Machine Learning Algorithms in Quantum

Computing: A Survey
Somayeh Bakhtiari Ramezani Alexander Sommers Harish Kumar Manchukonda
Computer Science and Engineering Dept. Computer Science and Engineering Dept. Computer Science and Engineering Dept.
Mississippi State University Mississippi State University Mississippi State University
Starkville, US Starkville, US Starkville, US
[email protected] [email protected] [email protected]

Shahram Rahimi Amin Amirlatifi

Computer Science and Engineering Dept. Dave C. Swalm School of Chemical Eng.
Mississippi State University Mississippi State University
Starkville, US Starkville, US
[email protected] [email protected]

Abstract —Machine Learning (ML) aims at designing models is an abstract concept of existence, or lack of, connectivity
that learn from previous experience, without being explicitly between two elements in the processing unit, or magnetic
formulated. Applications of machine learning are inexhaustible, orientation of memory sectors.
including recognizing patterns, predicting future trends and
making decisions, and they are capable of handling sizable Similar to the classical computers, the quantum computers
quantities of multi-dimensional data in the form of large vectors use a “quantum bit” or “qubit” [5] which is the probability of
and tensors. To perform these operations on classical computers, “spin up” and “spin down” of an electron after going through a
however, requires vast time and computational resources. Unlike magnetic field. The spin can be thought as an equivalent of the
the classical computers that rely on computations using binary value of a bit in classical computing.
bits, Quantum Computers (QC) benefit from qubits which can
hold combinations of 0 and 1 at the same time via superposition A. Quantum Computation
and entanglement. This makes QCs powerful at handling and post Quantum mechanics deals with an infinite dimensional
processing large tensors, making them a prime target for Hilbert space ( ℋ ) and as a result, requires an infinitely
implementing ML algorithms. While several models used for ML dimensional vector notation. The state of qubit is shown by a
on QCs are based on concepts from their classical computing two dimensional vector in Hilbert space [5], using Dirac’s Bra-
counterparts, utilization of the QC’s potential has made them the Ket notation, which was created by Paul Dirac in 1939 [6]. In
superior of the two. This paper presents an overview of the current this notation, |1 (read “Ket one”) is used for the qubit “on” or
state of knowledge in application of ML on QC, and evaluates the “spin up” state, and |0 (read “Ket zero”) is used for qubit “off”
speed up, and complexity advantages of using quantum machines.
or “spin down”.
Keywords—Machine Learning, Quantum Computing,
1 0
Algorithms |0 = , |1 = (1)
0 1
I. INTRODUCTION AND BACKGROUND
1
The concept of Quantum Computation (QC) was originally As an example, Ket zero shows that the qubit is at a
0
presented in 1982 by Richard Feynman [1] following an spin down state. Here, the first element represents the
observation of the exponential complexity involved in modeling probability of spin down, and the second element shows the
the behavior of a quantum system with the existing knowledge probability of spin up.
of classical computing. Unlike classical computers where the
state of any unit of computation is deterministically prescribed In quantum physics the | notation is used to denote the
to be either a zero or one, the state of the unit of computation in arbitrary state of qubit through superposition. In superposed
a quantum computer could be zero, one or anything in between. qubit, the state of a qubit is the combination of being 0 and at
This unique feature of QC, results in their ability to pursue the same time being 1.
several parallel paths simultaneously in a single calculation unit,
a concept that is not physically possible in classical machines, | = |0 + |1 , ∈ ℂ (2)
and would rather require several passes to be achieved, hence
where | | + | | = 1 . The complex numbers , are
causing higher orders of complexity [2]–[4].
amplitudes of the basic states of |1 and |0 [7]. The continuous
In classical computing, data storage, handling, and linear combination of these two states could be one of any
calculations are established on a binary basis called “bit”, which possible points on a Bloch sphere, presented in Figure 1.

978-1-7281-6926-2/20/$31.00 ©2020 IEEE

Additionally, in Dirac’s notation, the Bra is used for inner ∑ | ( ), (5)
product operations on two vectors or qubits, which is a
Hermitian conjugate of a Ket. For example, | (read bra of ) Such an operation suggests that by applying the inner
is a complex conjugate vector of Ket , where the sign of every product operator matrix on the state matrix that is comprising of
imaginary part is changed, and every Ket is changed into bras all input qubits, the operation is applied to the whole system in
and vice versa. Similarly, the inner product of these two vectors one step; this is known as quantum parallelism [11]. Since
measurement of an unknown state of the system will result in a
collapse of the values, the challenge of using quantum
parallelism lays in taking advantage of parallelism before
measuring the system [13].
One of the strengths of quantum computers lays in their
ability to process large amounts of high dimensional vectors in
polynomial order [3], which makes them a prime target for
machine learning applications. In the present paper, we will
investigate current studies on machine learning algorithms in
quantum computing.
B. Grover’s Algorithm
A key component of almost all NP-complete algorithms is
searching for a particular key or entity in the input set, and lack
| is a scaler [8]. of an efficient search algorithm leads to their dreary
Fig 1. Qubit sate in the Bloch Sphere [9] performance. Grover algorithm is a quantum specific algorithm
that quadratically increases the search performance in a
Quantum mechanics laws dictate that measurement of the randomly ordered dataset, thus surpassing any classical
value of a qubit will collapse the arbitrary state | of the qubit counterparts. While classical algorithms can take 0.5N or worse
to one of the two ground states of |0 or |1 [9]. The probability to find the match in a dataset with N members, the Grover’s
of the final state can be calculated by equation 3 [10]: algorithm can accomplish the same task on a quantum computer
in (√ ) steps [14].
∑ ∈ , | | , (3) C. Quantum Machine Learning
()
Machine Learning (ML) tasks often involve analysis and
where classification of large number of vectors in multi-dimensional
spaces and the trend in ML, already a set of practices which rely
Pr( ) = ∑ , (4) on a sufficient abundance and expressiveness (read complexity)
∈ ,
of data is to apply it to "big data". Quantum Machine Learning
(QML) is a growing field that brings quantum computing and
In equation 3, which is generalized for multi qubit systems,
machine learning together. The main premise of QML is to use
Pr( ) is the probability of the system, as a whole, to be in the |
the inherent advantages of quantum computing (including
state, once the measurement is performed. In other words, after superposition, entanglement, and quantum parallelism), to
the measurement, the state of all qubits will change to |0 or |1 . improve the performance of classical machine learning
As an example, in a two-qubit system, the final state could be algorithms.
|00 , |01 , |10 or |11 .
Two of the biggest challenges that QC faces, are limited
In a multi qubit system, when all qubits are at superposition, number of qubits available, and the high level of noise that is
the value of one qubit could be connected and intertwined to the experienced when using them. Quantum computers need to
value of another qubit; in this case, the change or measurement expand in terms of available qubits, and fault tolerance;
of one qubit will reveal the value of the other qubit; additionally, however, it is anticipated that the Noisy Intermediate-Scale
the square magnitude of probability of all qubits together will be Quantum (NISQ) [15] computers that will become available in
1. This concept is called entanglement [11]. Regardless of the near future, are capable of performing computations that
physical distance of the two entangled quantum systems, a showcase the inherent capabilities of quantum computers.
change in either quantum system, can change the other quantum
system, even far from the first one [12]. As an example, in an In section II we briefly explain the current algorithms that
entangled two-qubit system, called the Bell state, upon are proposed for NISQ era quantum computers. Majority of the
measurement of the first qubit, the second qubit collapses to the proposed algorithms that can be run on near-term devices are
state of the first qubit, thus the final state could be either |00 or heuristic, i.e. we do not know their performance beforehand and
|11 [10]. cannot prove their speed up. Algorithms that have provable
performance advantages over classical computers require vastly
An operator U is defined as a matrix [10], such that it can larger, high fidelity and fault tolerant quantum computers.
apply function ( ) to the system as a whole. Assuming that the Section III offers a comparison between different quantum
qubits in the system are entangled, the computed result of machine learning algorithms, in terms of their relative speedup,
applying Uf on all qubits of | is expressed as [13]: use of Grover algorithm, and existence of an implementation.
 II. QUANTUM MACHINE LEARNING ALGORITHMS whole network. Complexity reduction of the randomized
Majority of quantum machine learning algorithms are based algorithm makes it suitable for large problems with many nodes
on the concepts and architectures borrowed from classical and connections between them.
algorithms, while utilizing quantum physical concepts such as Panella and Martinelli [23] proposed a technique that uses
entanglement and supper position. In some cases, hybrid Boolean functions and proper implementation of nonlinear
systems which are combinations of classical and quantum quantum circuits to transform non-linear data into a linear form.
computers, are used to overcome limitations in the NISQ era Nonlinear quantum operators are used to overcome the lack of
quantum computers. Examples include, the use of Quantum exhaustive search in the network. Their approach had the
Approximate Optimization Algorithm (QAOA) [16] where potential of overcoming the aforementioned hurdles faced in
parameters from classical solution to a problem are used as the QNN. Similar to other implementations, however, their
initial guess and starting point for larger problems, or variational technique included evaluating every possible configuration to
quantum eigen solvers [17], [18] . For the purpose of find the optimized network configuration, but the main
classification, this paper looks at the hierarchy of quantum advantage of their technique was the combination of linear and
machine learning algorithms as Supervised, Semi-Supervised, nonlinear components in the implementation of quantum circuit.
and Un-supervised.
The linear nature of unitary input-output gates, counteracts
A. Supervised Algorithms the efficiency of QC machines in handling linear inputs [24]–
Quantum Neural Network (QNN) is the prime example of [26]. Panella and Martinelli have also presented a model for
supervised quantum algorithm. As an early approach in this Quantum Neuro-Fuzzy Networks (QNFN) based on Binary
field, Narayanan and Menneer [19], showed the theoretical form Neuro-Fuzzy Networks (BNFN) with a tailored binary
of a QNN architecture and how the components of such a system membership function that benefits from inherent quantum
would perform compared to classical counter parts. Although computing features [25]. Their model employs a quantum
their work is high level description of the components, needed, Oracle to get superposition as a way of ensuring parallelism, and
it laid the foundation for future implementations of QNN. to mark the optimal solutions with a dedicated qubit. Likewise,
entanglement is used to tie the superposed solutions to their
Artificial Neural Networks (ANNs) are based on respective network performances. The exponential speedup
aggregation of non-linear functions applied to neurons that are gained by the nonlinear nature of quantum algorithm makes
laid out in layers and sequences. The linear nature of quantum QNFN capable of performing an exhaustive search for the
mechanics, however, makes the implementation of non-linear optimal solution from a staggering number of different QNFNs
activation functions challenging. Cao et al. [20] proposed a in superposition, that is not feasible in classical models. Such an
concept for the building block of QNNs. They have presented a approach is optimized where parameter values are considered,
quantum neuron concept, based on a quantum circuit that but it is not optimal from a complexity or number of rules
natively simulates threshold activation of neurons, and can perspective.
replicate the response from a wide range of ANN settings. Their
proposed model honors inherent quantum advantages, i.e. Similar to classical computers, training and learning of a
superpositions of inputs, quantum coherence and entanglement, QNN relies on existence of a quantum memory. Achieving a
and can be used to construct several classical network Quantum Random Access Memory (QRAM) that is independent
arrangements, including supervised network, unsupervised of the classical counter parts has been an intersection of quantum
network and reinforcement learning. computing and neural networks.
Schuld et al. [21] have presented a general procedure for Ventura and Martinez [27] introduced a quantum associative
modeling quantum perceptron that simulates the step-activation memory neural network architecture through wave functions
function, which is a key component of ANNs. Their approach and operators, that compared to a classical Hopfield network,
needs O(n) qubits, with n being the size of the input, and giving would exponentially improve storage capacity. This technique
an efficiency that is equivalent to classical models. uses patterns as quantum operators applied on two state (half
Implementation of such step activation functions is the key to spin) quantum systems but lacks efficiency in recalling patterns.
neural networks that can be trained on quantum computers, Trugenberger [28] presented a model for quantum
which can potentially benefit from superposition-based memories, where capacity increased exponentially by the
learning, thus training in quantum parallel space. number of qubits and used a probabilistic approach for memory
Ricks and Ventura [22] proposed an algorithm for training retrieval. He later presented [29] a model for quantum
quantum based neural networks, capable of producing high associative memory to overcome the capacity shortage of
accuracy solutions. Their approach uses a minimum set of associative memory in classical models, by benefiting from the
entangled qubits to cover the training data set and can return fact that quantum memory, free from spurious memories. He
results that encompass a predetermined portion of the training used an n-qubit quantum superposition state to store n-bit binary
data. The downside of their approach is that complexity will patterns, but information retrieval in this model was not exact
increase exponentially, as the problem size increases, leading to and relied on a probabilistic approach of finding a pattern in the
extremely complicated scenarios, regardless of the reduction in quantum state memory that most closely resembled the input.
complexity that is offered through the use of quantum machines. Bisio, et al. [30], showed an implementation of the state of a
As a possible solution to this problem, they have proposed a quantum memory through learning to store and retrieve an
modified random version of the algorithm that is less complex, unknown unitary transformation value from a number of
unlike the full model that used composite weight vector for the
training examples. Similarly, [31] used unitary transformation to analysis on MNIST dataset shows that such a classifier would
establish quantum coherence between various instances of a be 98.5% accurate in differentiating handwritten digits. Their
quantum system and conduct quantum Principal Component proposed dimensionality reduction algorithm has the potential
Analysis (PCA), leading to exponential speed up in of direct application on other algorithms that produce quantum
determination of eigenvectors of large systems. data, thus eliminating the need for QRAM, and has a
polylogarithmic execution order in the dimension and number
Oliveira [32] introduced a quantized equivalent for RAM of inputs. While this approach is not necessarily a faster way of
based Neural Networks (RbNNs) that was debuted as quantum classifying data, it is assumed to be more accurate, and can
RAM based Neural Networks (q-RbNN). The main advantage simplify implementation of quantum-based techniques, by
of q-RbNNs is their ability to use the classical learning eliminating the need for QRAM.
algorithms, while benefiting from the physical advantages of
quantum-based machines. Silva, et. al [33]–[35] later expanded Recommender systems use ratings given by a group of users
this idea and presented a model for neural networks based on (M) to several products purchased in the past (N), to
weightless neural nodes, to serve as a Quantum Random Access recommend new products that may align well with preferences
Memory (QRAM) that concurrently uses the ensemble of of a specific user. Such systems can be considered as supervised
training patterns in superposition. Through modification of algorithms, if they provide new suggestions. In these systems,
Grover search algorithm [36] and applying it on a probabilistic it is assumed that users belong to one of K user types, where K
quantum memory, they have been able to run forward and is the rank of ∗ matrix. Kerenidis and Prakash [40] have
backward propagation schemes. proposed an algorithm for recommender systems that
eliminates the need to access all information at once, and just
Silva, et. al [13], presented a supervised learning algorithm
focuses on sampling parts of the matrix that are relevant to the
for QNN that is optimized for quantum learning and is also
row of interest. This has been done through mapping of a vector
capable of operating on classical models. Their approach
to the row space of the matrix. This approach was exponentially
includes a quantum Weightless Neural Network (qWNN) based
faster than known classical algorithms at the time, with an order
on the quantization of the QRAM, capable of receiving all
of O(poly(K)polylog(M.N)). This algorithm, however, inspired
training set patterns concurrently in superposition. They have
a new classical algorithm that is much faster than previous
also presented a Superposition-based Learning Algorithm
classical algorithms [41] and is only polynomially slower than
(SLA) for supervised learning. The SLA determines capability
the Kerenidis and Prakash quantum algorithm.
of the QRAM to accept qubits while in superposition mode. It
also employs a probabilistic approach in determining the Improvements in Support Vector Machines (SVMs), which
optimal configuration for the network, out of a cohort of are used for regression and pattern recognition, have led them to
configurations that are received in superposition. being used in situations where Quadratic-Programming (QP),
which allowed for the application of algorithms to training the
Kerenidis et al. [37] have offered a shallow circuit algorithm
SVMs [42], is no longer possible. These QP-algorithms enabled
for deep Quantum Convolutional Neural Networks (QCNN).
efficient training ([42], [43]) but advances in the use of
Their proposed algorithm is capable of handling non linearities
Rademacher estimates to reduce the complexity of the models
and pooling operations, thus mimicking the behavior of
disallow the use of quadratic programming optimization and this
classical CNN, with potential for offering more sophisticated
can turn optimization into a task with NP-complete complexity
kernels, and handling larger or deeper inputs. A unique feature
[42]. Anguita, et. al. [42] explored the possibility of using QC
of their approach lies in a new quantum tomography where the
based optimization in these cases and compared their
most significant data is extracted with higher likelihood, thus
performance to those of a Monte Carlo classical method. The
decreasing the complexity of the system.
authors point out proof that QC can break the NP-completeness
Similarly, Cong et al. [38] have proposed a QNN barrier is not given yet; still, benefits are evident when applying
architecture for QCNN that uses O(log(N)) on NISQ devices. In QC, especially as the problem complexity increases, presumably
addition to multi-scale entanglement renormalization ansatz, because conventional computing becomes less useful due to the
their approach uses quantum error correction to identify impossibility of its runtime being acceptable.
quantum states for one dimensional topological phases that are
The premise of quantum computing, with its ability to solve
protected by the symmetry. They used sample complexity to
complex problems more rapidly, could solve the complex
demonstrate the performance of their Quantum Phase
optimization problem of SVM, more efficiently. As an example
Recognition (QPR), over String Order Parameters (SOP), where
backing this claim, Kerenidis, et al. [44], have used Interior
QCNN is shown to outperform conventional methods by
Point Method (IPM) method for Second Order Conic
requiring much less repetitions. Additionally, they demonstrate
Programming (SOCP) to train SVM binary classifier on real-
a quantum error correction system based on QCNN that is
world data and have shown that the algorithm works more
optimized for a particular error model. It is shown that their
easily, than in theory. Their results for the performance
approach for concurrent optimization of encoding and decoding
comparison between traditional IPM for SOCPs and quantum
procedures, results in a scheme that is superior over other
IPM for SOCPs show that the quantized instance converged
quantum codes.
around the same iteration, but the runtime of quantum SOCPs
Kerenidis and Luongo [39] presented a quantum classifier are much better than traditional IPM.
based on quantum equivalent of Slow Feature Analysis (QSFA)
Farhi and Gutmann [45] presented a performance
for dimensionality reduction, and Quantum Frobenius Distance
comparison between classical decision tree traverse, and a
(QFD) as an algorithm for classification. Their simulated
similar walk using quantum interference. They show that Support Vector Clustering (SVC) correlates data-points with
quantum interference can result in exponential speed up in states in Hilbert space. These states are represented by Gaussian
traversing a class of trees, rather than randomly walking through wave functions and can allow for the weighting of certain points
them, as is the case for the classical algorithm. They use a single to give them more emphasis, presumably as cluster center
time dependent Hamiltonian to create a quantum state that is possibilities. This is valuable if one has a method, such as SVC,
initially centered around the root of a decision tree but evolves which can be unduly influenced by outliers. With this additional
through its nodes. The Hamiltonian calculation is further sped information computations could be weighed against being
up by determining energy-dependent transmission coefficients. influenced by these outlier points when determining cluster
centers. The one controlled parameter proposed in this method
Building upon the PCA method presented in their earlier controls the width of the clusters that are being sought. Both it
work [31], Rebentrost, et al. [46], used non-sparse matrix and the scale-space method search for cluster centers rather than
simulation to perform PCA and inversion of the training kernel boundaries but the Horn’s method produces minima indicating
matrix, thus reaching exponential speed up with the size of input the centers which are more defined (i.e. deep and robust) than
vectors and number of available training examples for quantized the maxima the alternative approach produces. The complexity
SVM. of this method is order N squared, independent of
B. Unsupervised Algorithms dimensionality. Horn’s method, which easily identifies cluster
Dimensionality reduction and clustering algorithms are centers, is part of a hybrid approach which would then work on
primary examples of unsupervised quantum machine learning the other aspects of cluster identification.
algorithms. One use case of quantum clustering algorithms can Lloyd, et al. [3] presented QML algorithms for assigning
be in privacy enhancement, since the database holding the points to clusters and for finding clusters. They have presented
vectors to be clustered needs to be queried less in frequency and a quantum computer algorithm that can assign vectors of N
volume by virtue of the lower number of calls needed by the features/dimensions to one of M clusters in O(log (M.N)) time,
quantum algorithms used. This exposes less of the database's compared to the best classical algorithms run time of
data to the user of the algorithm. Classical algorithms take O(poly(M.N)). In their approach for k-means clustering, a
polynomial time in vector number and dimension to solve these quantized version of Lloyd’s algorithm [52] is presented in
problems, while QML algorithms can take logarithmic time in O(k.log (k.M.N)) time where M is the number of vectors to
both, thus achieving exponential speed-up. The amount of classify into k clusters and N is the number of features per vector
information O(log MN) that needs to be accessed for QML (the dimensionality of the space all this is going on in). Wiebe,
operations, compared to the size of information O(MN) being et al. [53], have presented algorithms for calculating nearest
applied to classical machine learning, means that privacy of that neighbors and k-means based on the Euclidean distance of the
data is also increased by the use of QML over classical ML. points, and amplitude estimation that precludes the need for
Aïmeur, et. al. [47], [48] have described three quantum measurement. They show that use of these techniques can result
algorithms that could be substituted for components of classical in polynomial reduction in complexity, compared to classical
algorithms and achieve exponential speedup in clustering, over Monte Carlo simulations.
classical algorithms; their quantization presupposes existence of Similarly, Kerenidis, et al. [54], have presented a quantized
a black-box quantum circuit which acts as a distance oracle, variant of k-means algorithm (Q-means) which shows
giving the distance between vector inputs. While such an comparable performance and convergence to classical k-means,
assumption may not be valid, or at least readily available, their and uses k-means with Grover-type quadratic speed up. Such
subroutines can be used, respectively: (1) To find the two most algorithms can provide substantial saving compared to the
distant points from one another in a vector dataset, (2) To find classical algorithms particularly for the case of large datasets.
the n closest points to another, specified point, all in a vector
dataset, (3) To produce neighborhood graphs of vector datasets, Suzuki, et al. [55], have presented a kernel-based quantum
all in times superior to classical counterparts. With these classifier where Pauli decomposition is used for evaluation and
capabilities, their proposed algorithms can be used, respectively creation of the feature map, using the real-value representations.
for quantizing: (1) Divisive clustering, (2) K-medians clustering, They have presented a general formula that calculates the lower
(3) Unsupervised learning algorithms. These subroutines are limit of the accuracy, which helps in finding the selected feature
based on Grover's algorithm and use Grover iterations to isolate map, and separating it linearly.
desired outputs from the results of computing with super C. Semi-Supervised Algorithms
positioned inputs.
Semi-supervised quantum algorithms can be viewed as
Horn [49], [50] has proposed a general solution to the optimization algorithms, quantum enhanced Reinforcement
unsupervised learning problem of clustering using quantum Learning (RL), and generative models.
mechanics. Here, the space of the points is used to create a scale-
We can also consider QC for solving systems of linear
space probability function, which in turn utilizes a Parzen-
window estimator (a means of estimating an unknown equations or determination of Stochastic Gradient Descent
continuous function using a sample of its output [51]) to estimate (SGD), where gradient itself is affine, and doing so with less
the probability distribution that would produce the witnessed demand for quantum memory. Such algorithms, which are
data. The minima of this potential function are then used to widely used in ML, are examples of semi-supervised algorithms.
determine cluster centers. To this end, Kerenidis and Prakash [56] have used the HHL
algorithm [57] to solve systems of linear equations.
 Although researchers have examined ways that Artificial Impatient environments are those in which the learner's
Intelligence (AI) can benefit from Quantum Information response must come within some time limit. In these systems the
Processing (QIP) in supervised and unsupervised learning rather models of environments become less simplified as the
exhaustively, RL still remains as an area that needs to be computation time for a response adds up against the impatience
explored. Learning from experience is a defining signature of an that the scenario includes. Quantum computing could provide
intelligent agent, artificial or otherwise. The volume and the speed-up needed to fix this. Autonomous learning agents,
complexity of information needed to describe real-life situations which are closer to general AIs and combine several sub
in which AI might find itself, still often presents the agent with problems have not yet been demonstrated to be so assisted.
so much data to compute on, as to make the learning take too Paparo, et. al. [64] present a view of AI based, on the view of
long to be useful. intelligence taken by cognitive science and robotics, which is
behavior based and puts a strong emphasis on the physical
To utilize superposition principle and quantum parallelism, aspects of the agent. They present learning agents situated in
Dong, et al. [58], proposed combining RL with quantum theory, classical environments which use quantum memory to process
and introduced Quantum Reinforcement Learning (QRL). They previous experiences of an environment which is unknown
observed that quantum parallelism and probability amplitude (opaque) but gives rewards. The agent learns by taking walks
can help to speed up the learning.
over a directed weighted graph, which are built by its episodic
Dunjk, et al. [59], have produced a schema which can be memory. The use of superposition in performing these walks
used to indicate where quantum computing might be applied to allows exponential speedup over classical analogues methods of
enhance AI and ML, and where it might not. They have doing so. They proposed that general AI will emerge as a result
identified broad settings in RL where a quantum learner interacts of growth from experience of such a system, rather than by
with a quantum environment. They have presented quantum deliberate design of the agent as generally intelligent.
enhanced RL agents that outperform classical counterparts Being able to model complex quantum dynamic systems is
quadratically, and exponential boost in performance over becoming more important and is an operation that only quantum
limited periods. The agent needed less interaction to learn and computers can do, having no classical analog. Quantum
consequently earn the rewards, and it could do so with high classification is considered as a supplier of two qubit states,
probability, thus making the learning much more efficient. without prior knowledge, and a 50-50 chance. These states are
McKieran, et al. [60], defined a framework for using RL to later labeled as either a 0 or a 1. This results in 2n such labeled
improve hybrid quantum-classical computing. The framework states with roughly n of each. Quantum classification would then
has learning environment where the state, action and rewards are be performed by a system that uses these 2n labeled states as a
decided based on the problem. The RL agent generates the training set and can label new states produced by this source.
quantum program and the interaction between agent and the Such a machine would need to store all 2n training states in
quantum resource is repeated until the agent rewards exceed the quantum memory, perform some measurements of them, and the
threshold value. The result in the performance of the trained unknown qubit, and then classify it and this would need to be
agent exceeds the untrained agent as well as that of the QAOA, repeated for each unknown qubit. This would require the
which is a leading near-term hybrid quantum algorithm typically generation of a new training set and storing it in quantum
used to solve combinatorial optimization problems. memory until the moment of classification, for each such
classification.
Generative Adversarial Network (GAN) [61] is a subset of
ML were two networks compete with each other and generate Sentı´s et. al. [65] propose an improved system where a
new data based on the observed training data. Wasserstein Stern-Gerlach measurement [66] is performed on the unknown
GANs (WGAN) is a variation of such networks that uses qubit while all the information needed to control said
Wasserstein distance to determine the distance between the measurement is stored in classical memory. This means that a
actual and generated distributions and is optimistic for new training set does not need to be made and stored in quantum
improving the training stability of GANs because of the memory; instead, information gathered from the previous
continuity and smoothness. Chakrabarti, et al. [62] presented a training sets are used. The excess risk in this model is reduced
quantized alternative to WGANs (qWGAN), to improve the as the number of trained qubits increases. Another advantage of
stability and scalability of the adversarial training of quantum this approach is that for training sets that are big enough, the
generative models on NISQ era machines. In their method, a system will maintain its optimal performance, even when noise,
qWAN implementation is used to replicate the response from a depolarization or statistical variations are imposed to it. In other
3-qubit 1-d Hamiltonian simulation circuit that requires over words, it can attain its robustness and tolerance to noise and
10,000 gates, with a 3-qubit quantum circuit of 50 gates. other anomalies in the training set.
Concept learning is approximating of a membership function D. Quantum Inspired Classical Algorithms
such that members of the mapped from set are mapped to 1, else The concepts of quantum computing, and algorithms presented
to 0. Hunziker, et. al, [63] combined optimal impatient learning in quantum computing have inspired several researchers to
algorithms with a variant of Grover's algorithm, which uses present new or improved classical algorithms. A few examples
queries to an equivalence oracle instead of a membership oracle, of such algorithms include [41], [67]–[69]. Since analysis of
to produce Amplified Impatient Learning (AIL). In their
quantum inspired algorithms is out of scope of the current
approach, task of concept learning is performed using quantum
paper, readers are referred to [70] which presents an excellent
computing and has a query complexity asymptotically smaller
than those of the classical algorithms. review of several quantum inspired classical algorithms.
 III. QUANTUM MACHINE LEARNING ALGORITHM SPEEDUP neural networks. Using the HHL algorithm for solving systems
COMPARISON of linear equations can benefit QML algorithms that use SGD,
and we anticipate that it will capture a lot of attention over the
Table I shows a general overview of current QML
next few years. The probabilistic nature of QC and its
algorithms and denotes whether they are based on or have used
parallelism will give birth to new and quantized version of
the Grover algorithm. This table also reports the speedup
classical algorithms that heavily rely on randomness and parallel
compared to their classical counterpart, and possible
computation. This is especially interesting for semi-supervised
implementations (as opposed to theoretical presentation) of
approaches such as reinforcement learning, deep MPC,
such algorithms. As illustrated in Table I, several of the
Bayesian networks, Markov processes, etc. The ability of
published works have not reported their speed up, or at best,
quantum machines to capture large chunks of data in one pass,
have only reported on their expectations for speed ups.
and applying the algorithm on the data at once, makes them the
TABLE I. GENERAL OVERVIEW OF CURRENT QML ALGORITHMS
perfect conduit for big data analysis in near future.
Algorithm Grover speedup V. REFERENCES
Associative Memory [27] Yes Exponentially [1] R. P. Feynman, “Simulating physics with computers,” Int. J. Theor. Phys.,
Improvement vol. 21, no. 6–7, pp. 467–488, 1982.
in Capacity
Neural Networks [19] Yes Not Specified [2] T. Albash and D. A. Lidar, “Adiabatic quantum computation,” Rev. Mod.
Phys., vol. 90, no. 1, p. 15002, 2018.
Probabilistic Memory [28] No Not Specified
Pattern Recognition [29] Yes Exponentially [3] S. Lloyd, M. Mohseni, and P. Rebentrost, “Quantum algorithms for
Improvement supervised and unsupervised machine learning,” pp. 1–11, 2013.
in Capacity [4] T. F. Rønnow et al., “Defining and detecting quantum speedup,” Science
Support Vector Machines [42] Yes Quadratic (80-. )., vol. 345, no. 6195, pp. 420–424, 2014.
Boosting [71] No Quadratic [5] B. Schumacher, “Quantum coding,” Phys. Rev. A, vol. 51, no. 4, pp.
Unitary Transformation [30] No Not Specified 2738–2747, 1995.
Weightless NN [13] Yes Not Specified [6] P. A. M. Dirac, “A new notation for quantum mechanics,” Math. Proc.
K-medians [47] Yes Quadratic Cambridge Philos. Soc., vol. 35, no. 3, pp. 416–418, 1939.
Divisive Clustering [47] Yes Quadratic [7] P. Wittek, Quantum Machine Learning What Quantum Computing Means
K-means [3] Optional Exponential to Data Mining, 1st ed. San Diego, CA, USA: Elsevier, 2014.
Principal Components Analysis [31] No Exponential [8] J. C. Solem and L. C. Biedenharn, “Understanding geometrical phases in
Support Vector Machine [46] No Exponential quantum mechanics: An elementary example,” Found. Phys., vol. 23, no.
Nearest neighbors [53] Yes Quadratic 2, pp. 185–195, 1993.
Neuro-Fuzzy Network [23] No Not Specified
[9] R. A. Servedio and S. J. Gortler, “Quantum versus classical learnability,”
RAM-based NN[23] Yes Not Specified Proc. Annu. IEEE Conf. Comput. Complex., pp. 138–148, 2001.
Deep Convolutional NN [37] No Not Specified
[10] A. Atıcı, “Advances in Quantum Computational Learning Theory,”
Convelutional NN [38] No Not Specified
Columbia University, 2006.
Feature Reduction [39] No No
Recommender Systems [40] Yes Exponential [11] V. Vedral and M. B. Plenio, “Basics of quantum computation,” Prog.
Quantum Electron., vol. 22, no. 1, pp. 1–39, 1998.
Interior Point Method for Second No Asymptotic
Order Conic Programming [44] [12] A. Steane, “Quantum computing,” Reports Prog. Phys., vol. 61, no. 2, pp.
Q-means [54] No Exponential 117–173, Feb. 1998.
Pauli Decomposition [55] No Not Specified [13] A. J. Silva, W. R. Oliveira, and T. B. Ludermir, “Classical and superposed
Reinforcement Learning [58] Yes Not Specified learning for quantum weightless neural networks,” Neurocomputing, vol.
Gradient Descent [56] No Exponential 75, no. 1, pp. 52–60, 2012.
Wasserstein GAN [62] No Not Specified [14] L. K. Grover, “Quantum mechanics helps in searching for a needle in a
Proximal Policy Optimization [60] No Not Specified haystack,” Phys. Rev. Lett., vol. 79, no. 2, pp. 325–328, 1997.
[15] J. Preskill, “Quantum Computing in the NISQ era and beyond,” Quantum,
IV. CONCLUSION vol. 2, no. July, p. 79, 2018.
Quantum computing and its unique properties, including [16] F. G. S. L. Brandao, M. Broughton, E. Farhi, S. Gutmann, and H. Neven,
“For Fixed Control Parameters the Quantum Approximate Optimization
superposition, quantum parallel computation and entanglement, Algorithm’s Objective Function Value Concentrates for Typical
have attracted broad interest from different fields of science and Instances,” pp. 1–16, Dec. 2018.
engineering. While several attempts are made at transforming [17] A. Peruzzo et al., “A variational eigenvalue solver on a photonic quantum
classical machine learning algorithms into their quantum processor,” Nat. Commun., vol. 5, no. 2, pp. 1–10, 2014.
computing equivalents, especially in the NISQ era, significant [18] J. R. McClean, J. Romero, R. Babbush, and A. Aspuru-guzik, “The theory
gaps exist in implementation, usability and utilization of pre- of variational hybrid quantum-classical algorithms,” New J. Phys., vol.
quantum supremacy era hardware. This work offered a general 18, no. 2, pp. 1–20, 2016.
comparison between viability, performance improvement and [19] A. Narayanan and T. Menneer, “Quantum artificial neural network
architectures and components,” Inf. Sci. (Ny)., vol. 128, no. 3, pp. 231–
current implementation of QC-ML algorithms. It was shown 255, 2000.
that QNN is one of the most extensively investigated supervised
[20] Y. Cao, G. G. Guerreschi, and A. Aspuru-Guzik, “Quantum Neuron: an
quantum learning algorithms, however, earlier implementations elementary building block for machine learning on quantum computers,”
of QNN did not fully exploit the power of quantum computing pp. 1–30, 2017.
and parallelism. The linear nature of unitary input-output gates [21] M. Schuld, I. Sinayskiy, and F. Petruccione, “Simulating a perceptron on
makes QC efficient when handling linear input. These two a quantum computer,” Phys. Lett. Sect. A Gen. At. Solid State Phys., vol.
factors adversely affect the performance of quantum based 379, no. 7, pp. 660–663, 2015.
[22] B. Ricks and D. Ventura, “Training a quantum neural network,” Adv. - ICML ’07, 2007, vol. 227, pp. 1–8.
Neural Inf. Process. Syst., 2004. [48] E. Aïmeur, G. Brassard, and S. Gambs, “Quantum speed-up for
[23] M. Panella and G. Martinelli, “Neurofuzzy networks with nonlinear unsupervised learning,” Mach. Learn., vol. 90, no. 2, pp. 261–287, 2013.
quantum learning,” IEEE Trans. Fuzzy Syst., vol. 17, no. 3, pp. 698–710, [49] D. Horn and A. Gottlieb, “The Method of Quantum Clustering,” no. April,
2009. 2002.
[24] A. De Vos and Y. Van Rentergem, “Synthesis of reversible logic for [50] D. Horn and A. Gottlieb, “Algorithm for Data Clustering in Pattern
nanoelectronic circuits,” Int. J. Circuit Theory Appl., vol. 35, no. 3, 2007. Recognition Problems Based on Quantum Mechanics,” Phys. Rev. Lett.,
[25] M. Panella and G. Martinelli, “Neural networks with quantum vol. 88, no. 1, p. 4, 2002.
architecture and quantum learning,” Int. J. Circuit Theory Appl., vol. 39, [51] H. Schløler and U. Hartmann, “Mapping neural network derived from the
no. 1, pp. 61–77, Jan. 2011. parzen window estimator,” Neural Networks, vol. 5, no. 6, pp. 903–909,
[26] R. P. Feynman, “Quantum mechanical computers,” Found. Phys., vol. 16, 1992.
no. 6, 1986. [52] S. P. Lloyd, “Least Squares Quantization in PCM,” IEEE Trans. Inf.
[27] D. Ventura and T. Martinez, “Quantum Associative Memory,” Int. J. Theory, vol. 28, no. 2, pp. 129–137, 1982.
Mod. Phys. B, vol. 17, no. 12, pp. 2447–2472, Jul. 1998. [53] N. Wiebe, A. Kapoor, and K. M. Svore, “Quantum algorithms for nearest-
[28] C. A. Trugenberger, “Probabilistic Quantum Memories,” Phys. Rev. Lett., neighbor methods for supervised and unsupervised learning,” Quantum
vol. 87, no. 6, pp. 67901-1-67901–4, 2001. Inf. Comput., vol. 15, no. 3–4, pp. 318–358, 2015.
[29] C. A. Trugenberger, “Quantum Pattern Recognition,” in Invited Talk at [54] I. Kerenidis, J. Landman, A. Luongo, and A. Prakash, “q-means: A
the 1st Feynman Festival, Univ. of Maryland, College Park, August 2002, quantum algorithm for unsupervised machine learning,” 2018.
2002. [55] Y. Suzuki et al., “Analyzing and synthesis of feature map for quantum
[30] A. Bisio, G. Chiribella, G. M. D’Ariano, S. Facchini, and P. Perinotti, classifier,” pp. 2–7, 2019.
“Optimal quantum learning of a unitary transformation,” Phys. Rev. A - [56] I. Kerenidis and A. Prakash, “Quantum gradient descent for linear
At. Mol. Opt. Phys., vol. 81, no. 3, 2010. systems and least squares,” 2017.
[31] S. Lloyd, M. Mohseni, and P. Rebentrost, “Quantum principal component [57] A. W. Harrow, A. Hassidim, and S. Lloyd, “Quantum algorithm for linear
analysis,” Nat. Phys., vol. 10, no. 9, pp. 631–633, 2014. systems of equations,” Phys. Rev. Lett., vol. 103, no. 15, pp. 1–4, 2009.
[32] W. R. de Oliveira, “Quantum RAM based neural networks,” ESANN 2009 [58] D. Dong, C. Chen, H. Li, and T. J. Tarn, “Quantum reinforcement
Proceedings, 17th Eur. Symp. Artif. Neural Networks - Adv. Comput. learning,” IEEE Trans. Syst. Man, Cybern. Part B Cybern., vol. 38, no. 5,
Intell. Learn., no. April, pp. 331–336, 2009. pp. 1207–1220, 2008.
[33] A. J. Silva, T. B. Ludermir, and W. R. Oliveira, “Superposition based [59] V. Dunjko, J. M. Taylor, and H. J. Briegel, “Quantum-Enhanced Machine
learning algorithm,” Proc. - 2010 11th Brazilian Symp. Neural Networks, Learning,” Phys. Rev. Lett., vol. 117, no. 13, pp. 1–6, 2016.
SBRN 2010, pp. 1–6, 2010.
[60] K. A. McKiernan, E. Davis, M. S. Alam, and C. Rigetti, “Automated
[34] A. Silva, W. De Oliveira, and T. Ludermir, “A weightless neural node quantum programming via reinforcement learning for combinatorial
based on a probabilistic quantum memory,” Proc. - 2010 11th Brazilian optimization,” 2019.
Symp. Neural Networks, SBRN 2010, no. June 2015, pp. 259–264, 2010.
[61] I. Goodfellow et al., “Generative adversarial nets,” in Advances in neural
[35] W. R. De Oliveira, A. J. Silva, T. B. Ludermir, A. Leonel, W. R. Galindo, information processing systems, 2014, pp. 2672–2680.
and J. C. C. Pereira, “Quantum logical neural networks,” Proc. - 10th
Brazilian Symp. Neural Networks, SBRN 2008, no. October, pp. 147–152, [62] S. Chakrabarti, Y. Huang, T. Li, S. Feizi, and X. Wu, “Quantum
2008. Wasserstein Generative Adversarial Networks,” no. NeurIPS, pp. 1–23,
2019.
[36] L. K. Grover, “A fast quantum mechanical algorithm for database search,”
Jpn. J. Hum. Genet., vol. 41, no. 3, pp. 333–338, May 1996. [63] M. Hunziker, D. A. Meyer, and J. Park, “The geometry of quantum
learning,” no. August, pp. 1–20, 2003.
[37] I. Kerenidis, J. Landman, and A. Prakash, “Quantum Algorithms for Deep
Convolutional Neural Networks,” pp. 1–40, Nov. 2019. [64] G. D. Paparo, V. Dunjko, A. Makmal, M. A. Martin-Delgado, and H. J.
Briegel, “Quantum speedup for active learning agents,” Phys. Rev. X, vol.
[38] I. Cong, S. Choi, and M. D. Lukin, “Quantum convolutional neural 4, no. 3, pp. 0–14, 2014.
networks,” Nat. Phys., pp. 1–12, 2019.
[65] G. Sentís, J. Calsamiglia, R. Muñoz-Tapia, and E. Bagan, “Quantum
[39] I. Kerenidis and A. Luongo, “Quantum classification of the MNIST learning without quantum memory,” Sci. Rep., vol. 2, no. 1, p. 708, Dec.
dataset via Slow Feature Analysis,” pp. 1–25, 2018. 2012.
[40] I. Kerenidis and A. Prakash, “Quantum recommendation system,” Leibniz [66] W. Gerlach and O. Stern, “Der experimentelle Nachweis der
Int. Proc. Informatics, LIPIcs, vol. 67, pp. 1–22, 2017. Richtungsquantelung im Magnetfeld,” Zeitschrift f r Phys., vol. 9, no. 1,
[41] E. Tang, “A quantum-inspired classical algorithm for recommendation pp. 349–352, Dec. 1922.
systems,” Proc. Annu. ACM Symp. Theory Comput., pp. 217–228, 2019. [67] O. P. Patel, N. Bharill, A. Tiwari, and M. Prasad, “A Novel Quantum-
[42] D. Anguita, S. Ridella, F. Rivieccio, and R. Zunino, “Quantum inspired Fuzzy Based Neural Network for Data Classification,” IEEE
optimization for training support vector machines,” Neural Networks, vol. Trans. Emerg. Top. Comput., pp. 1–1, 2019.
16, no. 5–6, pp. 763–770, 2003. [68] Kuk-Hyun Han and Jong-Hwan Kim, “Quantum-inspired evolutionary
[43] G. Zanghirati and L. Zanni, “A parallel solver for large quadratic algorithm for a class of combinatorial optimization,” IEEE Trans. Evol.
programs in training support vector machines,” Parallel Comput., vol. 29, Comput., vol. 6, no. 6, pp. 580–593, Dec. 2002.
no. 4 SPEC., pp. 535–551, 2003. [69] K. H. Han and J. H. Kim, “Genetic quantum algorithm and its application
[44] I. Kerenidis, A. Prakash, and D. Szilágyi, “Quantum algorithms for to combinatorial optimization problem,” Proc. IEEE Conf. Evol. Comput.
Second-Order Cone Programming and Support Vector Machines,” pp. 1– ICEC, vol. 2, pp. 1354–1360, 2000.
29, 2019. [70] O. H. Montiel Ross, “A Review of Quantum-Inspired Metaheuristics:
[45] E. Farhi and S. Gutmann, “Quantum computation and decision trees,” Going from Classical Computers to Real Quantum Computers,” IEEE
Phys. Rev. A - At. Mol. Opt. Phys., vol. 58, no. 2, pp. 915–928, 1998. Access, vol. 8, pp. 814–838, 2020.
[46] P. Rebentrost, M. Mohseni, and S. Lloyd, “Quantum support vector [71] H. Neven, V. S. Denchev, G. Rose, and W. G. Macready, “Training a
machine for big data classification,” Phys. Rev. Lett., vol. 113, no. 3, pp. Large Scale Classifier with the Quantum Adiabatic Algorithm,” pp. 1–11,
1–5, 2014. 2009.
[47] E. Aïmeur, G. Brassard, and S. Gambs, “Quantum clustering algorithms,”
in Proceedings of the 24th international conference on Machine learning