See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/345821742

Machine Learning: Quantum Vs Classical

Article in IEEE Access · November 2020

DOI: 10.1109/ACCESS.2020.3041719

CITATIONS READS

94 8,390

2 authors:

Tariq M. Khan Antonio Robles-Kelly

UNSW Sydney Deakin University
116 PUBLICATIONS 1,951 CITATIONS 253 PUBLICATIONS 3,414 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Tariq M. Khan on 14 November 2020.

The user has requested enhancement of the downloaded file.

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.DOI

Machine Learning: Quantum Vs

Classical
TARIQ M. KHAN1 (Member, IEEE), ANTONIO ROBLES-KELLY1 (Senior Member, IEEE)
1
School of Information Technology, Deakin University, Australia
Corresponding author: Tariq M. Khan (e-mail: [email protected]).

ABSTRACT Encouraged by growing computing power and algorithmic development, machine learning
technologies have became powerful tools for a wide variety of application areas, spanning from agriculture
to chemistry and natural language processing. The use of quantum systems to process classical data using
machine learning algorithms has given rise to an emerging research area, i.e. quantum machine learning.
Despite its origins in the processing of classical data, quantum machine learning also explores the use of
quantum phenomena for learning systems, the use of quantum computers for learning on quantum data and
how machine learning algorithms and software can be formulated and implemented on quantum computers.
Quantum machine learning can have a transformational effect in computer science. It may speed up the
processing of information well beyond the existing classical speeds. Recent work has seen the development
of quantum algorithms that could serve as foundations for machine learning applications. Despite its great
promise, there are still significant hardware and software challenges that need to be resolved before quantum
machine learning becomes practical. In this paper, we present an overview of quantum machine learning in
the light of classical approaches. Departing from foundational concepts of machine learning and quantum
computing, we discusses various technical contributions, strengths and similarities of research works in this
domain. We also elaborate upon recent progress of different quantum machine learning approaches, their
complexity and applications in various fields such as physics, chemistry and natural language processing.

INDEX TERMS Quantum machine learning, quantum computing, quantum algorithms, QuBit

I. INTRODUCTION complex mathematical problems which traditional computers

find hard and, sometimes, unfeasible to solve. Moreover,
Its estimated that, every day, 2.5 exabytes of data are gener-
quantum computing operations are expected to give rise to
ated. This is equal to 250,000 Congress Libraries and reflects
a new family of algorithms in Quantum-enhanced machine
the fact that, every minute, more than 3.2 billion internet
learning. This is since Quantum calculations are based on the
users are feeding the data banks with 347,222 tweets, 4.2
notion that subatomic particles exist at any time in more than
million Facebook likes, 9,722 pins on Pinterest plus the users
one state at a time. This contrasts with traditional computers,
data created by taking pictures and videos, saving documents,
which are binary in nature, i.e. 0 or 1, and can neither be
opening accounts and more. With these huge volumes of
superposed nor intertwined as quantum bits (QuBits) can
data at our disposal, we are able to investigate not only
be. QuBits run in a much richer world, making a quantum
business issues such as financial strategies, but also climate
computer more powerful than any traditional computer.
change factors, complex biological processes or economic
phenomena. This amount of data is ever growing, outpacing
The rapid development in the field of GPUs for parallel
computational power growth.
processing has addressed some of the issues in big data in
Despite Moore’s Law, which assumes that the capacity traditional computation by giving a big boost to deep neural
of integration should double every two years on integrated networks. That said, GPUs also benefit from Moore’s Law.
circuits, has proved to be remarkably resilient since 1965, it Moreover, as the complexities of the problems to be solved
will likely come to an end as transistors on traditional modern have grown, AI stacks have begun to include more and more
computers can will soon not be able to go any smaller. This different hardware back-ends, as shown in Figure 1. These
makes quantum computing attractive. Quantum computers include off-the-shelf CPU’s, Tensor Processing Units (TPUs)
make use of quantum phenomena and can efficiently tackle and neuromorphic chips. Yet, there are still problems that

VOLUME 4, 2016 1
 T. M. khan and A. Robles-Kelly: Machine Learning: Quantum Vs Classical

CLASSICAL COMPUTING QUANTUM

CPU GPU FPGA QPU

Supervised Unsupervised
Discriminative Generative
Deterministic Probabilistic
Parallizable ASIC Difficult on GPU

FIGURE 1: Mix of heterogeneous back-end hardware currently used by AI. In the future, this will also incorporate quantum
technologies to enable certain learning algorithms [1]

remain too difficult to solve and some of these are amenable computationally complex but also costly. It is very easy to
to quantum computing. In other words, quantum computers do such calculations on a quantum computer as underlying
can enable the creation of more complex machine learning physical objects representing QuBits match and arise from
and deep learning models in the same way that GPU’s have quantum phenomena. Despite its promise, existing code and
done in the past. Figure 1 describes the role of quantum algorithms do not often apply to a quantum processor. This is
computing in machine learning. due to the natural difference between QuBits and bits. This is
a problem at the core of QML work, whereby quantum algo-
Quantum machine learning (QML) is the intersection of
rithms or quantum-inspired may be quite different in nature
machine learning and quantum computing. QML attempts to
as compared to their counterparts executed on conventional
use the capacity of quantum computers to process data at
computers or GPUs.
much faster speeds than traditional computers. QML refers
the use of quantum systems to incarnate algorithms that allow
II. LITERATURE REVIEW
computer programs to improve through experience. Thus,
As near-term quantum devices move beyond the point of
in QML, quantum computers are used to tackle machine
classical simulability, also known as quantum supremacy
learning problems making use of the natural efficiency of
[5], it is of utmost importance to discover new applications
quantum computers. Recall that quantum states in superpo-
for Noisy Intermediate Scale Quantum devices [6] which
sition can lead to important speedups as a result of the ability
are expected to be available in the next few years. These
to evaluate multiple states simultaneously. As noted in [2],
are expected to have an architecture like the one shown in
there is no comprehensive theory of QML yet. Sasaki and
Figure 2. Among the most promising applications for near-
Carlini [3] have delved into the notion of semi-classical and
term devices is Quantum Machine Learning (QML) [7]–[10],
universal strategies whereby in the former classical methods
Quantum Simulation (QS) [11]–[13] and Quantum-enhanced
are adapted to work on quantum systems whereas in the
Optimization (QEO) [14]. Recent advances in these three
latter the methods are purely quantum in nature. In [4], the
areas have been dominated by a class of algorithms called
authors explain how learning algorithms may be accelerated
hybrid quantum classical variational algorithms. In these
via quantisation, i.e. the partial or total translation of the
algorithms, a classical computer aids the quantum computer
algorithm to its quantum counterpart. Both [3] and [4] argue
in a search over a parameterized class of quantum circuits.
for the use of quantum primitives.
These parameterized quantum circuits are sometimes called
Morevoer, note that quantum computers may not be a quantum neural networks [7]. Key to the success of quantum
complete replacement for traditional computers. That said, classical algorithms is hybridization. This follows the notion
they will allow us to extend the categories of computer- that, in the near-term, quantum computers will be used as co-
tractable problems. First of all, there are certain conven- processors for classical devices. Here we also explore ways
tional tasks that are ideally suited for quantum computers. to hybridize certain quantum simulation and QML tasks in
Quantum computers will be able to solve these tasks with a way that fully takes advantage of the strengths of both
extremely less computationally complexity. Further, quantum devices. The rise of variational quantum algorithms can be
computers can perform computations and solve problems traced back to the invention and implementation of the Varia-
conventional computers can not do. For example, no con- tional Quantum Eigensolver [12], which sparked a Cambrian
ventional hardware can generate a truly random number. explosion of works in the field of near-term algorithms.
That’s why the generators in conventional computers are re- Many improved quantum enhanced algorithms have been
ferred to as spontaneous pseudo generators. This is, however, proposed to speed up certain machine learning tasks [15] and
an straightforward task for a quantum computer. Likewise, intermediates sizes have now been achieved with existing
when trying to imitate natural processes with conventional quantum processors. For fault-tolerant quantum computing,
bits (e.g. subatomic component behaviour), it is not only the vast majority of quantum enhanced algorithms have been
2 VOLUME 4, 2016
T. M. khan and A. Robles-Kelly: Machine Learning: Quantum Vs Classical

Quantum chipset

Quantum Electronic

Classical Computer

Software and data

Quantum register
gates writing

physical/logical

Measurement Electronic
interface reading

Refrigerated enclosure

FIGURE 2: Global architecture of a quantum computer.

developed. However, near-term quantum devices will not be with these problems are Markov chain Monte Carlo (MCMC)
corrected for error and will thus necessarily encounter high processes. However, even with MCMC, drawing samples
levels of error during operation. This raises the necessary costs of models with multiple neurons can be high. That said,
question as to whether or not quantum devices with pre- quantum computers can draw samples in a natural manner
fault-tolerance-threshold noise levels are useful for industrial due to their capacity to generate true random sequences and,
applications. Notice that the classical-quantum hybrid algo- hence, will reduce training costs [21].
rithms, such as Quantum Eigensolver [12], show that there The RBM training comprises two major quantum tech-
can be a quantum advantage partially robust to noise with niques classes. The first is about quantum linear algebra
some applications in optimisation and quantum chemistry. As and quantum methods of sampling. Wiebe et al. [22] have
for machine-learning applications, it has been demonstrated developed two RBM algorithms, based on the amplification
that annealers can perform some forms of machine learning, amplitude [23] and quantum Gibbs sampling, to effectively
but there is a strong question as to whether the quantum train an RBM. These improve the number of examples
computing model of a near-term circuit can perform similar needed to train the RBM quadratically but their algorithm
tasks. scaling in the number of edges is quadratically lower than
Many quantum-classical variational algorithms consist of that of the contrastive divergence [24]. Another advantage
optimizing the parameters of a parameterized quantum cir- of the approach suggested by [25] is its use for training
cuit to extremize a cost function, which often comprises of full Boltzmann machines. A conventional version of such
the expectation value of a certain observable at the output an algorithm has also been proposed in [22]. A complete
of the circuit. This optimization of parameterized functions Boltzmann machine is a type of Boltzmann machine where
is similar to the methods of classical deep learning with neurons correspond to the complete graph nodes, i.e. when
neural networks [16]–[18]. Furthermore, the training and they are fully connected. While Boltzmann machines have a
inference processes for classical deep neural networks have more comprehensive range of RBM parameters, these ma-
been shown to be “embeddable” into this quantum-classical chines are not practical in many cases due to high computing
Parameterized Quantum Circuit (PQC) optimization frame- costs.
works [19]. Given these connections with deep learning and The second class of RBM training is based on Quantum
neural nets, it has become common to sometimes refer to Annealer (QA), a quantum computing model that encodes
certain PQC ansatz as Quantum Neural Networks (QNN’s) problems in the energy function of the Ising model. Specifi-
[15]. In the other hand, the literature on accelerated neural cally, the authors of [26] use the spin configurations produced
network training using quantum resources focused primarily by the quantum annealer to draw Gibbs samples that can
on restricted Boltzmann machines (RBM). The RBMs are then be used to train the RBM. Such types of physical
generative models that, because of their strong connections implementations of RBMs pose a number of challenges.
with the Ising model, are particularly suitable to be studied Benedetti et al. [27] pointed out that it was difficult to de-
from a quantum perspective. The computation of the log- termine the effective temperature of the physical machine. A
likeness and sampling from an RBM has been shown to be algorithm to calculate the effective temperature and compare
computationally hard [20]. The basic approaches for dealing the output of a physical system with some simple problems
VOLUME 4, 2016 3
 T. M. khan and A. Robles-Kelly: Machine Learning: Quantum Vs Classical

was implemented to overcome this difficulty. Dumoulin et and a better overall response to the global landscape. Good
al. [21] conducted a second critical analysis of quantum initialization is therefore crucial in order to improve the
RBM training. The authors have shown how the limitations convergence of local optimisers to local extremes with rea-
of first-generation quantum machines are likely to seriously sonably good local minima.
restrict the applicability of quantum methods in terms of
noise, synchronisation and parameter tuning. The quantum III. BACKGROUND
Boltzmann machine proposed by Amin et al. [28] is a hybrid Quantum is the smallest unit possible of any physical entity,
approach between ANN training and a fully quantum neural for instance energy or mass. In 1900, Max Planck suggested
network. This model gives the standard RBM energy function that the energy of any body at the atomic and subatomic
a pure quantum term that, according to the authors, makes levels be composed of a discrete packet called quanta [35].
it possible to model a rich class of problems (i.e. problems Recall that wave-particle duality is characteristic of how,
that would otherwise be classically difficult to model, such depending on the environment, matter and light sometimes
as quantum states). It is not known whether these models can behave like a wave and sometime as a particle. In this setting,
give any advantage to classical tasks. Kieferova & Wiebe [29] quantum theory is characterised by the determination of the
suggest that quantum Boltzmann machines could provide ad- probability of a particle being at a given point x in space,
vantages for tasks such as rebuilding a quantum state density not its exact location. This notion is at the center of quantum
matrix from a set of measurements (this operation is known mechanics and will be important in the following sections.
as quantum state tomography in the quantum information In quantum mechanics, vectors are expressed in the bra-
literature). ket notation, i.e. ha| and |ai, whereby a the former is a row
Although there is no consensus on the characteristics of vector and the latter is a column vector. In this notation,
quantum ANN, a number of works have tried in recent for convenience, the inner product is expressed as hb|ai,
decades to build an artificial neural network (ANN) whose where hb| and |ai are bra-ket vectors. Similarly, let be H
updating rules and elements are based on the quantum me- a Hermitian matrix, its eigenvector equation is given by
chanics laws. The Review by Schuld et al. [30] gives a critical Hha| = λa |ai, where ha| is the eigenvector and λa is the
overview of the various strategies for the development of corresponding eigenvalue. Our choice of a Hermitian matrix
quantum artificial neural networks and points out how the here stems from the importance of this sort of matrix in
majority of the approaches are not meeting the requirements quantum mechanics. Recall a Hermitian matrix is defined as
with what can reasonably be defined as quantum ANN. In that which is equal to its transposed complex conjugate. For
fact, Schuld et al. [30] did not reproduce the fundamental each Hermitian matrix, there are two separate eigenvectors
characteristics of the ANN (for example, Hopfield’s attractive each of which will give rise to a separate eigenvalue. In the
dynamics). On the other hand, it can be argued that the physical world, Hermitian matrices represent a measurable
single biggest challenge for a quantum ANN is that quantum observable. Eigenvectors represents the possible state of the
mechanics are linear, but artificial neural networks require system. Eigenvalues then represent the actual results we get
nonlinearity [31]. when we undertake the measurements. One thing that is
The optimization of quantum neural networks in the NISQ notable here is that, in many cases, the Hermitian operator
(Noisy Intermediate Scale Quantum) era is currently facing cannot be directly measured. This also applies to the eigen-
two major challenges. The first is local optimization. This vector. The eigenvalue in contrast is the measurable, whereby
is, the stochastic nature of the objective function, combined the result of the measurement can determine the eigenvector
with readout complexity considerations, has made the direct was and, therefore, indicates the state of the system.
translation of classical local optimization algorithms chal-
lenging. Gradient-based optimisers either rely on quantum A. QUBIT
back-spreading of errors [19] requiring additional gate depth
The QuBit is a complex vector pair that points to a point on a
and quantum memory, or use finite-difference gradients [32]
unit sphere. We have illustrated this in Figure 4. A QuBit can
that typically require multiple quantum circuit evaluations for
also be viewed as a zeros and ones in a fuzzy combination.
each gradient descent iteration. Recent work has proposed
Quantum bits could in theory consist of anything acting like
sampling of analytical gradients [33] to reduce this cost.
an electron or atom in a quantum manner. Thus, QuBits can
However, these approaches also require a large number of
be represented by:
measurement runs and therefore remain expensive.
The second major challenge for optimization of quantum • An electron orbiting a nucleus: where |0i and |1i are the
neural networks regards the initialization of their param- ground state and excited state, respectively
eter space. Although some proposals have been made for • A photon: where |0i and |1i are its polarization states.
the QNN parameter initialization heuristics [34], there is In practice, QuBits in the computing industry are tiny
still a need for more efficient and flexible versions of such superconductive circuits that are made from nano-technology
heuristics. By initialising cost landscape parameters in the that can run in two directions. The nature of QuBits makes
neighbourhood of a local minimum, one can ensure more them ideal for multi-tasking, this is one of the reason why
consistent local optimization convergence in fewer iterations computing power could explode exponentially if computa-
4 VOLUME 4, 2016
T. M. khan and A. Robles-Kelly: Machine Learning: Quantum Vs Classical

Classical data Quantum data

1
0 off QuBit
1 on 0
Suspension of states

Quantum gates are basic building blocks for

Logic gate are basic building blocks manipulating QuBits, e.g. Hadamard gate
1
1
& 1
|0 > H |0 > +|1 >

Logic gate are made using transistors

Quantum gates are represented by
+5V 1 unitary matrices
1 1

A microprocessor is a package of several

The richness of the values of the qubit
transaistors linked between them to process
is manifested only during the calculation
the information

Bits: 0 or 1 Qubits: 0 and 1

State Two possible exclusive states Two simultaneous possible states

Initialization 0 or 1 0

Representation 0 or 1 vector of two dimension

Dimensionality 1 binary digit two floating numbers

Modification logical gates quantum gates

Read 0 or 1, deterministic 0 or 1, probabilistic

FIGURE 3: Comparison between classical and Quantum data.

tions are based upon QuBits rather than just bits. Figure 3 |0iand |1i. If a QuBit is measured1 , the QuBit collapses to
compares classical data to quantum data. one of its eigenstates and the measured value reflects that
QuBits are the basic unit of quantum information. They particular state. For instance, a measurement will cause a
are, in effect, the quantum incarnation of the classic binary QuBit to collapse into one of its two basic states |0i and |1i in
bit with a two-state device. A QuBit is a quantum mechanical a superposition of equal weights with an equal probability of
device of two different levels. The key distinctions between 50%. Here, |0i is the state that when measured, and therefore
a bit and a QuBit is shown in Figure 3. Moreover, the mathe- collapsed, will always give the result 0. Similarly |1i will
matical model for QuBits is related to the representation, by always convert to 1. Also note that, fundamentally, a quantum
a two-dimensional vector, of the state of a quantum object superposition is different from superposing classical waves.
over two states, which always has a norm of unit length. This A quantum computer that consists of n QuBits can exist in
function has two numbers α and β, whereby one of these is 2n superposition states, i.e. from |000...0i to |111...1i. On
real, i.e. α, and the second one is complex, i.e. β. the other hand, playing n musical sounds with all different
frequencies can only give a superposition of n frequencies.
The key concept that can be recalled from this mathe- Thus, adding classical wave scales is linear, where the super-
matical representation is that a QuBit can be written as a position of quantum states is exponential.
superposition of two states at any time, independently of the Quantum entanglement is a phenomenon in which quan-
initialization and moment of reading, as: tum particles interact with each other and are represented by
|ψi = α |0i + β |1i (1) reference to one another, even though they are separated by a
1 To be more precise, it is possible only to measure observables and, hence,
QuBits can be in a superposition of both the basis states we assume this is the case in this particular expample

VOLUME 4, 2016 5
 T. M. khan and A. Robles-Kelly: Machine Learning: Quantum Vs Classical

Bit Probabilistic Bit Angles defining the

Qubit
vectors of the
intermediary state σz |0 > Ground state of the Qubit
0 0
|ψ > Intermediary state

ϕ σy

σx

1 1
|1 > Excited state of the Qubit

Probability Probability
State of of the 0 of the 1 The relationship between α and β accroding to
the qubit state state
the Max Born rule, related to the Schrodinger
wave function that defines the state |0 > and |1 >

 
|ψ >= α|0 > +β|1 > α cos θ
2


|ψ >=  

2 2
β exp (iϕ) sin θ
2
|α| + |β| = 1
FIGURE 4: QuBits and their mathematical representation

large distance. At the time of measurement, if one entangled A quantum register of n size is simply a n QuBits array. A
particle in a pair is decided to be in the spin state of ‘down’, detailed overview of the n QuBits register compared to a n
this decision shall be immediately transmitted to the other bits register is given in Figure 5.
coupled particle which now assumes the opposite spin state We do use logical gates in quantum computing, but, again,
of ‘up’. Quantum entanglement allows QuBits, even those logical gates have different purposes to those used in quan-
that are faraway, to interact instantaneously with each other. tum computing. Quantum logic gates which are essentially
Consider a system of two electrons, A and B, which are a transformation of one or more QuBits. We have given a
entangled. Assume something interacts with just one of these description of the quantum gate types in [20] in Figure 6.
electrons in the system, not both, and that both are quite far Further, in quantum computing, a quantum algorithm is
apart spatially from each other. Also recall that, if we have an based on a practical model of quantum computation. A
entangled system, then the probability distribution of spins classical algorithm, i.e. non-quantum, is a finite sequence
whether that is up or down for B will change after the spin of of instructions or a step-by-step problem solving procedure,
particle A is measured. Therefore, if something interacts with where a classic computer will execute any step or instruction
just one part of our system, for example particle A, then the that can be defined using binary operations. In the same way,
whole system is affected. More specifically, the probability a quantum algorithm is a step-by-step procedure where each
distribution of an up or down spin for the other particle in step can be performed on a quantum computer. Although it is
this particular case will be unduly affected. So in general, also possible to do all classic algorithms on a quantum com-
the interaction of something external with one part of our puter [36], the term quantum algorithm is usually used for
entangled system causes the entire system to be affected. In those algorithms that are inherently based on quantum object
conclusion, quantum entanglement enables particles to affect properties. These are expected to use some of the key features
each other instantaneously across any distance. Entangled of quantum computation, such as quantum superposition or
particles would remain “connected” even if they were on quantum entanglement.
opposite sides of the universe. This example is relevant since Figure 7 shows the four big classes of quantum algorithms
it opens up the possibility to employ the properties of wave- that are commonly used in quantum computing. These algo-
particle duality to make possible, in certain cases, to interact rithms are primarily based upon quantum annealing, simula-
with QuBits or to interact with QuBits by interference in tion or the versatility and properties for universal computers.
quantum algorithms. This is since quantum annealing and simulation often employ
the notion that the quantum system should evolve naturally in
B. QUANTUM REGISTERS AND GATES time, whereby quantum universal computers are expected to
In quantum computing, a quantum register is a system be able to control the evolution process itself. Each category
comprising multiple QuBits. A quantum computer stores its also varies by the amount of processing power (QuBits)
quantum data in one or more quantum registers. Note that needed and the number of potential applications, as well as
quantum and classical registers are conceptually different. An the time necessary to become commercially viable. In the
array of n flip-flops is referred to in the n classic register. following sections, we examine each of these.
6 VOLUME 4, 2016
T. M. khan and A. Robles-Kelly: Machine Learning: Quantum Vs Classical

Classical Registers Quantum Registers

Register of n bits Register of n qubits

n n
2 possible states (one at a time) 2 possible states (simultaneously)
Evaluable Partially evaluable
Independent copies Incopiables independently
Individually erasable Indelible individually
Lecture non destructive Lecture modify the present value
Deterministic Probabilistic

Register of 4 Bits
0 0 0 0
0 0 0 1 Register of 4 QuBits
0 0 1 0
0 0 1 1 |0000>+|0001>+|0010>+|0011>
0 1 0 0 +
0 1 0 1
0 1 1 0 |0100>+|0101>+|0110>+|0111>
0 1 1 1 +
1 0 0 0 |1000>+|1001>+|1010>+|1011>
1 0 0 1 +
1 0 1 0
|1100>+|1101>+|1110>+|1111>
1 0 1 1
1 1 0 0
1 1 0 1
1 1 1 0
1 1 1 1
This register could represent 16 states that are in a
superposition → we have 16 state in the same time

This register could represent 16 states

FIGURE 5: Quantum register vs classical register

C. QUANTUM COMPUTING minimum energy states correspond to local minima. Quan-

1) Quantum Annealing tum annealing, therefore, uses quantum physics to find the
minimum energy state. Sampling problems are related to
Quantum Annealing is essentially a way to use quantum’s
optimization, but are slightly different. Instead of focusing
intrinsic effects to find the best solution to problems involv-
on trying to recover the minimum energy state, in sampling
ing a wide range of solutions. The type of problems it solves
problems, we aim at characterising the energy landscape.
are either related to optimization or probabilistic sampling.
This is useful for applications such as machine learning,
Optimization is a problem where we try to find the best
where the aim to learn and predict the state of a given system
configuration out of many different possible combinations.
based upon known instances. These samples provide us with
It solves these types of problems by exploiting properties
knowledge about what the model is and how it evolves over
specific to quantum physics such as quantum tunnelling,
time. Thus, quantum annealing aims at harnessing the natural
entanglement and superposition. Each state may be repre-
evaluation of quantum states. The main issue here resides in
sented as an energy level in the quantum annealers. These
the fact that the evolution of the energy landscape is often
states are simulated in a short time, taking advantage of the
related to the quantum object itself. That is, the problem
superposition and entanglement properties of the QuBits and
should be set up at the begining so as to allow the quantum
achieving the lowest energy result. The lowest state of energy
physics to evolve naturally, where the configuration at the end
provides the optimal solution, which, in quantum annealing
of the process should correspond to the aim of computation.
corresponds to the most likely one.
The reason why physics can be used to solve optimization Quantum annealing starts at the lowest energy eigenstate
problems is based on the key element of physics, whereby of the initial Hamiltonian. When we anneal, we are intro-
VOLUME 4, 2016 7
 T. M. khan and A. Robles-Kelly: Machine Learning: Quantum Vs Classical

Quantum gate Input Classical gate Circuit representation Matrix Short discription

1 qubit None 1
 1 1

Hadamard gate Create superposition
H √
2 1 −1

 0 1

Pauli-X gate 1 qubit Not gate Create x rotation
X 1 0

 0 −i

Pauli-Y gate None Create y rotation
1 qubit Y i 0

 1 0
 Create z rotation
Pauli-Z gate None (special case of
1 qubit Z 0 −1
phase shift gate)
 1 0 0 0

Not gate 0 1 0 0 Used to measure
Controlled not gate 2 or more qubits 0 0 0 1 2nd qubit
0 0 1 0

FIGURE 6: Examples of quantum gates.

ducing the Hamiltonian problem, which involves biases and as quantum chemistry. The basic foundation of quantum
couplers, and reducing the effect of the initial Hamiltonian as chemistry simulation has been set by the universal quantum
it evolves. At the end of the annealing process, the final state method [37] and the first quantum algorithm for simulating
will correspond to an eigenstate of the Hamiltonian problem. fermions [38]. Quantum chemistry simulation applications
Ideally, during the quantum annealing, the evolution will have been actively evolving. Quantum chemistry simulation
correspond to a geodesic of minimal energy throughout the is considered a true application in NISQ computers [39], [40].
energy landscape. Further, every QuBit is a classic object at Quantum computers can help calculate a large number of
the end of the annealing process. potential protein folding sequences to make more efficient
medications. Quantum simulations will allow rapid drug
2) Quantum Simulations designer testing in the future by accounting for any possible
Quantum simulations explore particular problems in quan- protein-to-drug combination. The main processing task in
tum mechanics beyond the capabilities of classical systems. quantum simulation is typically the time evolution of the
Simulation of complex quantum phenomena can be one of quantum system under the Hamiltonian,
the most important applications for quantum computing. One |Ψ(t)i = exp(iĤt)|Ψ(0)i (2)
highly exciting area is the simulation of the effect of chemical
stimulation on a large number of subatomic particles known where Ψ(0) is the initial state and Ĥ, which is time dependent
used to calculate Ψ(t) at time t. In certain cases, it is the
properties of a system governed by a particular Hamiltonian
that are sought, and pure quantum evolution is sufficient.
In others, quantum simulation and annealing are required to
reach the desired solution.

3) Universal Quantum Computing

Universal quantum computers are the most powerful and
the ones that most profit from quantum phenomena, they
are also the most difficult to build. A universal quantum
computing system relies on building up robust QuBits where
simple quantum circuit operations, similar to the classical
operations we all know, can be put together to construct
any sequence of instructions, running increasingly complex
algorithms. A genuinely universal quantum computer will
certainly use more than 100,000 QuBits - some figures place
it at 1 M QuBits. Note that, the maximum number of QuBits
we can integrate today is not even 128. The basic concept
behind the universal quantum computer is that it can lead to
massively complex computations with very fast solutions. It
includes solving the annealing equations, simulating quan-
FIGURE 7: Quantum algorithms. tum phenomena, and more. In universal quantum computing
8 VOLUME 4, 2016
T. M. khan and A. Robles-Kelly: Machine Learning: Quantum Vs Classical

UNIVERSAL
chine learning [46] since these methods are often designed
QUANTUM COMPUTING Measurement to run on conventional machines. Quantum-inspired machine
Based
Shor’s Algorithm learning is, for example, focused on quantum processing to
Grover’s Algorithm generate new classical learning models and new ways to train
and evaluate them. Examples here are ideas involving tensor
Gate Model Adiabatic networks as learning models [47], the approaches of Tang
Polynomial
Shor’s Algorithm Mapping Shor’s Algorithm for fast stochastic linear-algebraic manipulations [48], and
Grover’s Algorithm Grover’s Algorithm
inspirations behind the projective simulation of Briegel [49].
Quantum-generalized machine learning generalizes the
Topological fundamental concepts like density matrices generalize clas-
Shor’s Algorithm sical notions of information. Quantum machine learning
Grover’s Algorithm
questions what machine learning can look like if data or
environments really are quantum objects. This is particularly
FIGURE 8: Universal quantum computing. useful in applications related to chemistry and biology. In
Table 2, we list existing work on quantum machine learn-
ing for physics, chemistry and natural language processing.
the aim is a lot more ambitious than just allowing the system Moreover, quantum weirdness might help us to understand
to evolve naturaly or simulate the quantum phenomena. What the strange features of nature itself. Quantum systems are
we’re trying to do is to control and manipulate the evolution popular for creating odd and counter-intuitive patterns that
of that quantum state over time. This is much more com- cannot be generated by any classical machine, even a clas-
plex than quantum annealing since quantum systems tend sic computer. Besides being able to produce unusual and
to be prone to noise, faults and loss of quantum coherence. counter-intuitive patterns, quantum computers would also
However, having this amount of control means a much larger be able to recognise patterns that no classic computer can
class of problems can be tackled. Many quantum algorithms, recognize.
such as Shor’s [41] to crack RSA cryptography and Grover’s Machine learning algorithms can broadly classified into
faster search [42], [43], can also run on a universal quantum supervised learning, semisupervised learning, unsupervised
computer. These algorithms are way more computationally learning and reinforcement learning. In Figure 9, we show
efficient than anything we can possibly run on a classical a simplified panoramic view of ML and the applications.
computers. It is important to note that, due to the nature of supervised learning algorithms involve direct operational
quantum states, in a universal quantum computer there is a supervision. In this case, the sample data are labelled and
polynomial time and resource mapping from the classical the algorithm operates at strict boundaries. The primary aim
gate model of these other approaches to a quantum computer. of supervised learning is to scale the data scope and to
We illustrate this in Figure 8 for Shor’s [44] and Grover’s [45] predict data that are unavailable, based on the labelled sample
algorithms. data. Supervised machine learning are used for mainly clas-
sification and regression tasks. Classification is the process
IV. QUANTUM MACHINE LEARNING where input data are labeled based on past data samples and
A. CONTEXT the algorithm is trained to identify and categorise certain
Quantum Machine Learning (Quantum ML) is an interdis- object types. Regression is the mechanism by which patterns
ciplinary field incorporating Quantum Physics and Machine are identified and the effects of continued predictions are
Learning (ML). This is a symbiotic association using the calculated.
power of Quantum Computing to generate quantum versions Moreover, note that unsupervised learning is a type of
of ML algorithms and applying classical ML algorithms to learning that is based upon unlabeled data. Semisupervised
analyse quantum systems. More comprehensively, quantum learning employs a mix of supervision with unlabelled data.
machine learning is a field where the power of quantum In unsupervised learning, the desired results are not known
computing and the properties of quantum physics are used a priori rather they should be inferred based upon predefined
for machine learning tasks and applied accordingly to related criteria. It does not require any labelled data. Unsupervised
fields. The two main types of quantum machine learning machine learning is used mainly for two purposes: clustering
are the quantum-applied machine learning and quantum- and dimensionality reduction. Clustering is an exploration
enhanced machine learning. Further, quantum-inspired ma- of the data used to segment it into meaningful groups, i.e.
chine learning and quantum-generalized learning ideas stand clusters, based on their internal affinity, without prior knowl-
out as two very promising lines of research in line with the edge of group memberships. Reinforced learning, in the other
QML philosophy. In Table 1 we provide a brief survey of hand, reflects the agent’s ability to obtain the result that
books and review articles for the benefit of the interested yields the maximum cumulative reward. The agent shall be
reader. rewarded or punished for a correct or incorrect answer on
Note that the speed-ups that can be achieved using quan- the basis of a reinforcement function. Reinforcement learning
tum computers are not achievable by quantum-inspired ma- differs from supervised learning in the sense that training data
VOLUME 4, 2016 9
 T. M. khan and A. Robles-Kelly: Machine Learning: Quantum Vs Classical

TABLE 1: Books and surveys on quantum machine learning

Type Title Year Reference
Quantum Machine Learning: What Quantum Comput- 2016 [50]
ing Means to Data Mining
Books Supervised Learning with Quantum Computers 2018 [51]
Principles of Quantum Artificial Intelligence 2013 [52]
Quantum Robotics: A Primer on Current Science and 2017 [53]
Future Perspectives
An introduction to Quantum Machine Learning 2014 [54]
The quest for a Quantum Neural Network 2014 [30]
Quantum machine learning 2017 [15]
Quantum machine learning: a classical perspective 2018 [55]
A survey of quantum learning theory 2017 [56]
Surveys and review articles Machine learning & artificial intelligence in the quan- 2018 [57]
tum domain: a review of recent progress
Parameterized quantum circuits as machine learning 2019 [58]
models
Quantum Deep Learning Neural Networks 2019 [59]
Machine learning and the physical sciences 2019 [60]
Special topic on data-enabled theoretical chemistry 2018 [61]
Learning in quantum control: High-dimensional global 2017 [62]
optimization for noisy quantum dynamics.

does not have a sense of reward. In a similar manner, agents Further, recall that two classical bits can take one of four
are not trained to achieve a result consistent with labeled data states: 00 or 01 or 10 or 11. Each of these, the first bit and
rather to learn actions so as to maximise a measure of reward. the second bit, combine to represent only one binary config-
Reinforced machine learning is mainly used for control and uration at a given time in a classical computer, representing a
classification. single binary configuration. However, one QuBit can exist si-
In Table 3 we have summarised supervised, unsupervised multaneously. Thus, all 4 binary configurations can be saved
and reinforcement quantum machine learning approaches at a time by two interacting QuBits. In general, ‘n’ QuBits
elsewhere in the literature. In the table, we also show the can represent classical binary settings in 2n simultaneously.
year of publication for each of the references under consid- Thus, a 300-QuBit quantum computer can simultaneously
eration. Note that a number of these quantum algorithms can explore 2300 possible solutions. This implies immense par-
significantly reduce the computational capacity of learning allelism, unlike any conventional computer, whereby the
algorithms in classical computers. In Figure 10 we show a addition of more QuBits to a quantum computer will increase
graphic view of the prospective applications that are expected the computer’s power exponentially.
to benefit from quantum machine learning.
Moreover, several of these quantum algorithms are based B. COMPLEXITY
on the mapping of a classical data vector to a quantum Computational complexity theory concerns with both the
state. Figure 11 shows the processing schemes for classical general and problem-specific scalability and computational
machine learning (CML) and quantum machine learning cost of algorithms. The word “scalability” is, in short, the
(QML). In the figure, an example of a datas tream of N bits cost in time and/or space needed to increase the volume
with N=3 is given for both CML and QML. For CML, data or complexity of the aim of computation. Making use of
is directly input to the algorithm, where CML process it and the Big-O notation, an algorithm that is O(n3 ) is said to
delivers an output. QML, in the other hand, first requires the be “harder" than one that is O(n2 ) because the former will
stream to be transformed into Quantum Data (QD). QML generally require more operations to be effected than the
takes this QD at input, processes it and produces an output latter, regardless of the speed at which these operations are
in QD form. Then this QD is transformed to classical data. performed. A problem that can be solved in polynomial time
Compared to classical machine learning, this process implies is said to be solvable if an algorithm is available that could
a complex encoding from classical data to QD so as to allow expressed in a O(np ) format. Otherwise, the problem is
QML to operate. This requires building quantum interface assumed to be non-polynomial.
devices that enable classical information to be encoded into a A given complexity class consists of problems, all of
quantum mechanical form. Difficulties such as the problems which have similar characteristics with regard to their hard-
of “input” or “output” then become major technical chal- ness. By far, the two most important classes of complexity are
lenges that need to be addressed. P and NP. P is defined as the class of problems for which there
10 VOLUME 4, 2016
T. M. khan and A. Robles-Kelly: Machine Learning: Quantum Vs Classical

Image
Structure classification
Meaningful discovery
compression
Frud
detection
Diagnostic
Big data
visualization
DIMENSIONALITY
CLASSIFICATION
REDUCTION

Recommentation Weather
forcasting

Customer
segmentation
CLUSTRING Market
UNSUPERVISED SUPERVISED REGRESSION forcasting
LEARNING LEARNING

Targetted
marketing
Estimating life
expectancy
MACHINE
Lane-finding LEARNING Traffic light
on GPS data control

Structure
detection CLUSTRING SEMISUPERVISED REINFORCEMENT
CONTROL Robot
LEARNING LEARNING navigation

Medical
prediction

Real-time
decisions
Text CLASSIFICATION
classification CLASSIFICATION

Skill
Aquisition Bidding and
Speech Advertising
Analysis
Optimized
Internet marketing
sequence
classification

FIGURE 9: A panoramic view of Machine Learning and its applications.

is an efficient, i.e. polynomial time, deterministic algorithm. by a polynomial-time algorithm, the probability of which
NP is known as the class of problems for which an effi- is limited by half over all instances. A quantum computer
cient deterministic certification algorithm exists regardless is said to solve a problem if, for each instance, it has a
of the difficulty that solving the problem entails. The main high probability of answering correctly. If the solution is
issue with NP problems concerns the nature of the solution, running in polynomial time, the problem is in BQP. Fig
whereby it is indeed a valid one to the problem in polynomial 12 summarises the assumption held by the most complex
time [88]. A problem is said to belong to the NP-Complete theorists regarding the relationship between P, BQP and NP-
class if every NP problem can be reduced to a canonical complete class of problems.
form that can be solved in polynomial time. Therefore, NP- Quantum algorithms are also faster than their classical
Complete problems are of special interest to the research counterparts since they are capable of maintaining a super-
community since an optimal solution to this canonical form position of all the states of a given system, and then pick
can be used to solve all NP problem of the same family a particular state from a list with only one operation. On a
efficiently. classical machine, to perform the same function, O(n) op-
The class of problems that can be effectively solved by erations are required. This advantage is exploited by Grover
quantum computers is called BQP for “bound error, quan- [45] to reduce the time spent searching an unsorted database
tum, polynomial time”. Recall that quantum computers run from O(n) to O(n(1/2) . However, how BQP problems related
algorithms that are probabilistic in nature. Thus, BQP on to NP-Complete ones or whether there is a well defined or
quantum computers is the equivalent of BPP (‘bound er- unique relationship remains an open question.
ror, probabilistic, polynomial time’) on classical computers. To better illustrate the complexity advantages of quan-
These are defined as a set of problems that can be resolved tum computing over conventional computation, consider six
VOLUME 4, 2016 11
 T. M. khan and A. Robles-Kelly: Machine Learning: Quantum Vs Classical

TABLE 2: Quantum machine learning for physics, chemistry and natural language processing
Type Title Year Reference
Detecting quantum speedup by quantum walk with 2019 [63]
convolutional neural networks
Machine learning for long-distance quantum commu- 2019 [64]
Quantum machine learning in Physics nication
Discovering physical concepts with neural networks 2018 [65]
Automated discovery of characteristic features of 2019 [66]
phase transitions in many-body localization
Convex optimization of programmable quantum 2019 [67]
computers
Machine learning and the physical sciences 2019 [60]
Quantum Chemistry in the Age of Machine Learning 2020 [68]
Unifying machine learning and quantum chemistry 2019 [69]
Quantum machine learning in Chemistry with a deep neural network for molecular wave func-
tions
ANI-1: an extensible neural network potential with 2017 [70]
DFT accuracy at force field computational cost
Quantum Machine Learning in Chemical Compound 2017 [71]
Space
Quantum-chemical insights from deep tensor neural 2017 [72]
networks
Learning Hidden Quantum Markov Models 2018 [73]
Word Vectors and Quantum Logic: Experiments with 2003 [74]
Natural Language Processing negation and disjunction
Quantum Algorithms for Compositional Natural 2016 [75]
Language Processing
Quantum Language Processing 2019 [76]

complex space since 26 = 64. On a quantum computer,

this computation in such a state space would only take four
iterations. If we evaluate the same state space using classical
computing, it would take thousands of flops. Since the classic
machine can be in a single state at a time, a non-quantum
machine must follow a branch-down process along a tree. On
the other hand, quantum computers process an entire level of
the tree at each step, therefore, to compute the 26 states only
takes 6 branch computations.

Moreover, when a quantum computer is in superposition,

it is effectively processing a set of states in parallel. After
processing these, it will only return one state which is de-
termined probabilistically based on its final superposition.
This means that we may have to run the computation a few
times to gain the confidence required. Despite this, it would
still be way less computationally intensive than trying to
compute the state space using a classical computer. When
the problem is scaled up, it takes exponentially less time
on a quantum computer as compared to classical computing
methods. Probably the most famous example of this is Shor’s
FIGURE 10: Quantum machine learning applications algorithm [41]. Recall that the quantum Fourier transform
(QFT) is used in Shor’s algorithm. The QFT can compute
the prime factorization of an n-bit integer in a complexity
QuBit quantum computation |010001i. This computation is, O(n3 ) whereby the best known classical algorithm requires
1/3
in general, one that corresponds to a sphere in 64 dimensional 2O(n ) time. This algorithm is important for quantum
12 VOLUME 4, 2016
T. M. khan and A. Robles-Kelly: Machine Learning: Quantum Vs Classical

TABLE 3: Quantum machine learning approaches grouped based upon their use of supervised, unsupervised and reinforcement
learning strategies
Type Title Year Reference
Continuous-variable quantum neural networks 2018 [77]
Bayesian Deep Learning on a Quantum Computer 2019 [78]
Quantum algorithms for feedforward neural net- 2018 [79]
works
Supervised learning Quantum Convolutional Neural Networks 2018 [80]
Sublinear quantum algorithms for training linear and 2019 [81]
kernel-based classifiers
A Universal Training Algorithm for Quantum Deep 2018 [19]
Learning
Classification with Quantum Neural Networks on 2018 [32]
Near Term Processors
Barren plateaus in quantum neural network training 2018 [7]
landscapes
Quantum classification of the MNIST dataset via 2018 [82]
Slow Feature Analysis
A Derivative-free Method for Quantum Perceptron 2020 [18]
Training in Multi-layered Neural Networks
Quantum Enhanced Inference in Markov Logic Net- 2017 [83]
Unsupervised learning works
Unsupervised classification of quantum data 2019 [84]
Quantum Algorithms for Solving Dynamic Program- 2019 [85]
Reinforcement learning ming Problems
Quantum gradient estimation and its application to 2019 [86]
quantum reinforcement learning
Reinforcement learning with neural networks for 2018 [87]
quantum feedback

N=3 3 QUBITS
3 BITS 2N = 8 POSSIBLE STATES
Classical
Data Quantum Data
CD tranform to QD

CML
QML

QD tranform to CD

FIGURE 11: Processing schemes of classical machine learning and quantum machine learning

VOLUME 4, 2016 13
 T. M. khan and A. Robles-Kelly: Machine Learning: Quantum Vs Classical

If A is sparse, i.e. has no more than S << n non-null entries

PSPACE per row, then the compexity can be reduced to O(n log(n)).
This kind of sparse problem arises when, for instance, we
have a discrete version of some partial differential equation.
NP-Complete Now lets examine the quantum version of this equation.
The linear equation above can be written using bra-ket no-
tation as A |xi = |bi where |xi = A−1 |bi. This is can be
NP viewed as a physical process where we can start with |bi and

by quantum computers
Complexity

Efficently solved
BQP then evolve it dynamically in a way that is governed by A. For
instance, if A is Hermitian (if A = At ) then we can think of
A as been the Hamiltonian for a a particular system realised
t
by the equation. In this way, we can consider e−iA |bi so as
P to construct |xi.
Viewed in this manner, the problem becomes that of
finding |xi as an n dimensional problem where then |xi =
n logN N
2 and |bi = ntlog2 can be represented in QuBits. If A
Efficently solved is sparse, then e −iA
|bi takes O(logN
2 ) steps of computation
by claissical computers
in a quantum computer. In the end, we can compute the
FIGURE 12: Suspected (though unproven) relationship be- expected value of hx| M |xi to as to read the desired output,
tween P, NP, and BQP [89]. where M is some operator. To see how this is effected, first,
lets assume A = At and write
    
computing both practically and historically. It was the first 0 A 0 ~b
=
polynomial-time quantum algorithm proposed for a difficult At 0 ~x 0
problem on traditional computers with a superpolynomial
quantum speedup [90]. In Table 4, we show the speedups Then we can solve A~x = ~b for any A, where A is sparse
t

for a number of quantum√approaches with respect to their and |xi = e−iA |bi. Moreover, if A is diagonalisable, we
classical counterparts, O( n) implies a quadratic speedup have  
λ1 0
and O(log(n)) involves an exponential gain as compared to uAu−t =
0 λ−1N
their classical counterparts whereas in Figure 13 we compare
the complexity of commonly used computational operations  
−t t λ1 0
for both, classical and quantum computing. A =u u
0 λ−1N

1) Linear Equations To solve this equation, we use a quantum phase algorithm

Lets take an example of quantum algorithm for solving a N|ψj i and eigenvectors |λj i of A. Fur-
and find the eigenvalues
linear equation A~x = ~b having N number of variables. ther, lets assume e−iA ρ , where ρ is a momentum operator.
Where, A is known, ~b is known and ~x is unknown. This is This operator can be veiwed as controlling the influence of
a standard problem which is often expressed as finding ~x = the eigenvalues of A in the decay of the exponential function.
A−1~b. This arises very frequently as a sub-routine for larger After applying this operator on |bi |0i we get
X
computational problems involving linear programming. In
N
e−iA ρ |bi |0i = Bj |ψj i |λj i (3)
this case, classical computing algorithms have a complexity j
of O(n3 ), which is achieved through Gaussian elimination. X −1
= Bj ei∆λj |ψj i |0i
j
TABLE 4: Speedups for a number of quantum algorithms √ in i∆λ−1
the literature, O(log(n)) means exponential whereas O( n) = e j |bi |0i
means quadratic speedup relative to their classical counter-
part = (1 + i∆A−1 ) |bi |0i
Method Speedup and, finally, we get A−1 |bi
√
Bayseian Inference [91], [92] O( n) Note the number of steps, i.e. the complexity, to solve this
Least squares fitting [93] O(log(n))
√ problem are  2 
Online perception [94] O( n) S n
O k log2 ,
Quantum BM [28], [29] O(log(n))
√ ε
Classical BM [25] O( n) where
Quantum SVM [95] O(log(n)) λmax
√ k= .
Quantum PCA [96] O(√n) λmin
Quantum reinforcement learning [97] O( n) This contrasts with the classical algorithm, which has a
14 VOLUME 4, 2016
T. M. khan and A. Robles-Kelly: Machine Learning: Quantum Vs Classical

Classical Computing Quantum Computing

Subroutine Complexity Subroutine Complexity

Matrix inversion: Matrix inversion:

O(N × log(N )) → O(N 2 )∗ O((log(N ))2
Ax = b → x = A−1 b A|x >= |b >→ |x >= A−1 |b >

Eigenvectors and eigenvalues

O(N 2 ) Q Phase O((log(N ))2
of sparse/low-rank matrices

FFT: Fast Fourier QFT: Quantum Fourier

O(N × log(N )) O((log(N ))2
Transform Transform

FIGURE 13: Computing complexity of Quantum vs classical computers.

complexity of   arrive to
kS N −1
! N −1
O n log n X X
− log ε QF T xk |ri = xr QF T (|ri) (8)
k=0 r=0
N −1 N −1
!
2) Quantum Fourier Transform X 1 X i2πkr/N
= xr e |ki
We can consider yet another example. In a similar fashion r=0
N
k=0
as before [36], [98], [99], we can arrive to the the quantum N −1
X
version of the discrete Fourier Transform. The first step is = yj |ji
to do this consists in finding the computational basis of r=0
the Fourier Transform that spans a quantum superposition.
Let us therefore assume that we use n-QuBits representing To obtain an objective comparison of the Quantum Fourier
N = 2n states and that we conveniently have denoted this Transform’s algorithmic advantages over its classical coun-
computational basis as: terpart, we need to examine the circuit and algorithmic
complexity for both. The circuit representation of the QFT
is shown in Figure 14. When a QuBit comes in, as repre-
|0i , |1i , . . . , |N − 1i (4) sented by the ket-vector in the figure, a Hadammard gate H
is applied followed by a set of phase transformations Rk ,
Now, the computational basis of the Quantum Fourier
k = {1, 2, . . . , n} [100]. These phase transformations are
Transforms can be defined in a similar manner to that of
given by " #
the discrete Fourier Transform. Thus we denote this quantum
1 0
basis as Rk ≡ k (9)
N −1
1 X j2πjk/N 0 e2πi/2
QF T |ki = e |ji (5)
N j=0
The result is then passed on to a final Hadamard gate, as
for which a general quantum superposition is given by depicted in the bottom-most lane of the circuit in the figure
for |jn i. This allows for the circuit itself to be quite efficient.
N
X −1 N
X −1
This can be viewed more clearly by calculating the number
xk |ki → yj |ji (6)
of gates for the circuit and the complexity of the operation
k=0 j=0
itself. The number of gates needed for this quantum circuit
where xk is, as usual in quantum mechanics, a position are
variable on the complex domain and yk is defined as follows
• For the 1st QuBit, 1 Hadamard gate and (n-1)Rk gates
N
X −1
1 are needed, where Rk , k = {1, 2, . . . , n} are phase
yj = ei2πr/N xr (7)
N transformations as defined above.
r=0
• For the 2nd QuBit, 1 Hadamard gate and (n-2)Rk gates
Making use of the equations above, its straightforward to are needed.
VOLUME 4, 2016 15
 T. M. khan and A. Robles-Kelly: Machine Learning: Quantum Vs Classical

|j1 i |0i + 2i2π0.j1 j2 ...jn |1i

H R2 Rn

|j2 i
H Rn−1
|0i + 2i2π0.j2 j3 ...jn |1i

|0i + 2i2π0.jn−1 jn |1i

|jn−1 i
H R2

|jn i
H

|0i + 2i2π0.jn |1i

FIGURE 14: Quantum circuit of Quantum Fourier Transform on the state |j1 . . . jn i

• ... This is done by applying the Hadamard gates and initialising

• For the nth QuBit, 1 Hadamard gate and 0 Rk gates are the QuBits to the zero-position as shown in Figure 15.
needed. Once the initialisation is effected and the Hadamard gates
applied, the circuit aims to obtain a function f (r) = ar
Thus, for this circuit, n Hadamard gates and n(n+1)2 mod N making use of repeated squaring. The circuit then
Rk gates are needed [100]. In addition, QuBits have to
applies an inverse QFT to the input register so as to be able
be swapped during the computation process as the trans-
to obtain a measurement as shown in the right-hand side of
formation is in inverse order. Therefore, we require 3n/2
the figure. This, in effect yields two outcomes, one in the
swap gates to accomplish this. Thus, the total number of
input registers and the other one in the measurement result.
gates required are approximately O(n2 ). The gate complex-
Given these two results, the method uses a classical fraction
ity of O(n2 ) is, in fact, an exponential speedup over the
expansion to find irreducible

2 factors. This

can be done with

a
O (N log (N )) complexity of classical FFT algorithm.
complexity of O log N log log N log log(log N ) .
Note the general number field sieve method has a complexity
3) Shor’s Factorisation Algorithm
2/3


of O exp 1.9(log N )1/3 log(log N ) .
Figure 15 illustrates the order-finding circuit of Shor’s algo-
rithm [101] for quantum factorization as presented in [102]. C. REGULARISATION
Shor’s algorithm solves the problem of, given an integer N , Different methods of regularisation, each providing a differ-
finding its prime factors in polynomial time. This quantum al- ent view of the problem, have been proposed in literature
gorithm is exponentially faster than the general number field [103]–[105]. Among these, it is worth considering those that
sieve method, which is the most efficient known classical explicitly place restrictions on the hypotheses class of can-
factoring algorithm. Shor’s algorithm [101] is effected using didate predictors or those that implement the regularisation
a circuit which consists of two main blocks: A modular expo- effect by introducing noise into the problem. These ideas led
nentiation circuit Ua that implements |xi → |(ax) mod N i to common ML methods currently used in practice, such as
and inverse quantum Fourier transform QF T −1 circuit. Gaussian process (GP) regression and classification [106],
The actual algorithm consists of a reduction step that can regularized least squares [107], support vector machines
be effected on a classical computer and an order finding step (SVMs) [104] and logistic regression [103] to name a few.
that is a quantum primitive. The reduction step consists in Regularization-based approaches use optimisation tech-
finding a number a which shares it greater common divisor niques from a computer-based perspective to solve the learn-
with the integer N under consideration. This is important ing problem that typically requires a number of simple linear
since Shor’s solution employs quantum phase estimation on algebra operations, including matrix multiplication and inver-
a superposition of states such that Ua |xi = | (ax) mod N i. sions. Many related operations are needed for most classical
16 VOLUME 4, 2016
T. M. khan and A. Robles-Kelly: Machine Learning: Quantum Vs Classical

|0> H

−1
|0> H QF T2n

|0> H

0 1
|1> n U a2 U a2 U a2
2n−1

FIGURE 15: Quantum subroutine in Shor’s algorithm.

algorithms, like GPs or SVM, to invert in sets of examples dom features [108] and Nystrom approaches [109], [110].
a square array equal to number N . Normally, this leads to It has recently been shown that even in these settings [111],
a complexity of O(N 3 ), which can be improved according it is possible to obtain comparable generalisation efficiency
to the scarcity and condition of the particular optimization for classical methods. For all of these strategies, training
problem. Nevertheless, with the increasing size of modern times will usually be reduced from O(n3 ) of standard ap-
data sets, the above approaches still approach the limits of proaches to Õ(n2 ) while maintaining the statistical efficiency
their practical use. of the trained estimator essentially unchanged. That said,
Alternate regularisation approaches have recently been since the size of modern data sets is continuously growing,
suggested to reduce the computational costs of learning. the time complexity of the order of Õ(n2 ) might still be too
Such methods are focused on the assumption that, instead challenging for realistic applications. In this regard, quantum
of treating the optimization as a separate process, from the computing may give the potential to further increase the
statistical process the computational burden of the learning efficiency of these methods, enabling them to be substantially
algorithm can be considered a form of regularisation of scaled up. Indeed, by means of a variety of quantum algo-
its own. In case of early stopping approaches, the iterative rithms for linear algebra, sampling and optimization tech-
optimization algorithm, e. g. gradient descent, takes only a niques, we may, in theory, obtain up to exponential speed-ups
small number of steps to avoid over fitting on the training over traditional methods. Nonetheless, current QML methods
set. This technique obviously requires fewer operations, i.e. require fast access to memory and complex data structures
a lesser number of steps, but can theoretically be shown to that may restrict their applicability in real settings. However,
lead to the same generalisation of approaches as Tikhonov’s under various assumptions, a number of findings in quantum
regularisation [105]. learning theory lead to a strong difference between classical
Divide and conquer, in the other hand, is based on the and quantum learning paradigms in certain different settings
concept of distributing portions of the training data to sep- [15].
arate machines, each solving a smaller learning problem, and
then combining individual predictors into a larger one. This V. DISCUSSION AND CONCLUSIONS
approach profits computationally from the parallelization and As mentioned above, quantum computing can help solve
the reduced dimension of the distributed data sets and it has specialized scientific problems such as the modelling of high-
been shown that the same statistical guarantees of classical temperature superconductors, the selection of molecules for
methods can be obtained under the correct training data the creation of organic batteries and drug modelling and
partitions [106]. An alterntive, which has recently received testing. There are several challenges in quantum machine
considerable attention from the ML community, is focused learning that need to be addressed on both hardware as
on the concept of restricting the learning problem to a limited well as on software. Firstly, to get the benefits of quantum
set of candidate predictors, obtained by randomly sampling algorithms, highlighted in this review, quantum hardware
directions in larger, universal space hypotheses that can be will be required to be practical. Secondly, QML requires
realised in a dense space over the domain of a continuous the integration of interface devices such as qRAM in order
function. Depending on how such sampling is done, various to encode the classical information in quantum mechanical
methods have been proposed. The most popular being ran- form. These hardware challenges are non trivial in nature
VOLUME 4, 2016 17
 T. M. khan and A. Robles-Kelly: Machine Learning: Quantum Vs Classical

and must be resolved. Thirdly, in order to fully realise QML [13] O. Higgott, D. Wang, and S. Brierley, “Variational Quantum Computation
approaches, the caveats in the applicability of quantum al- of Excited States,” Quantum, vol. 3, p. 156, 2019.
[14] Z. Jiang, E. Rieffel, and Z. Wang, “Near-optimal quantum circuit for
gorithms need to be resolved. There are four fundamental grover’s unstructured search using a transverse field,” Physical Review,
problems associated with quantum algorithms: input, output, vol. 95, 2017.
cost and benchmarking. [15] J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and
S. Lloyd, “Quantum machine learning,” Nature, vol. 549, no. 7671, pp.
At present, there is very little knowledge, if any, regarding 195–202, 2017. [Online]. Available: https://doi.org/10.1038/nature23474
the true number of gates that are required to implement an [16] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature,
algorithm in QML. Since this methods are purely conceptual vol. 521, no. 7553, pp. 436–444, May 2015. [Online]. Available:
https://doi.org/10.1038/nature14539
at the moment, the complexity of these in terms of integra- [17] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MIT Press,
tion is purely theoretical, being also difficult to predict at 2016.
algorithm level the efficiency between a quantum methods [18] T. M. Khan and A. Robles-Kelly, “A derivative-free method for quantum
perceptron training in multi-layered neural networks,” in To appear,
and classical ones as there are no practical benchmarks ICONIP 2020: International Conference on Neural Information Process-
against modern heuristic methods. Moreover, there are ample ing, 2020.
questions regarding the application of quantum computing to [19] G. Verdon, J. Pye, and M. Broughton, “A universal training algorithm for
quantum deep learning,” 2018.
data arising from quantum phenomena rather that that gath- [20] P. Long and R. Servedio, Restricted Boltzmann Machines are Hard to
ered from practical, non-quantum settings such as consumer Approximately Evaluate or Simulate, Aug. 2010.
applications. [21] V. Dumoulin, I. Goodfellow, A. Courville, and Y. Bengio, “On the
challenges of physical implementations of rbms,” Proceedings of the
Note that, while quantum computing has a great promise in National Conference on Artificial Intelligence, vol. 2, Dec. 2013.
terms of efficiency and scale relative to classical computing, [22] N. Wiebe, A. Kapoor, C. E. Granade, and K. M. Svore, “Quantum
it is still to be seen if this can be fully realised in practice. inspired training for boltzmann machines,” ArXiv, vol. abs/1507.02642,
2015.
Indeed, it is commonly assumed that any problem that can [23] G. Brassard, P. Hoyer, M. Mosca, and A. Tapp, “Quantum amplitude
be solved by quantum computing paradigm can be resolved amplification and estimation,” AMS Contemporary Mathematics Series,
by a classic Turing machine. This, however, would require a vol. 305, Jun. 2000.
[24] G. Hinton, “Training products of experts by minimizing contrastive
large scale of integration, whereby quantum computers are
divergence,” Neural computation, vol. 14, pp. 1771–800, Sep. 2002.
expected to achieve efficiencies that demand much lower [25] N. Wiebe, A. Kapoor, and K. M. Svore, “Quantum deep learning,” 2014.
quantum integration requirements than those in classical [26] S. Adachi and M. Henderson, “Application of quantum annealing to
machines for comparable computational problems. training of deep neural networks,” Oct. 2015.
[27] M. Benedetti, J. Realpe-Gmez, R. Biswas, and A. Perdomo-Ortiz,
“Estimation of effective temperatures in quantum annealers for sampling
REFERENCES applications: A case study with possible applications in deep learning,”
[1] R. S. Smith, M. J. Curtis, and W. J. Zeng, “A practical quantum instruc- Physical Review A, vol. 94, no. 2, p. 022308, Aug. 2016. [Online].
tion set architecture,” 2016. Available: https://link.aps.org/doi/10.1103/PhysRevA.94.022308
[2] M. Schuld, I. Sinayskiy, and F. Petruccione, “The quest for a [28] M. H. Amin, E. Andriyash, J. Rolfe, B. Kulchytskyy, and R. Melko,
quantum neural network,” Quantum Information Processing, vol. 13, “Quantum boltzmann machine,” Physical Review X, vol. 8, no. 2, May
no. 11, pp. 2567–2586, Nov. 2014. [Online]. Available: https: 2018. [Online]. Available: http://dx.doi.org/10.1103/PhysRevX.8.021050
//doi.org/10.1007/s11128-014-0809-8 [29] M. Kieferova and N. Wiebe, “Tomography and generative data modeling
[3] M. Sasaki and A. Carlini, “Quantum learning and universal quantum via quantum boltzmann training,” Dec. 2016.
matching machine,” Physical Review A, vol. 66, no. 2, 2002. [30] M. Schuld, I. Sinayskiy, and F. Petruccione, “The quest for a
[4] E. A¨meur, G. Brassard, and S. Gambs, “Quantum speed-up for unsuper- quantum neural network,” Quantum Information Processing, vol. 13,
vised learning,” Machine Learning, vol. 90, no. 2, pp. 261–287, 2013. no. 11, pp. 2567–2586, Nov. 2014. [Online]. Available: https:
[5] S. Boixo, S. V. Isakov, V. N. Smelyanskiy, R. Babbush, N. Ding, Z. Jiang, //doi.org/10.1007/s11128-014-0809-8
M. J. Bremner, J. M. Martinis, and H. Neven, “Characterizing quantum [31] G. Cybenko, “Approximation by superpositions of a sigmoidal function,”
supremacy in near-term devices,” Nature Physics, vol. 14, no. 6, pp. 595– Mathematics of Control, Signals and Systems, vol. 2, no. 4, pp. 303–314,
600, Jun. 2018. Dec. 1989. [Online]. Available: https://doi.org/10.1007/BF02551274
[6] J. Preskill, “Quantum computing in the nisq era and beyond,” Quantum, [32] E. Farhi and H. Neven, “Classification with quantum neural networks on
vol. 2, no. 79, 2018. near term processors,” 2018.
[7] J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Babbush, and H. Neven, [33] M. Schuld, V. Bergholm, C. Gogolin, J. Izaac, and N. Killoran,
“Barren plateaus in quantum neural network training landscapes,” Nature “Evaluating analytic gradients on quantum hardware,” Physical Review
communications, vol. 9, no. 4812, 2018. A, vol. 99, no. 3, Mar 2019. [Online]. Available: http://dx.doi.org/10.
[8] P.-L. Dallaire-Demers and N. Killoran, “Quantum generative adversarial 1103/PhysRevA.99.032331
networks,” Physical Review A, vol. 98, no. 012324, 2018. [34] Z.-C. Yang, A. Rahmani, A. Shabani, H. Neven, and C. Chamon,
[9] S. Khatri, R. LaRose, A. Poremba, L. Cincio, A. T. Sornborger, and “Optimizing variational quantum algorithms using pontryagin’s mini-
P. J. Coles, “Quantum-assisted quantum compiling,” Quantum, vol. 3, no. mum principle,” Phys. Rev. X, vol. 7, p. 021027, May 2017. [Online].
140, 2019. Available: https://link.aps.org/doi/10.1103/PhysRevX.7.021027
[10] R. LaRose, A. Tikku, ONeel-Judy, L. Cincio, and P. J. Coles, [35] “Physics: The quantum hypothesis. "scientific thought: In context.
“Variational quantum state diagonalization,” npj Quantum Information, . retrieved from encyclopedia.com",” Aug. 2020. [Online].
vol. 5, no. 1, p. 57, Jun. 2019. [Online]. Available: https://doi.org/10. Available: https://www.encyclopedia.com/science/science-magazines/
1038/s41534-019-0167-6 physics-quantum-hypothesis
[11] J. R. McClean, J. Romero, R. Babbush, and A. Aspuru-Guzik, “The [36] M. Lanzagorta and J. Uhlmann, Quantum Computer Science, 2008,
theory of variational hybrid quantum-classical algorithms,” New Journal vol. 1, no. 1.
of Physics, vol. 18, no. 2, 2016. [37] D. E. Weller, “Patterns Linked references are available on JSTOR for this
[12] A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung, X.-Q. Zhou, P. J. article :,” Ecology, vol. 68, no. 4, pp. 813–821, 1987.
Love, A. Aspuru-Guzik, and J. L. OBrien, “A variational eigenvalue [38] D. S. Abrams and S. Lloyd, “Simulation of many-body fermi systems
solver on a photonic quantum processor,” Nature Communications, on a universal quantum computer,” Phys. Rev. Lett., vol. 79, pp.
vol. 5, no. 1, p. 4213, Jul. 2014. [Online]. Available: https: 2586–2589, Sep 1997. [Online]. Available: https://link.aps.org/doi/10.
//doi.org/10.1038/ncomms5213 1103/PhysRevLett.79.2586

18 VOLUME 4, 2016
T. M. khan and A. Robles-Kelly: Machine Learning: Quantum Vs Classical

[39] D. Wecker, M. B. Hastings, and M. Troyer, “Progress towards practical [61] M. Rupp, O. A. von Lilienfeld, and K. Burke, “Guest editorial:
quantum variational algorithms,” Phys. Rev. A, vol. 92, p. 042303, Oct Special topic on data-enabled theoretical chemistry,” J. Chem. Phys.,
2015. [Online]. Available: https://link.aps.org/doi/10.1103/PhysRevA. vol. 148, no. 24, p. 241401, May 2020. [Online]. Available:
92.042303 https://doi.org/10.1063/1.5043213
[40] L. Mueck, “Quantum reform,” Nature Chemistry, vol. 7, no. 5, pp. [62] P. Palittapongarnpim, P. Wittek, E. Zahedinejad, S. Vedaie, and
361–363, 2015. [Online]. Available: https://doi.org/10.1038/nchem.2248 B. C. Sanders, “Learning in quantum control: High-dimensional global
[41] J. W. Cooper, “A re-evaluation of shor’s algorithm,” 2006. optimization for noisy quantum dynamics,” Neurocomputing, vol. 268,
[42] A. Chattopadhyay and V. Menon, “Fast simulation of grover’s quantum pp. 116–126, Dec. 2017. [Online]. Available: http://www.sciencedirect.
search on classical computer,” 2020. com/science/article/pii/S0925231217307531
[43] L. K. Grover, “A fast quantum mechanical algorithm for database search,” [63] A. A. Melnikov, L. E. Fedichkin, and A. Alodjants, “Predicting quantum
1996. advantage by quantum walk with convolutional neural networks,” New
[44] P. R. Giri and V. E. Korepin, “A review on quantum search algorithms,” Journal of Physics, vol. 21, no. 12, p. 125002, Dec. 2019. [Online].
Quantum Information Processing, vol. 16, no. 12, Nov 2017. [Online]. Available: http://dx.doi.org/10.1088/1367-2630/ab5c5e
Available: http://dx.doi.org/10.1007/s11128-017-1768-7 [64] J. Wallnfer, A. A. Melnikov, W. Dr, and H. J. Briegel, “Machine learning
[45] L. Luan, Z. Wang, and S. Liu, “Progress of grover quantum search for long-distance quantum communication,” 2019.
algorithm,” Energy Procedia, vol. 16, pp. 1701 – 1706, 2012, [65] R. Iten, T. Metger, H. Wilming, L. del Rio, and R. Renner,
2012 International Conference on Future Energy, Environment, and “Discovering physical concepts with neural networks,” Physical Review
Materials. [Online]. Available: http://www.sciencedirect.com/science/ Letters, vol. 124, no. 1, p. 010508, Jan. 2020. [Online]. Available:
article/pii/S1876610212002731 https://link.aps.org/doi/10.1103/PhysRevLett.124.010508
[46] J. M. Arrazola, A. Delgado, B. R. Bardhan, and S. Lloyd, “Quantum- [66] P. Huembeli, A. Dauphin, P. Wittek, and C. Gogolin, “Automated
inspired algorithms in practice,” 2019. discovery of characteristic features of phase transitions in many-body
[47] E. Stoudenmire and D. J. Schwab, “Supervised learning with localization,” Physical review B, vol. 99, no. 10, p. 104106, Mar.
tensor networks,” in Advances in Neural Information Processing 2019. [Online]. Available: https://link.aps.org/doi/10.1103/PhysRevB.
Systems 29, D. D. Lee, M. Sugiyama, U. V. Luxburg, 99.104106
I. Guyon, and R. Garnett, Eds. Curran Associates, Inc., [67] L. Banchi, J. Pereira, S. Lloyd, and S. Pirandola, “Convex optimization
2016, pp. 4799–4807. [Online]. Available: http://papers.nips.cc/paper/ of programmable quantum computers,” 2019.
6211-supervised-learning-with-tensor-networks.pdf [68] P. O. Dral, “Quantum chemistry in the age of machine learning,” J.
[48] E. Tang, “A quantum-inspired classical algorithm for recommendation Phys. Chem. Lett., vol. 11, no. 6, pp. 2336–2347, Mar. 2020. [Online].
systems,” in Proceedings of the 51st Annual ACM SIGACT Symposium Available: https://doi.org/10.1021/acs.jpclett.9b03664
on Theory of Computing, ser. STOC 2019. New York, NY, USA: [69] K. T. Schütt, M. Gastegger, A. Tkatchenko, K.-R. Müller, and
Association for Computing Machinery, 2019, p. 217228. [Online]. R. J. Maurer, “Unifying machine learning and quantum chemistry
Available: https://doi.org/10.1145/3313276.3316310 with a deep neural network for molecular wavefunctions,” Nature
[49] H. J. Briegel and G. De las Cuevas, “Projective simulation for artificial Communications, vol. 10, no. 1, p. 5024, Nov. 2019. [Online]. Available:
intelligence,” Scientific Reports, vol. 2, no. 1, p. 400, 2012. [Online]. https://doi.org/10.1038/s41467-019-12875-2
Available: https://doi.org/10.1038/srep00400 [70] J. S. Smith, O. Isayev, and A. E. Roitberg, “Ani-1: an extensible neural
[50] P. Wittek, Quantum Machine Learning: What Quantum Computing network potential with dft accuracy at force field computational cost,”
Means to Data Mining, Aug. 2014. Chem. Sci., vol. 8, no. 4, pp. 3192–3203, 2017. [Online]. Available:
[51] M. Schuld and F. Petruccione, Supervised Learning with Quantum Com- http://dx.doi.org/10.1039/C6SC05720A
puters. Springer, 2018. [71] O. A. von Lilienfeld, “Quantum machine learning in chemical compound
[52] A. Wichert, Principles of Quantum Artificial Intelligence. WORLD space,” Angew. Chem. Int. Ed., vol. 57, no. 16, pp. 4164–4169, May
SCIENTIFIC, May 2020. [Online]. Available: https://doi.org/10.1142/ 2020. [Online]. Available: https://doi.org/10.1002/anie.201709686
8980 [72] K. T. Schütt, F. Arbabzadah, S. Chmiela, K. R. Müller, and
[53] P. Tandon, S. Lam, B. Shih, T. Mehta, A. Mitev, and Z. Ong, “Quantum A. Tkatchenko, “Quantum-chemical insights from deep tensor neural
robotics: A primer on current science and future perspectives,” ynthesis networks,” Nature Communications, vol. 8, no. 1, p. 13890, Jan. 2017.
Lectures on Quantum Computing, vol. 6, no. 1, pp. 1–149, 2017. [Online]. Available: https://doi.org/10.1038/ncomms13890
[54] M. Schuld, I. Sinayskiy, and F. Petruccione, “An introduction to quantum [73] S. Srinivasan, G. Gordon, and B. Boots, “Learning hidden quantum
machine learning,” Contemporary Physics, vol. 56, no. 2, pp. 172–185, markov models,” 2017.
Apr. 2015. [Online]. Available: https://doi.org/10.1080/00107514.2014. [74] D. Widdows and S. Peters, “Word vectors and quantum logic:
964942 Experiments with negation and disjunction,” Proceedings of the 8th
[55] C. Ciliberto, M. Herbster, A. D. Ialongo, M. Pontil, A. Rocchetto, Mathematics of Language Conference, no. June, pp. 141–154, 2003.
S. Severini, and L. Wossnig, “Quantum machine learning: a classical [Online]. Available: http://citeseerx.ist.psu.edu/viewdoc/download?doi=
perspective,” Proceedings of the Royal Society A: Mathematical, 10.1.1.80.3862{&}rep=rep1{&}type=pdf
Physical and Engineering Sciences, vol. 474, no. 2209, p. 20170551, [75] W. Zeng and B. Coecke, “Quantum algorithms for compositional
May 2020. [Online]. Available: https://doi.org/10.1098/rspa.2017.0551 natural language processing,” Electronic Proceedings in Theoretical
[56] S. Arunachalam and R. de Wolf, “A survey of quantum learning theory,” Computer Science, vol. 221, p. 6775, Aug 2016. [Online]. Available:
2017. http://dx.doi.org/10.4204/EPTCS.221.8
[57] V. Dunjko and H. J. Briegel, “Machine learning & artificial intelligence [76] N. Wiebe, A. Bocharov, P. Smolensky, M. Troyer, and K. M. Svore,
in the quantum domain: a review of recent progress,” Reports on “Quantum language processing,” 2019.
Progress in Physics, vol. 81, no. 7, p. 074001, Jun. 2018. [Online]. [77] N. Killoran, T. R. Bromley, J. M. Arrazola, M. Schuld, N. Quesada,
Available: http://dx.doi.org/10.1088/1361-6633/aab406 and S. Lloyd, “Continuous-variable quantum neural networks,” Phys.
[58] M. Benedetti, E. Lloyd, S. Sack, and M. Fiorentini, “Parameterized Rev. Research, vol. 1, p. 033063, Oct 2019. [Online]. Available:
quantum circuits as machine learning models,” Quantum Science and https://link.aps.org/doi/10.1103/PhysRevResearch.1.033063
Technology, vol. 4, no. 4, p. 043001, Nov. 2019. [Online]. Available: [78] Z. Zhao, A. Pozas-Kerstjens, P. Rebentrost, and P. Wittek, “Bayesian
http://dx.doi.org/10.1088/2058-9565/ab4eb5 deep learning on a quantum computer,” Quantum Machine Intelligence,
[59] A. Kamruzzaman, Y. Alhwaiti, A. Leider, and C. C. Tappert, “Quantum vol. 1, no. 1, pp. 41–51, May 2019. [Online]. Available: https:
deep learning neural networks,” in Advances in Information and Com- //doi.org/10.1007/s42484-019-00004-7
munication, K. Arai and R. Bhatia, Eds. Cham: Springer International [79] J. Allcock, C.-Y. Hsieh, I. Kerenidis, and S. Zhang, “Quantum algorithms
Publishing, 2020, pp. 299–311. for feedforward neural networks,” 2018.
[60] G. Carleo, I. Cirac, K. Cranmer, L. Daudet, M. Schuld, N. Tishby, [80] I. Cong, S. Choi, and M. D. Lukin, “Quantum convolutional neural
L. Vogt-Maranto, and L. Zdeborová, “Machine learning and the networks,” Nature Physics, vol. 15, no. 12, pp. 1273–1278, Dec. 2019.
physical sciences,” Reviews of Modern Physics, vol. 91, no. 4, p. [Online]. Available: https://doi.org/10.1038/s41567-019-0648-8
045002, Dec. 2019. [Online]. Available: https://link.aps.org/doi/10.1103/ [81] T. Li, S. Chakrabarti, and X. Wu, “Sublinear quantum algorithms for
RevModPhys.91.045002 training linear and kernel-based classifiers,” 2019.

VOLUME 4, 2016 19
 T. M. khan and A. Robles-Kelly: Machine Learning: Quantum Vs Classical

[82] I. Kerenidis and A. Luongo, “Quantum classification of the mnist dataset [107] F. Cucker and S. Smale, “On the mathematical foundations of learning,”
via slow feature analysis,” 2018. Bulletin of the American Mathematical Society, vol. 39, Nov. 2001.
[83] P. Wittek and C. Gogolin, “Quantum enhanced inference in markov [108] A. Rahimi and B. Recht, “Random features for large scale kernel ma-
logic networks,” Scientific Reports, vol. 7, no. 1, p. 45672, Apr. 2017. chines,” NIPS, vol. 20, pp. 1177–1184, Jan. 2007.
[Online]. Available: https://doi.org/10.1038/srep45672 [109] A. Smola and B. Schlkopf, Sparse Greedy Matrix Approximation for
[84] G. Sentís, A. Monràs, R. Muñoz Tapia, J. Calsamiglia, and E. Bagan, Machine Learning, Jan. 2000.
“Unsupervised classification of quantum data,” Physical Review X, [110] T. Leen, T. Dietterich, V. Tresp, C. Williams, and M. Seeger, “Using the
vol. 9, p. 041029, Nov 2019. [Online]. Available: https://link.aps.org/ nystroem method to speed up kernel machines,” in 13th Int. Conf. on
doi/10.1103/PhysRevX.9.041029 Neural Information Processing Systems, Jan. 2000.
[85] P. Ronagh, “Quantum algorithms for solving dynamic programming [111] A. Rudi, R. Camoriano, and L. Rosasco, Less is More: Nystrm Compu-
problems,” 2019. tational Regularization, Dec. 2015.
[86] A. Cornelissen, “Quantum gradient estimation and its application to
quantum reinforcement learning,” Master’s thesis, Electrical Engineer-
ing, Mathematics and Computer Science, Delft University of Technology,
2018.
[87] T. Fösel, P. Tighineanu, T. Weiss, and F. Marquardt, “Reinforcement
learning with neural networks for quantum feedback,” Physical
Review X, vol. 8, p. 031084, Sep 2018. [Online]. Available:
https://link.aps.org/doi/10.1103/PhysRevX.8.031084
[88] J. Kleinberg and E. Tardos, Algorithm Design. Boston: Pearson TARIQ M. KHAN is working as Research Fellow
Education, 2006. at the Faculty of Sci Eng Built Envat, School
[89] “Quantum computer,” Wikipedia, The Free Encyclopedia, May of Info Technology, Deakin University, Geelong
2007. [Online]. Available: https://en.wikipedia.org/w/index.php?title= Waurn Ponds Campus, Australia. He received his
Quantumcomputer&oldid=127764104 Ph.D. degree in electronic engineering from Mac-
[90] F. R. F. Pereira, E. B. Guedes, and F. M. d. Assis, “Simulating the quan- quarie University Sydney, Australia in 2016. He
tum fourier transform,” in 2013 2nd Workshop-School on Theoretical received a degree in B.S. Computer Engineering
Computer Science, 2013, pp. 40–44. from COMSATS Institute of Information Tech-
[91] N. Wiebe and C. Grandade, “Can small quantum systems learn,” Quan- nology, Islamabad Pakistan and M.Sc. in Com-
tum Info. Comput., vol. 17, no. 78, p. 568594, Jun. 2017. puter Engineering from University of Engineering
[92] G. H. Low, T. J. Yoder, and I. L. Chuang, “Quantum inference on Technology, Taxila Pakistan. His research interests include most aspects
bayesian networks,” Phys. Rev. A, vol. 89, p. 062315, Jun 2014. [Online]. of machine learning, pattern recognition, medical image analysis, scene
Available: https://link.aps.org/doi/10.1103/PhysRevA.89.062315 understanding and deep learning methods for image analysis.
[93] N. Wiebe, D. Braun, and S. Lloyd, “Quantum algorithm for data fitting,”
Phys. Rev. Lett., vol. 109, p. 050505, Aug 2012. [Online]. Available:
https://link.aps.org/doi/10.1103/PhysRevLett.109.050505
[94] A. Kapoor, N. Wiebe, and K. Svore, “Quantum perceptron models,”
in Advances in Neural Information Processing Systems 29, D. D.
Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett, Eds.
Curran Associates, Inc., 2016, pp. 3999–4007. [Online]. Available:
http://papers.nips.cc/paper/6401-quantum-perceptron-models.pdf
[95] P. Rebentrost, M. Mohseni, and S. Lloyd, “Quantum support vector ANTONIO ROBLES-KELLY received a B.Eng.
machine for big data classification,” Physical Review Letters, vol. 113, degree in Electronics and Telecommunications
07 2013. with honours in 1998 and a PhD in Computer
[96] S. Lloyd, M. Mohseni, and P. Rebentrost, “Quantum principal component Science from the University of York, UK, in 2003.
analysis,” Nature Physics, vol. 10, no. 9, pp. 631–633, 2014. [Online]. He remained in York under the MathFit-EPSRC
Available: https://doi.org/10.1038/nphys3029 framework and, in 2005, he moved to Australia
[97] V. Dunjko, J. Taylor, and H. Briegel, “Quantum-enhanced machine and took a research scientist appointment with
learning,” Physical Review Letters, vol. 117, 09 2016. National ICT Australia (NICTA). In 2006 he be-
[98] A. A. Yu Kitaev H Shen M N Vyalyi, S. G. Krantz David Saltman came the project leader at NICTA and, from 2007
David Sattinger Ronald Stern A Kitaev, A. Xenn, and M. I. VVlyy to 2009, he was a Postdoctoral Research Fellow
KLASSIQESKIE KVANTOVYE VYQISLENII, “Classical and Quan- of the Australian Research Council. In 2016, he joined CSIRO where he
tum Computation Graduate Studies in Mathematics,” vol. 12, no. 21, pp. was a Principal Researcher with Data61 and, in 2018 he moved to Deakin
68–2, 2000.
University, where he is a Professor of Machine Learning and the Associate
[99] M. Vogel, Quantum Computation and Quantum Information, by M.A.
Head of School of IT (Research).
Nielsen and I.L. Chuang, 2011, vol. 52, no. 6.
Dr Robles-Kelly’s research has been applied to areas such as logistics,
[100] H.-F. Wang, X.-X. Jiang, S. Zhang, and K.-H. Yeon, “Efficient
quantum circuit for implementing discrete quantum fourier transform in
infrastructure planning, biosecurity, forensics, food quality assurance and
solid-state qubits,” Journal of Physics B: Atomic, Molecular and Optical biometrics and is now being commercialised under the trademark of Scyl-
Physics, vol. 44, no. 11, p. 115502, may 2011. [Online]. Available: larus (www.scyllarus.com). He has served as the president of the Australian
https://doi.org/10.1088%2F0953-4075%2F44%2F11%2F115502 Pattern Recognition Society (APRS) and is an associate editor of the Pattern
[101] Recognition Journal and the IET Computer Vision Journal. He is a Senior
[102] A. Pavlidis and D. Gizopoulos, “Fast quantum modular exponentiation Member of the IEEE, the president of the TC2 (Technical Committee on
architecture for shor’s factorization algorithm,” 2012. structural and syntactical pattern recognition) of the International Associa-
[103] C. Bishop, Pattern Recognition and Machine Learning. New York, NY: tion for Pattern Recognition (IAPR), an Adjunct Academic at the ANU and
Springer, 2006. a Visiting Scientist at CSIRO Astronomy and Space. He has also been a
[104] V. Vapnik, Statistical learning theory. New York, NY: Wiley, 1998, technical committee member, area and general chair of several mainstream
vol. 1. machine learning, computer vision and pattern recognition conferences.
[105] F. Bauer, S. Pereverzev, and L. Rosasco, “On regularization algorithms
in learning theory,” Journal of Complexity, vol. 23, no. 1, pp. 52–72,
Feb. 2007. [Online]. Available: http://www.sciencedirect.com/science/
article/pii/S0885064X06000781
[106] Y. Zhang, J. Duchi, and M. Wainwright, “Divide and conquer kernel ridge
regression.” in Conference on Learning Theory, 2013.

20 VOLUME 4, 2016

View publication stats