Programmable quantum optic arrays with around 100 tunable
requirements needed for quantum error correction. Heuristic search interferometers have been constructed using integrated photonics in sil-
methods can help to achieve this in a supervised learning scenario68,69 icon, but loss of quantum effects increases as such circuits are scaled up.
(for instance in the case of nearest-neighbour-coupled ­superconducting A particularly important challenge for quantum machine learning is the
­artificial atoms69 with gate fidelity above 99.9% in the presence of noise) construction of interface devices such as qRAM that allow classical infor-
and thus to reach an accepted threshold for fault-tolerant quantum mation to be encoded in quantum mechanical form52. A qRAM to access
­computing. A similar methodology has been successful in constructing a N pieces of data consists of a branching array of 2N quantum switches,
single-shot Toffoli gate, again reaching gate fidelity above 99.9%70. Genetic which must operate coherently during a memory call. In principle, such
algorithms have been employed to reduce digital and experimental errors a qRAM takes time O(logN) to perform a memory call, and can tolerate
in quantum gates71. They have been used to simulate controlled-NOT error rates of up to O(1/logN) per switching operation, where logN is the
gates by means of ancillary qubits and imperfect gates. Besides outper- depth of the qRAM circuit. Proof-of-principle demonstrations of qRAM
forming protocols for digital quantum simulations, it has been shown that have been performed, but constructing large arrays of quantum switches
genetic algorithms are also useful for suppressing experimental errors is a difficult technological problem.
in gates72. Another approach used stochastic gradient descent and two- These hardware challenges are technical in nature, and clear paths exist
body interactions to embed a Toffoli gate into a sequence of quantum towards overcoming them. They must be overcome, however, if quantum
operations or gates without time-dependent control using the natural machine learning is to become a ‘killer app’ for quantum computers. As
dynamics of a quantum network73. Dynamical decoupling sequences help noted previously, most of the quantum algorithms that have been identi-
to protect quantum states from decoherence, which can be designed using fied face a number of caveats that limits their applicability. We can distill
recurrent neural networks74. the caveats mentioned above into four fundamental problems.
Controlling a quantum system is just as important and complex.
Learning methods have also been very successful in developing ­control (1) Th
 e input problem. Although quantum algorithms can provide dra-
sequences to optimize adaptive quantum metrology, which is a key matic speedups for processing data, they seldom provide advantages
­quantum building block in many quantum technologies. Genetic in reading data. This means that the cost of reading in the input can
­algorithms have been proposed for the control of quantum molecules to in some cases dominate the cost of quantum algorithms. Under-
overcome the problem caused by changing environmental ­parameters standing this factor is an ongoing challenge.
­during an experiment75. Reinforcement learning algorithms using (2) The output problem. Obtaining the full solution from some quan-
­heuristic global optimization, like the algorithm used for designing tum algorithms as a string of bits requires learning an exponential
­circuits, have been widely successful, particularly in the presence of number of bits. This makes some applications of quantum machine
noise and ­decoherence, scaling well with the system size76–78. One can learning algorithms infeasible. This problem can potentially be side-
also exploit reinforcement learning in gate-based quantum systems. For stepped by learning only summary statistics for the solution state.
instance, adaptive controllers based on intelligent agents for q ­ uantum (3) The costing problem. Closely related to the input/output ­problems, at
information demonstrate adaptive calibration and compensation ­strategies present very little is known about the true number of gates ­required
to an ­external stray field of unknown magnitude in a fixed direction. by quantum machine learning algorithms. Bounds on the complexity
Classical machine learning is also a powerful tool with which to extract suggest that for sufficiently large problems they will ­offer huge
theoretical insights about quantum states. Neural networks have recently advantages, but it is still unclear when that crossover point occurs.
been deployed to study two central problems in condensed matter, (4) The benchmarking problem. It is often difficult to assert that a
namely phase-of-matter detection79,80 and ground-state search81. These quantum algorithm is ever better than all known classical machine
succeeded in achieving better performances than established numerical algorithms in practice because this would require extensive bench-
tools. Theoretical physicists are now studying these models to understand marking against modern heuristic methods. Establishing lower
analytically their descriptive power compared to traditional methods such bounds for quantum machine learning would partially address
as tensor networks. Interesting applications to exotic states of matter are this issue.
already on the market, and have been shown to capture highly non-trivial
features from disordered or topologically ordered systems. To avoid some of these problems, we could apply quantum computing
to quantum, rather than classical, data. One aim therein is to use ­quantum
Perspectives on future work machine learning to characterize and control quantum computers66. This
As we have discussed in this review, small quantum computers and would enable a virtuous cycle of innovation similar to that which occurred
larger special-purpose quantum simulators, annealers and so on seem to in classical computing, wherein each generation of processors is then
have potential use in machine learning and data analysis15,21,22,36,48,82–95. leveraged to design the next-generation processors. We have already
However, the execution of quantum algorithms requires quantum begun to see the first fruits of this cycle with classical machine learning
­hardware that is not yet available. being used to improve quantum processor designs23–31,102–104, which in
On the hardware side, there have been great strides in several enabling turn provide powerful computational resources for quantum-enhanced
technologies. Small-scale quantum computers with 50–100 qubits will machine learning applications themselves8,9,11,13,33–36.
be made widely available via quantum cloud computing (the ‘Qloud’).
Special-purpose quantum information processors such as quantum received 20 February; accepted 4 July 2017.
­simulators, quantum annealers, integrated photonic chips, nitrogen
vacancy centres (NV)-diamond arrays, qRAM, and made-to-order 1. Rosenblatt, F. The perceptron: a probabilistic model for information storage
and organization in the brain. Psychol. Rev. 65, 386 (1958).
superconducting circuits will continue to advance in size and complex- 2. LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521, 436–444
ity. Quantum machine learning offers a suite of potential applications (2015).

2 0 0 | N A T U RE | V O L 5 4 9 | 1 4 s e p t e m b e r 2 0 1 7
© 2017 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.
 Review INSIGHT

3. Le, Q. V. Building high-level features using large scale unsupervised learning. 37. Aïmeur, E., Brassard, G. & Gambs, S. in Machine Learning in a Quantum World
In IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP) 431–442 (Springer, 2006).
8595–8598 (IEEE, 2013). 38. Shor, P. W. Polynomial-time algorithms for prime factorization and discrete
4. Schuld, M., Sinayskiy, I. & Petruccione, F. An introduction to quantum machine logarithms on a quantum computer. SIAM J. Comput. 26, 1484–1509
learning. Contemp. Phys. 56, 172–185 (2015). (1997).
5. Wittek, P. Quantum Machine Learning: What Quantum Computing Means to Data 39. Nielsen, M. A. & Chuang, I. L. Quantum Computation and Quantum Information
Mining (Academic Press, New York, NY, USA, 2014). (Cambridge Univ. Press, 2000).
6. Adcock, J. et al. Advances in quantum machine learning. Preprint at https:// 40. Wossnig, L., Zhao, Z. & Prakash, A. A quantum linear system algorithm for
arxiv.org/abs/1512.02900 (2015). dense matrices. Preprint at https://arxiv.org/abs/1704.06174 (2017).
7. Arunachalam, S. & de Wolf, R. A survey of quantum learning theory. Preprint at 41. Giovannetti, V., Lloyd, S. & Maccone, L. Quantum random access memory.
https://arxiv.org/abs/1701.06806 (2017). Phys. Rev. Lett. 100, 160501 (2008).
8. Harrow, A. W., Hassidim, A. & Lloyd, S. Quantum algorithm for linear systems 42. Lloyd, S. Universal quantum simulators. Science 273, 1073–1078 (1996).
of equations. Phys. Rev. Lett. 103, 150502 (2009). 43. Vapnik, V. The Nature of Statistical Learning Theory (Springer, 1995).
9. Wiebe, N., Braun, D. & Lloyd, S. Quantum algorithm for data fitting. Phys. Rev. 44. Anguita, D., Ridella, S., Rivieccio, F. & Zunino, R. Quantum optimization for
Lett. 109, 050505 (2012). training support vector machines. Neural Netw. 16, 763–770 (2003).
10. Childs, A. M., Kothari, R. & Somma, R. D. Quantum linear systems algorithm 45. Dürr, C. & Høyer, P. A quantum algorithm for finding the minimum. Preprint at
with exponentially improved dependence on precision. Preprint at https:// https://arxiv.org/abs/quant-ph/9607014 (1996).
arxiv.org/abs/1511.02306 (2015). 46. Chatterjee, R. & Yu, T. Generalized coherent states, reproducing kernels, and
11. Lloyd, S., Mohseni, M. & Rebentrost, P. Quantum principal component quantum support vector machines. Preprint at https://arxiv.org/abs/1612.03713
analysis. Nat. Phys. 10, 631–633 (2014). (2016).
12. Kimmel, S., Lin, C. Y.-Y., Low, G. H., Ozols, M. & Yoder, T. J. Hamiltonian 47. Zhao, Z., Fitzsimons, J. K. & Fitzsimons, J. F. Quantum assisted Gaussian
simulation with optimal sample complexity. Preprint at https://arxiv.org/ process regression. Preprint at https://arxiv.org/abs/1512.03929 (2015).
abs/1608.00281 (2016). 48. Li, Z., Liu, X., Xu, N. & Du, J. Experimental realization of a quantum support
13. Rebentrost, P., Mohseni, M. & Lloyd, S. Quantum support vector machine for vector machine. Phys. Rev. Lett. 114, 140504 (2015).
big data classification. Phys. Rev. Lett. 113, 130503 (2014). 49. Whitfield, J. D., Faccin, M. & Biamonte, J. D. Ground-state spin logic. Europhys.
This study applies quantum matrix inversion in a supervised discriminative Lett. 99, 57004 (2012).
learning algorithm. 50. Farhi, E., Goldstone, J. & Gutmann, S. A quantum approximate optimization
14. Lloyd, S., Garnerone, S. & Zanardi, P. Quantum algorithms for topological and algorithm. Preprint at https://arxiv.org/abs/1411.4028 (2014).
geometric analysis of data. Nat. Commun. 7, 10138 (2016). 51. Aaronson, S. Read the fine print. Nat. Phys. 11, 291–293 (2015).
15. Dridi, R. & Alghassi, H. Homology computation of large point clouds using 52. Arunachalam, S., Gheorghiu, V., Jochym-O’Connor, T., Mosca, M. & Srinivasan,
quantum annealing. Preprint at https://arxiv.org/abs/1512.09328 (2015). P. V. On the robustness of bucket brigade quantum RAM. New J. Phys. 17,
16. Rebentrost, P., Steffens, A. & Lloyd, S. Quantum singular value decomposition 123010 (2015).
of non-sparse low-rank matrices. Preprint at https://arxiv.org/abs/1607.05404 53. Scherer, A. et al. Concrete resource analysis of the quantum linear system
(2016). algorithm used to compute the electromagnetic scattering cross section of a
17. Schuld, M., Sinayskiy, I. & Petruccione, F. Prediction by linear regression on a 2D target. Preprint at https://arxiv.org/abs/1505.06552 (2015).
quantum computer. Phys. Rev. A 94, 022342 (2016). 54. Denil, M. & De Freitas, N. Toward the implementation of a quantum RBM. In
Neural Information Processing Systems (NIPS) Conf. on Deep Learning and
18. Brandao, F. G. & Svore, K. Quantum speed-ups for semidefinite programming.
Unsupervised Feature Learning Workshop Vol. 5 (2011).
Preprint at https://arxiv.org/abs/1609.05537 (2016).
55. Dumoulin, V., Goodfellow, I. J., Courville, A. & Bengio, Y. On the challenges of
19. Rebentrost, P., Schuld, M., Petruccione, F. & Lloyd, S. Quantum gradient
physical implementations of RBMs. Preprint at https://arxiv.org/abs/1312.5258
descent and Newton’s method for constrained polynomial optimization.
(2013).
Preprint at https://arxiv.org/abs/1612.01789 (2016).
56. Benedetti, M., Realpe-Gómez, J., Biswas, R. & Perdomo-Ortiz, A. Estimation of
20. Wiebe, N., Kapoor, A. & Svore, K. M. Quantum deep learning. Preprint at
effective temperatures in quantum annealers for sampling applications:
https://arxiv.org/abs/1412.3489 (2014).
a case study with possible applications in deep learning. Phys. Rev. A 94,
21. Adachi, S. H. & Henderson, M. P. Application of quantum annealing to training
022308 (2016).
of deep neural networks. Preprint at https://arxiv.org/abs/arXiv:1510.06356
57. Biamonte, J. D. & Love, P. J. Realizable Hamiltonians for universal adiabatic
(2015). quantum computers. Phys. Rev. A 78, 012352 (2008).
22. Amin, M. H., Andriyash, E., Rolfe, J., Kulchytskyy, B. & Melko, R. Quantum This study established the contemporary experimental target for non-
Boltzmann machine. Preprint at https://arxiv.org/abs/arXiv:1601.02036 stoquastic (that is, non-quantum stochastic) D-Wave quantum annealing
(2016). hardware able to realize universal quantum Boltzmann machines.
23. Sasaki, M., Carlini, A. & Jozsa, R. Quantum template matching. Phys. Rev. A 64, 58. Temme, K., Osborne, T. J., Vollbrecht, K. G., Poulin, D. & Verstraete, F. Quantum
022317 (2001). metropolis sampling. Nature 471, 87–90 (2011).
24. Bisio, A., Chiribella, G., D’Ariano, G. M., Facchini, S. & Perinotti, P. Optimal 59. Yung, M.-H. & Aspuru-Guzik, A. A quantum–quantum metropolis algorithm.
quantum learning of a unitary transformation. Phys. Rev. A 81, 032324 Proc. Natl Acad. Sci. USA 109, 754–759 (2012).
(2010). 60. Chowdhury, A. N. & Somma, R. D. Quantum algorithms for Gibbs sampling
25. Bisio, A., D’Ariano, G. M., Perinotti, P. & Sedlák, M. Quantum learning and hitting-time estimation. Quant. Inf. Comput. 17, 41–64 (2017).
algorithms for quantum measurements. Phys. Lett. A 375, 3425–3434 (2011). 61. Kieferova, M. & Wiebe, N. Tomography and generative data modeling via
26. Sentís, G., Calsamiglia, J., Muñoz-Tapia, R. & Bagan, E. Quantum learning quantum Boltzmann training. Preprint at https://arxiv.org/abs/1612.05204
without quantum memory. Sci. Rep. 2, 708 (2012). (2016).
27. Sentís, G., Guţă, M. & Adesso, G. Quantum learning of coherent states. 62. Lloyd, S. & Terhal, B. Adiabatic and Hamiltonian computing on a 2D lattice
EPJ Quant. Technol. 2, 17 (2015). with simple 2-qubit interactions. New J. Phys. 18, 023042 (2016).
28. Paparo, G. D., Dunjko, V., Makmal, A., Martin-Delgado, M. A. & Briegel, H. J. 63. Ventura, D. & Martinez, T. Quantum associative memory. Inf. Sci. 124,
Quantum speedup for active learning agents. Phys. Rev. X 4, 031002 (2014). 273–296 (2000).
29. Dunjko, V., Friis, N. & Briegel, H. J. Quantum-enhanced deliberation of learning 64. Granade, C. E., Ferrie, C., Wiebe, N. & Cory, D. G. Robust online Hamiltonian
agents using trapped ions. New J. Phys. 17, 023006 (2015). learning. New J. Phys. 14, 103013 (2012).
30. Dunjko, V., Taylor, J. M. & Briegel, H. J. Quantum-enhanced machine learning. 65. Wiebe, N., Granade, C., Ferrie, C. & Cory, D. G. Hamiltonian learning and
Phys. Rev. Lett. 117, 130501 (2016). certification using quantum resources. Phys. Rev. Lett. 112, 190501
This paper investigates the theoretical maximum speedup achievable in (2014).
reinforcement learning in a closed quantum system, which proves to be 66. Wiebe, N., Granade, C. & Cory, D. G. Quantum bootstrapping via compressed
Grover-like if we wish to obtain classical verification of the learning process. quantum Hamiltonian learning. New J. Phys. 17, 022005 (2015).
31. Sentís, G., Bagan, E., Calsamiglia, J., Chiribella, G. & Muñoz Tapia, R. Quantum 67. Marvian, I. & Lloyd, S. Universal quantum emulator. Preprint at https://arxiv.
change point. Phys. Rev. Lett. 117, 150502 (2016). org/abs/1606.02734 (2016).
32. Faccin, M., Migdał, P., Johnson, T. H., Bergholm, V. & Biamonte, J. D. 68. Dolde, F. et al. High-fidelity spin entanglement using optimal control.
Community detection in quantum complex networks. Phys. Rev. X 4, 041012 Nat. Commun. 5, 3371 (2014).
(2014). 69. Zahedinejad, E., Ghosh, J. & Sanders, B. C. Designing high-fidelity single-shot
This paper defines closeness measures and then maximizes modularity with three-qubit gates: a machine-learning approach. Phys. Rev. Appl. 6, 054005
hierarchical clustering to partition quantum data. (2016).
33. Clader, B. D., Jacobs, B. C. & Sprouse, C. R. Preconditioned quantum linear 70. Zahedinejad, E., Ghosh, J. & Sanders, B. C. High-fidelity single-shot Toffoli gate
system algorithm. Phys. Rev. Lett. 110, 250504 (2013). via quantum control. Phys. Rev. Lett. 114, 200502 (2015).
34. Lloyd, S., Mohseni, M. & Rebentrost, P. Quantum algorithms for supervised 71. Zeidler, D., Frey, S., Kompa, K.-L. & Motzkus, M. Evolutionary algorithms
and unsupervised machine learning. Preprint at https://arxiv.org/ and their application to optimal control studies. Phys. Rev. A 64, 023420
abs/1307.0411 (2013). (2001).
35. Wiebe, N., Kapoor, A. & Svore, K. M. Quantum algorithms for nearest-neighbor 72. Las Heras, U., Alvarez-Rodriguez, U., Solano, E. & Sanz, M. Genetic algorithms
methods for supervised and unsupervised learning. Quantum Inf. Comput. 15, for digital quantum simulations. Phys. Rev. Lett. 116, 230504 (2016).
316–356 (2015). 73. Banchi, L., Pancotti, N. & Bose, S. Quantum gate learning in qubit networks:
36. Lau, H.-K., Pooser, R., Siopsis, G. & Weedbrook, C. Quantum machine learning Toffoli gate without time-dependent control. npj Quant. Inf. 2, 16019
over infinite dimensions. Phys. Rev. Lett. 118, 080501 (2017). (2016).

1 4 s e p t e m b e r 2 0 1 7 | V O L 5 4 9 | N A T U RE | 2 0 1
© 2017 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.
 INSIGHT Review

74. August, M. & Ni, X. Using recurrent neural networks to optimize dynamical 97. Schuld, M., Fingerhuth, M. & Petruccione, F. Quantum machine learning
decoupling for quantum memory. Preprint at https://arxiv.org/ with small-scale devices: implementing a distance-based classifier with a
abs/1604.00279 (2016). quantum interference circuit. Preprint at https://arxiv.org/abs/1703.10793
75. Amstrup, B., Toth, G. J., Szabo, G., Rabitz, H. & Loerincz, A. Genetic algorithm (2017).
with migration on topology conserving maps for optimal control of quantum 98. Monràs, A., Sentís, G. & Wittek, P. Inductive supervised quantum learning.
systems. J. Phys. Chem. 99, 5206–5213 (1995). Phys. Rev. Lett. 118, 190503 (2017).
76. Hentschel, A. & Sanders, B. C. Machine learning for precise quantum This paper proves that supervised learning protocols split into a training and
measurement. Phys. Rev. Lett. 104, 063603 (2010). application phase in both the classical and the quantum cases.
77. Lovett, N. B., Crosnier, C., Perarnau-Llobet, M. & Sanders, B. C. Differential 99. Tiersch, M., Ganahl, E. J. & Briegel, H. J. Adaptive quantum computation in
evolution for many-particle adaptive quantum metrology. Phys. Rev. Lett. 110, changing environments using projective simulation. Sci. Rep. 5, 12874
220501 (2013). (2015).
78. Palittapongarnpim, P., Wittek, P., Zahedinejad, E., Vedaie, S. & Sanders, B. C. 100. Zahedinejad, E., Ghosh, J. & Sanders, B. C. Designing high-fidelity single-shot
Learning in quantum control: high-dimensional global optimization for noisy three-qubit gates: a machine learning approach. Preprint at https://arxiv.org/
quantum dynamics. Neurocomputing https://doi.org/10.1016/j.neucom. abs/1511.08862 (2015).
2016.12.087 (in the press). 101. Palittapongarnpim, P., Wittek, P. & Sanders, B. C. Controlling adaptive quantum
79. Carrasquilla, J. & Melko, R. G. Machine learning phases of matter. Nat. Phys. phase estimation with scalable reinforcement learning. In Proc. 24th Eur.
13, 431–434 (2017). Symp. Artificial Neural Networks (ESANN-16) on Computational Intelligence and
80. Broecker, P., Carrasquilla, J., Melko, R. G. & Trebst, S. Machine learning Machine Learning 327–332 (2016).
quantum phases of matter beyond the fermion sign problem. Preprint at 102. Wan, K. H., Dahlsten, O., Kristjánsson, H., Gardner, R. & Kim, M. S. Quantum
https://arxiv.org/abs/1608.07848 (2016). generalisation of feedforward neural networks. Preprint at https://arxiv.org/
81. Carleo, G. & Troyer, M. Solving the quantum many-body problem with artificial abs/1612.01045 (2016).
neural networks. Science 355, 602–606 (2017). 103. Lu, D. et al. Towards quantum supremacy: enhancing quantum control by
82. Brunner, D., Soriano, M. C., Mirasso, C. R. & Fischer, I. Parallel photonic bootstrapping a quantum processor. Preprint at https://arxiv.org/
information processing at gigabyte per second data rates using transient abs/1701.01198 (2017).
states. Nat. Commun. 4, 1364 (2013). 104. Mavadia, S., Frey, V., Sastrawan, J., Dona, S. & Biercuk, M. J. Prediction and
83. Cai, X.-D. et al. Entanglement-based machine learning on a quantum real-time compensation of qubit decoherence via machine learning.
computer. Phys. Rev. Lett. 114, 110504 (2015). Nat. Commun. 8, 14106 (2017).
84. Hermans, M., Soriano, M. C., Dambre, J., Bienstman, P. & Fischer, I. Photonic 105. Rønnow, T. F. et al. Defining and detecting quantum speedup. Science 345,
delay systems as machine learning implementations. J. Mach. Learn. Res. 16, 420–424 (2014).
2081–2097 (2015). 106. Low, G. H., Yoder, T. J. & Chuang, I. L. Quantum inference on Bayesian
85. Tezak, N. & Mabuchi, H. A coherent perceptron for all-optical learning. networks. Phys. Rev. A 89, 062315 (2014).
EPJ Quant. Technol. 2, 10 (2015). 107. Wiebe, N. & Granade, C. Can small quantum systems learn? Preprint at
86. Neigovzen, R., Neves, J. L., Sollacher, R. & Glaser, S. J. Quantum pattern https://arxiv.org/abs/1512.03145 (2015).
recognition with liquid-state nuclear magnetic resonance. Phys. Rev. A 79, 108. Wiebe, N., Kapoor, A. & Svore, K. M. Quantum perceptron models. Adv. Neural
042321 (2009). Inform. Process. Syst. 29, 3999–4007 (2016).
87. Pons, M. et al. Trapped ion chain as a neural network: error resistant quantum 109. Scherer, A. et al. Concrete resource analysis of the quantum linear-system
computation. Phys. Rev. Lett. 98, 023003 (2007). algorithm used to compute the electromagnetic scattering cross section of a
88. Neven, H. et al. Binary classification using hardware implementation of 2D target. Quantum Inform. Process. 16, 60 (2017).
quantum annealing. In 24th Ann. Conf. on Neural Information Processing
Systems (NIPS-09) 1–17 (2009). Acknowledgements J.B. acknowledges financial support from AFOSR
This paper was among the first experimental demonstrations of machine grant FA9550-16-1-0300, Models and Protocols for Quantum Distributed
learning using quantum annealing. Computation. P.W. acknowledges financial support from the ERC (Consolidator
89. Denchev, V. S., Ding, N., Vishwanathan, S. & Neven, H. Robust classification Grant QITBOX), Spanish Ministry of Economy and Competitiveness (Severo
with adiabatic quantum optimization. In Proc. 29th Int. Conf. on Machine Ochoa Programme for Centres of Excellence in R&D SEV-2015-0522 and
Learning (ICML-2012) (2012). QIBEQI FIS2016-80773-P), Generalitat de Catalunya (CERCA Programme and
90. Karimi, K. et al. Investigating the performance of an adiabatic quantum SGR 875), and Fundacio Privada Cellex. P.R. and S.L. acknowledge funding from
optimization processor. Quantum Inform. Process. 11, 77–88 (2012). ARO and AFOSR under MURI programmes. We thank L. Zheglova for producing
91. O’Gorman, B. A. et al. Bayesian network structure learning using quantum Fig. 1.
annealing. EPJ Spec. Top. 224, 163–188 (2015).
92. Denchev, V. S., Ding, N., Matsushima, S., Vishwanathan, S. V. N. & Neven, H. Author Contributions All authors designed the study, analysed data, interpreted
Totally corrective boosting with cardinality penalization. Preprint at data, produced Box 3 Figure and wrote the article.
https://arxiv.org/abs/1504.01446 (2015).
93. Kerenidis, I. & Prakash, A. Quantum recommendation systems. Preprint at Author Information Reprints and permissions information is available at
https://arxiv.org/abs/1603.08675 (2016). www.nature.com/reprints. The authors declare no competing financial
94. Alvarez-Rodriguez, U., Lamata, L., Escandell-Montero, P., Martín-Guerrero, J. D. interests. Readers are welcome to comment on the online version of the paper.
& Solano, E. Quantum machine learning without measurements. Preprint at Publisher’s note: Springer Nature remains neutral with regard to jurisdictional
https://arxiv.org/abs/1612.05535 (2016). claims in published maps and institutional affiliations. Correspondence should
95. Wittek, P. & Gogolin, C. Quantum enhanced inference in Markov logic be addressed to J.B. ([email protected]).
networks. Sci. Rep. 7, 45672 (2017).
96. Lamata, L. Basic protocols in quantum reinforcement learning with Reviewer Information Nature thanks L. Lamata and the other anonymous
superconducting circuits. Sci. Rep. 7, 1609 (2017). reviewer(s) for their contribution to the peer review of this work.

2 0 2 | N A T U RE | V O L 5 4 9 | 1 4 s e p t e m b e r 2 0 1 7
© 2017 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.