Received: 11 April 2024

DOI: 10.1049/qtc2.12100

ORIGINAL RESEARCH
- -Revised: 10 June 2024 Accepted: 24 June 2024

- IET Quantum Communication

Quantum machine learning with Qiskit: Evaluating regression

accuracy and noise impact

Amit Kumar1 | Neha Sharma1 | Nikhil Kumar Marriwala2 | Sunita Panda3 |

M. Aruna4 | Jeetendra Kumar5
1
Chitkara University Institute of Engineering & Technology, Chitkara University, Rajpura, Punjab, India
2
Electronics and Communication Engineering Department, University Institute of Engineering and Technology, Kurukshetra University, Kurukshetra, Haryana, India
3
Department of Electrical, Electronics and Communication Engineering, GITAM School of Technology, Bengaluru, Karnataka, India
4
Faculty of Engineering and Technology, Department of Computing Technologies, College of Engineering and Technology, SRM Institute of Science and Technology, Kattankulathur,
Tamilnadu, India
5
Department of Computer Science and Application, Atal Bihari Vajpayee University Bilaspur, Bilaspur, Chhattisgarh, India

Correspondence Abstract
Nikhil Kumar Marriwala. Quantum machine learning (QML) can be employed in solving complicated machine
Email: nikhilmarriwala@[Link];
nmarriwala@[Link]
learning tasks although the performance in examining the regression processes is only
barely understood. Knowledge gaps are intended to be closed by studying modelling
performance of QML in regression tasks, with emphasis being dedicated to scaling up and
ability to resist noise. The regression part offers the following functions that include
straight line and complex operations. Furthermore, the authors employ quantum neural
networks generated using Qiskit to perform experiments. The results demonstrate that
QML has a remarkable level of accuracy in basic regressions, reaching a maximum of
97%. Nevertheless, there are difficulties in representing intricate functions, such as
5 � cos(x), which results in a noticeable decline in performance. The work deals with the
influence of noise and IERs from imperfect hardware on the efficiency of QML algo-
rithms providing insight into the core obstacles. The result of a detailed examination of
the results that have tested the powers and limits of QML in the development of
regression applications is represented. The future direction of research and development
will be defined by the results obtained in it.

KEYWORDS
quantum communication, quantum computing, quantum computing techniques, quantum cryptography

1 | INTRODUCTION artificial intelligence in QML nowadays and then review the

related work for key technologies and checkpoints in this in-
Quantum machine learning (QML) stands for one of the most dustry. QML has been a growing sector in recent years with a
appealing intersections between quantum computing and ma- pace of rapid advancement being dictated by hardware and
chine learning, as it enables them to solve certain problems that algorithm breakthroughs in the field of quantum computing.
cannot be performed deterministically using classical com- Experts have designed diverse QML models and developed
puters. In the last decade, the study and activity of QML have algorithms within this umbrella of technology tackling a
gained not only a strong interest but the results of the study as multitude of machine learning tasks, such as classification,
well. Therefore, this introduction will first present the overall regression, clustering, and optimisation. These quantum

-
This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is
properly cited.
© 2024 The Author(s). IET Quantum Communication published by John Wiley & Sons Ltd on behalf of The Institution of Engineering and Technology.

310 IET Quant. Comm. 2024;5:310–321. [Link]/journal/qtc2

 26328925, 2024, 4, Downloaded from [Link] Wiley Online Library on [05/01/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
KUMAR ET AL.
- 311

models implement non‐classical principles of quantum me- authors, publication year, study title, and main contribution of
chanics for carrying out computations in ways that have the work. The table briefly presents some important quants
nothing in common with previous approaches; this may mean and machine learning fields, which are overview, analysis, image
an exponent in the order of the speedup for the performance. classification, novel model development, reviews and further
Studies projected in the most recent past examined different advanced studies [10–14]. This compilation is affordable
techniques in QML: quantum neural networks (QNNs) and because of the multiplicity of research areas that are covered.
kernel methods as well as variational algorithms. QNNs This multiplicity encourages specialisation, thereby making it
attracted attention from the experts in particular due to their possible to create unique perspectives concerning diverse
capacity to learn complex patterns and relationships in the data contributions made by researchers in the QML field [15, 16].
by transforming the classical inputs into the quantum circuits. This research aims to evaluate the performance of QNNs in
Conversely, quantum kernel methods obtain a reading order of regression tasks, particularly focusing on their scalability and
features using a quantum feature space that achieves the resistance to noise. Even minimal levels of noise can result in
completion of machine learning tasks rapidly and efficiently. In decoherence, which results in the loss of superposition and
the related line development, the probabilistic approaches, entanglement capabilities of qubits. The loss of quantum in-
such as quantum variational classifiers and quantum variational formation can invalidate computations and restrict the scale
autoencoders, provide hybridised structures for tissue and intricacy of viable quantum algorithms. We investigate how
approximation and optimising the network requirements. The hardware limitations and noise affect QML algorithms and
occurrence of QML represents a strong paradigm shift in propose methods to mitigate these issues. This paper has the
artificial intelligence theories with the aim of solving more goal of being a part of an ongoing discourse by analysing how
complex computational tasks. Kyriaki and her associates have a QML models work in regression tasks, discovering their
clear‐cut tutorial that explains what QML is, together with the weaknesses, and suggesting new avenues for research and
key steps and technologies of the methodology [1]. Subsequent development.
to this furnishing, the research piece of Seongwook presents a Research contributions: our primary contributions in this
different perspective whereby tensor network approaches are paper include the following:
used in the analysis of QML revealing the mystery side of
quantum mechanics and machine learning [2]. Furthermore, 1. Comprehensive Evaluation of QNNs: We provide a
another trend is the work of Alexander et al. with regard to the detailed analysis of QNN performance on regression tasks,
use of QML in image classification tasks, to show the possi- highlighting their accuracy and limitations.
bility of employment of QML in this field as a tool to improve 2. Noise Impact Analysis: We investigate the effects of noise
the traditional machine learning algorithms. Joseph Mawuli and imperfect error recovery on QNNs, providing insights
Lindsay and Ramtin Zand put forward a new LSTM model into how these factors degrade performance.
that draws inspiration from QML. They propose some 3. Hardware Limitations: We examine how current hard-
quantum‐friendly approaches for the recurrent neural net- ware constraints impact the efficacy of QML algorithms,
works for which they highlighted relevant strategies that are offering a pathway for future improvements.
based on quantum principles [3]. Admin iteration Sarkar ana- 4. Methodological Innovations: We propose novel meth-
lyses QML, applying new articles and defining basics for odologies for scaling QNNs and mitigating noise, which are
further study [4]. Meanwhile, A.C and V.M. contribute with validated through extensive simulations.
their study‐focused analysis of QML (a technical word used 5. Future Directions: We outline potential advancements in
here), they critically examine the current methodologies, and quantum error correction and fault‐tolerant quantum
then identify the most promising approach for future research computing, suggesting how these developments could
[5]. They stress out the need for bridging the dichotomy be- enhance QML applications.
tween theoretical ideas and real‐word utilisation and boost for
institutions to sophisticate the advancements on QML in their These contributions aim to bridge existing gaps in QML
respective sectors. Tancara et al. conduct the same research research, offering both theoretical insights and practical
throwing a light on QML possibilities to capture non‐ solutions.
Markovianity by employing the regressor models based on The initial step is the unfolding of the research method-
the kernel [6]. Towards the end, Manjunath and Bhowmik ology that has been utilised in this study as well as the analysis
share their thoughts on recent developments in QML, of the outcome and the implications that they have. Further-
discovering latest breakthroughs, and expanding trends as this more, it will be highlighted that the essay ends with a
field is still being developed [6]. Integrating these studies helps conclusion and more ideas are being explored with research.
build a stronger knowledge base on QML theory, which pro-
motes methodological innovation of practical applications in
future research. 2 | METHODOLOGY
Table 1 summarises review articles in the area of QML that
were identified from the literature. Every row is the repre- The methodology section considers the investigation of
sentation of a research independent of other ones describing regression tasks using QML models via methodology. QML
all the necessary information, for instance, name(s) of the stands for a new wave of machine learning based on the
 26328925, 2024, 4, Downloaded from [Link] Wiley Online Library on [05/01/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
312
- KUMAR ET AL.

TABLE 1 Summary of selected literature on quantum machine learning.

Author Year Study Main contribution

Kyriaki A., et al. [1] 2023 Quantum machine learning—an overview Provides an overview of quantum machine learning

Seongwook S., 2023 Analysing quantum machine learning using tensor network Investigates quantum machine learning using tensor network
et al. [2]

Alexander S., et al. [7] 2023 Quantum machine learning for image classification Explores quantum machine learning for image classification

Joseph Mawuli L., 2023 A novel stochastic LSTM model inspired by quantum Introduces a novel stochastic LSTM model inspired by quantum
et al. [3] machine learning machine learning

Soumya S. [4] 2023 Quantum machine learning: A review Offers a comprehensive review of quantum machine learning

A. C., V. M. [5] 2023 Research oriented reviewing of quantum machine learning Conducts a research‐oriented review of quantum machine learning

Liu J. [8] 2023 Towards real‐world implementations of quantum machine Explores the path towards real‐world implementations of quantum
learning machine learning

Tancara D., et al. [6] 2023 Kernel‐based quantum regressor models learning non‐ Investigates kernel‐based quantum regressor models for learning
Markovianity non‐Markovianity

Manjunath T. D., 2023 Quantum machine learning and recent advancements Discusses recent advancements in quantum machine learning
et al. [9]

principles of quantum mechanics, which may outpace the Methodology encompasses the following key steps:
classical methods and have higher performance due to expo-
nential speed up. Our approach aims to be as systematic as 1. Selection of Regression Tasks: Our analysis commences
possible in gauging the ability and weakness of the QML with the choice of a set of regression tasks that will indicate
models in performing regression tasks over an extensive set of the performance of the models on the other tasks. Here, we
conditions. will look at the variety of tasks carried out using math
Figure 1 outlines the process by visually representing each functions, some of which are simple linear, polynomial, or
step of the process in a sequential manner. Below is a sinusoidal/exponential functions. To this end, the regres-
simplified flowchart depicting the key steps involved in sion tasks will be modified to reflect the possibility of
evaluating the performance of QML models in regression capturing diverse data representations with the knowledge
tasks: of machine learning models being assembled from QML.
The flowchart graphically displays the methodology 2. Quantum Neural Network (QNN) Implementation: In
starting with the regression tasks selecting and ending with terms of our experiments, we simulate QNN models
training results analysis. Here comes model implementation, through the Qiskit package, which is a quantum computing
dataset preparation, training, evaluation, and the result per- platform quickly accessible to a community contributed by
formance metric calculation [17]. In this way, each step is developers. QNNs represent QML's leverage in the ability
logically connected to the next, and finally we arrive at the to process quantum data using the models of quantum
analysis of the reliability of QML models in regression circuits [18–21] parameterised by quantum gates. To fabri-
problems. cate the QNN architectures appropriately fitted to all
regression tasks, one will need to configure the necessary
parameters, such as the number of qubits and layers as well
Algorithm 1: QML regression task. as parameters particular to the complexity of the task.
3. Dataset Preparation: To generate synthetic datasets for
Input: Quantum dataset D, QNN parameters θ
each of the regression tasks so as to have enough data
Output: Trained QNN model
points that can reflect the inherent correlation among the
1: Initialise QNN with parameters θ
underlying function patterns, in order to effectively apply
2: for each epoch do
machine learning algorithms. The datasets are divided into
3: for each batch in D do
two parts for testing and training that enable to monitor and
4: Encode batch into quantum states
optimise the model’s accuracy and efficiency.
5: Apply quantum circuit with
4. Training and Evaluation: To run the QNN models on the
current θ
training datasets using the gradient‐based optimisation al-
6: Measure output and compute loss
gorithms, such as the stochastic gradient descent and the
7: Update θ using gradient descent
Adam optimisation algorithm. This adjustment is done in
8: end for
an iterative manner until a preset loss function has been
9: end for
achieved, which is a weighted measure of the error in the
10: Return trained QNN model
values of inputs and outputs [22–26].
 26328925, 2024, 4, Downloaded from [Link] Wiley Online Library on [05/01/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
KUMAR ET AL.
- 313

FIGURE 1 Methodology flowchart for evaluating quantum machine learning (QML) models in regression tasks.

5. Performance Metrics: To determine the performance of regression with continuous output values based on input data.
trained QNN models, we apply diagnostic metrics prepared The intended model comprises of the quantum circuits, the
specifically for regression tasks. These metrics include MSE optimisable parameters, and the optimisation techniques.
(mean squared error), MAE (mean absolute error), R² score Figure 2 depicts the model architecture together with the
(coefficient of determination), and accuracy percentage to structured diagram.
analyse the results. These metrics tell about the quality,
accuracy, and generalisation aspect of models in regression 1. Input Features: The model takes input features from the
tasking [11, 27–31]. dataset as its input. These features could include various
6. Analysis of Results: The analysis will be done to investi- numerical or categorical variables relevant to the regression
gate the strong and weak points of the QML models in task.
regression. The model's performance on the different 2. Quantum Circuit (QNN): The input features are encoded
regression jobs can be analysed for discerning patterns and into quantum states and processed through a parameterised
trends in the results, and the implications of the findings for quantum circuit shown in Figure 3, which forms the core of
the future research and development in the discipline can be the Quantum Neural Network [3]. The quantum circuit
discussed [17, 32–36]. consists of qubits and quantum gates, with trainable pa-
rameters that are optimised during the training process. The
This systematic approach is intended to deliver a detailed two qubits, (q0,q2), each undergoing a single‐qubit rotation
account of how QML models perform in regression tasks, around the Y‐axis of the Bloch sphere. The rotation angles
enriching our understanding of what QML is capable of and are parameterised by θ[0] for q0 and θ[1] for q1. There is no
what it cannot do in applied scenarios [14, 37–42]. direct interaction between the qubits in this circuit and each
rotation gate acts independently on its respective qubit. This
circuit can be utilised in variational quantum algorithms or
3 | PROPOSED MODEL: QUANTUM quantum state preparation tasks, where the rotation angles
NEURAL NETWORK FOR REGRESSION are adjusted to achieve desired quantum states or compu-
tational outcomes.
In this study, the development of the QNN model for
regression is being expounded. In QNN architecture, we are In QML, the efficacy of given parametrised quantum cir-
going to take advantage of the vast computational power of cuits strongly depends on their generalisation feature, which,
quantum computing to accurately solve the problems of ideally, should include most of the state members which can be
 26328925, 2024, 4, Downloaded from [Link] Wiley Online Library on [05/01/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
314
- KUMAR ET AL.

processes, which increases the efficiency and interpretability of

the algorithm.

3.1 | Mathematical model for noise analysis

In the context of quantum computing, noise can significantly
impact the performance of quantum algorithms, including
those used for machine learning. The primary sources of noise
in quantum systems are decoherence and gate errors. Here is a
mathematical model for analysing these noise effects.
Decoherence: Decoherence refers to the loss of quantum
coherence due to the interaction of the quantum system with
its environment. It can be modelled using the density matrix ρ
of the quantum state and the Lindblad equation:
� � ��
dt i 1
¼ − ½H; ρ� ¼ Σk Lk ρ Lk − Lk Lk ; ρ
† †
dp h 2

FIGURE 2 Proposed quantum neural network architecture for where H, is the Hamiltonian of the system, Lk are the Lindblad
regression.
operators representing different noise processes, and {⋅,⋅}
denotes the anticommutator.
Quantum Gate Errors: Quantum gate errors occur dur-
ing the implementation of quantum gates. These can be
modelled using a noise channel ε that acts on the quantum
state. For a single‐qubit gate, the noisy operation can be rep-
resented as follows:

I
EðρÞ ¼ ð1 − pÞUρU † þ p
2
FIGURE 3 Parameterised quantum circuits. where U is the ideal unitary operation, p is the probability of an
error occurring, and I is the identity matrix representing the
complete depolarisation.
found in the output Hilbert space. This ‘space’ which is virtual
in quantum mechanics is a multiple‐dimension inner product
space with the norm and completeness properties. An 3.2 | ZZ feature map
increased expressibility of a quantum circuit with circuit pa-
rameters has a direct impact on the area of the Hilbert space, The ZZ Feature Map, one of the gates in Qiskit, imports
which in other words defines the hypothesis space. These classical data into the quantum computer in a tailored way by
circuits, however, with their high degree of expressibility can mapping the operations of qubits. It is designed to extract the
depict all the diverse unitaries, therefore increasing their po- higher‐order interactions between the conventional data char-
tential to employ‐general. Parameterised quantum circuits acteristics, ultimately submerging 2‐feature pairs into the data
serve two primary functions in QML: quantum information set. Figure 4 depicts an optimal feature map that places a
science, data encodings, and quantum modelling. They realise sizeable emphasis on CZ gates or controlled Z‐rotations for
digital data as every parameter dynamically adjusts for input data relationship encoding while harnessing the inherent
features and plays the role of a quantum model by doing such quantum principles to compress such complex data in-
parameter optimisation processes which gives them a better teractions into a simple feature map. Qiskit provides access to
advantage in recognising these complex patterns. Feature maps tools such as TwoLocal, NLocal, and QAOA which allow users
are the fundamental ingredients that drive learning in machine to conveniently use parameterised circuits. Such schemes
learning as well as in image processing and pattern recognition provide a wide range of manipulating tools for swift assem-
tasks. The resultant compressed and interpreted data contains blage of customised‐to‐task quantum circuits used in machine
only relevant and non‐duplicated features which can be effi- learning or optimisation. Integrating these frameworks, re-
ciently used for developing the next stage such as feature se- searchers and practitioners are able to engineer quantum cir-
lection and dimensionality reduction—an important cuits with variable parameters; thus, these frameworks facilitate
component of succeeding learning and generalisation the implementation of flexibility and adaptability while meeting
 26328925, 2024, 4, Downloaded from [Link] Wiley Online Library on [05/01/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
KUMAR ET AL.
- 315

FIGURE 4 Parameterised circuit in Qiskit.

diverse computing requirements inside QML and optimisation classical optimisers with specific hardware. Such circuits
domains simultaneously. consist of a featurised map, with the training stanza connecting
In this case, the circuit is composed of Hadamard gates its inputs to its weights. In general, Do NNs embody the
(H), Pauli gates (P), and controlled gates (controlled‐NOT principle of a classical neural network in the quantum realm to
gates, or CNOT, represented by X). Each line represents a resolve challenging issues through the capabilities of quantum
qubit and the operations are applied sequentially from left to computation. Pi for QNA applications, Qiskit makes it all
right. Gates are applied to qubits indicated by the lines they're possible either through simulation or on real hardware.
connected to. q 0 has a Hadamard gate (H), followed by a Pauli Although a simulation provides an opportunity to utilise both
gate Pð2:0∗x½0�Þ, followed by some controlled operations flexibility and accessibility, being granted time on quantum
involving other qubits. q 1 has a similar structure with a hardware has its own dynamics because you then require
Hadamard gate (H), followed by a Pauli gate Pð2:0∗x½1�Þ, and joining a queue and logging into platforms such as IBM Qiskit.
further operations. q 2 again starts with a Hadamard gate (H), By the use of the NeuralNetwork class in qiskit‐ml, all QNNs
followed by a Pauli gate Pð2:0∗x½2�Þ, and more operations. The are provided with the interface for the forward and backward
controlled operations seem to involve some Pauli gates and trainable passes. The interface utilises data samples and train-
further controlled operations, possibly involving the able weights classification as in and out. The quantum net-
qubits q 0; q 1; and q 2. works are just out of state during the learning procedure, each
network in the quantum system itself has no learning capacity
and cannot hold the values for the training weights. Therefore,
3.3 | Training of parameterised QCs training progresses to specific algorithms or tools, that is,
classifiers or regressors, in which the determined trainable
While the classical machine learning method involves the parameters are used to accomplish inferences or predictions.
implementation of the learning process as the minimisation of EstimatorQNN initialises a QNN by taking a para-
a cost function or loss function that is mathematically meterised quantum circuit as input "input parameters” (q)
expressed, the QML setting is similar to training parameterised where input½1� indicates ‘input features’ and weight[3] “Weight
quantum circuits where the loss function or cost function is parameters” represent the tunable parameters within the QNN
shown in Figure 5. Yet, in QML, the purpose is to minimise along with an optional list of observables. It computes the
the expected value, not the excitement of the parameters. expectation value, offering a means to gauge the performance
While classical machine learning uses plenty of algorithms and of the QNN. Additionally, users can include a list of observ-
neural nets to solve learning tasks, QML uses QNNs in order ables to enhance the complexity of the neural network. In
to find the best way for working with qubit‐based computing. quantum mechanics, observables represent operators that
The QNNs' construction is targeted at parameterising ascertain properties of a quantum state through specific se-
quantum circuits that are optimised through variational quences of operations. As shown in Figure 6, the performance
 26328925, 2024, 4, Downloaded from [Link] Wiley Online Library on [05/01/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
316
- KUMAR ET AL.

FIGURE 5 Cost function/loss function (the objective function).

metrics indicate a high level of accuracy for basic regression 3. Optimisation: The optimisation step uses the adjustment
tasks. of the quantities of partial input parameters to reduce the
SamplerQNN is an example of ways around, rather than a loss function. This is usually achieved by the employment of
substitute of EstimatorQNN, in the initialising of QNNs with gradient‐based optimisation methods, like the stochastic
their own strategies of observation. Unlike EstimatorQNN, gradient descent or the Adam optimisation. Here, the
SamplerQNN does not have to be run along with the obser- optimisation technique is used repeatedly for the parameter
vation on its own. Hence, it does not need custom observables updates, such that the model is able to increasingly aim at
to use. Rather than this, the neuromorphic chip provides a the task of regression.
short‐and‐sweet way, that is, through trimming down the steps
and bypassing the need for substantial customisation. This Relation to Research The QNN model is our key research
makes the deploying and implementing processes of qubit methodological unit in this research play that is quantifying the
observables relatively simpler by avoiding complicated pro- performance of QML models in the case of regression prob-
cedures such as explicitly defining observables. Figure 7 lems. Through applying quantum power to the QNN model,
further illustrates the impact of noise on the regression the approach brings greater precision and adequacy that it
models. offers over traditional machine learning methods. We will use
In a QNN, the ‘input parameters’ (q0,q1) refer to the variables exhausting explorations in analysing the effectiveness of the
representing the input data or features provided to the QNN. In suggested model as we want to present the findings that would
this case, `input[0]` and `input [1]` indicate that there are two add knowledge to the general QML community.
input features. On the other hand, the ‘weight parameters’ Physical Realisation: The research presented here has,
represent the tunable parameters within the QNN that are further proved to be the right direction for QML for regres-
adjusted during the training process to optimise the network's sion tasks that introduced novel algorithms, and physicalised it
performance. These parameters influence the behaviour of the as well. The specific problem solving method which is referred
quantum circuit and its ability to make accurate predictions. The to as the QML framework consists of the development of
labels weight½0�; weight½1�; weight½2�; and weight½3� suggest that QNNs, which are actually QNNs that are possible thanks to
there are four trainable weights in the QNN model. the Qiskit tool. We push the boundaries; regression analysis is
not an exception, and we tackle complex issues with multiple
1. Output Prediction: The processed quantum states are high‐dimensional data and complex functions, where QNNs
then applied to formulate the forecast corresponding to the are the right tools for us. Lastly, we use both software emu-
desired targets' output values. These projections take into lations and quantum circuits of Qiskit (which stands for
account a series of non‐stop values that correspond to the Quantum Information Software Kit). We then place our QML
regression coefficients which are calculated for each feature models into Qiskit and test them with an available IBM's
space of the input. quantum system to perform experiments. Being able to
2. Loss Calculation: The predicted output values are conduct the benchmarks helps us to identify the limitations of
compared with corresponding true targets from the dataset our numerical matrix, such gate accuracy, and connectivity
to measure how much they differ. This defines their loss or among the qubits as well as error rates. With implementing this
error. Diverse loss functions may be used to define the approach, we are basically two concurrent processes: from one
deviation of the output from the actual instance, for side, we contribute to the development of the basic theory of
example, MSE or MAE. QML and from another side, and we also provide insights into
 26328925, 2024, 4, Downloaded from [Link] Wiley Online Library on [05/01/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
KUMAR ET AL.
- 317

FIGURE 6 EstimatorQNN [43].

FIGURE 7 SamplerQNN [43].

its practical application, which are very essential for the whole the network can be customised suitably to fact the data effi-
field. ciently, the QML process has therefore highlighted the range
and capabilities of these techniques.

3.4 | Results and discussion

3.5 | Regression analysis
The quantitative results of this work are presented in the
section that consists of the obtained outcomes after the QNN Regression analysis is a statistical method for assessing (a) one
training process on custom data. Through hard experimenta- or more independent variables and their relation (b) to a
tion and evaluation, we study the performance, transferability, dependent variable. It is most often applied for forecasting and
and functionality of the QNN applied to the task as trained. prediction in such areas within the economy, biology, me-
This section reveals the truth about the QNN accuracy, chanics, engineering, and social sciences. As synthetic data
computational ability, interpretability, and scalability which presented by Figure 8 depict, the regression task is in effect
gives an overview of its strengths and limitations in the with additive noise. The plot gives the output corresponding to
physical world. We will scope off and analyse the obtained the input variables and illustrates these in a dataset of simulated
results and thus we expect to get a deeper insight into the graphs.
QML capacities and issues of their applications to the elaborate The blue dots located in Figure 8 have been accepted from
real‐world problems. the dataset's synthetic points for the regression application.
Forward pass and Backward pass: In the first step of Here, the combined work of each blue dot equals the total
forward pass, the SamplerQNN network is with two input weight of a pair of input and output variables. These data
features and 4 trainable weights, resulting in (1, 4) output shape points are around it so the fact can be seen that due to the
and values [[0.23851185 0.1411135 0.3367757 0.28359895]]. presence of noise in such examples is what real life examples
These weights are purposefully used for the forward pass, look like where observations may not be aligned perfectly with
considering that the weighted sums precisely establish the the ideal function. A True red dotted line exhibits the real
values of the actual neural network during the training process. mechanics which are the cosine function ðcosðxÞÞ. This line
After the forward pass is done the backward pass phase begins. demonstrates the ideal relationship between the input and
During the backward pass, the input gradients are calculated output variables without any noise. It acts as a reference to
while the weight gradients are determined as [[[ 0.00609879 show how well the synthetic data points approximate the true
−0.30661818 −0.1834591 −0.28341665] [−0.02969169 function. Therefore, the blue dots indicate the synthetic data
0.17866236 0.1834591 −0.20004909] [−0.37812738 points, while the red dotted line indicates the true underlying
0.30266095 −0.30904568 0.28341665] [ 0.40172028 function.
−0.17470512 0.30904568 0.20004909]]]. These gradients allow Figure 9 represents the increase in a function value over
the parameters in the network to keep being tweaked, during time through which the optimisation indicates the process's
the training process. The gradient descent algorithm has been dynamics. The x‐axis labels the iterations, and they are shown
used in succession of forward and backward passes, to make successively to represent the process of step by step algorithm
the QNN trained with gradients. This underpins the fact that improvement. On the other hand, the x‐axis graphics show the
 26328925, 2024, 4, Downloaded from [Link] Wiley Online Library on [05/01/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
318
- KUMAR ET AL.

FIGURE 8 Synthetic data generated for regression with noise.

FIGURE 9 Objective function value against iteration.

optimisation objective function value which is closely associ-

ated with the effectiveness of the process. To begin with, the
value of objective function at iteration 0 ranges between 0.25
and 0.30, which is very close to the target value of the function.
The next following step, which is to be noticed during the
optimisation process, is that the objective function value re-
duces gradually. By the seventh iteration the function value has
definitely dropped to somewhere between 0.00 and 0.05, which
shows a noticeable improvement in the quality of the process
of optimisation. The monotonic decline of the objective
function rendering is an indication for the optimisation algo-
rithm's iteration process of correcting the parameters again and
again, thus improving the model performance and better
embodying the objectives.
Hadamard gates are used to create superposition states,
which are essential for exploring multiple possibilities simul-
taneously in quantum algorithms. Pauli gates (X, Y, Z) are used
to manipulate the qubit states in various ways, such as flipping FIGURE 10 Objective function value against iteration.
the state or introducing phase shifts, which are crucial for
implementing complex quantum operations.
Performance evaluation: The evaluation criterion, an the model's performance appears suboptimal, especially when
instrument of paramount importance in determining the confronted with simple data distributions such as (5 * times
viability and reliability of regressor in our scientific in- cos(x)).
vestigations, is the performance evaluation metric. The com- Such a gap indicates an underlying flaw in the simplicity
bination of an R squared value of 0.9756 is a reminder of the factor as an important driver of the ideal performances. In
close match between the predictions of the regressor and the addition, the most complex models founded on the use of
actual weather patterns. This number is a reflection of the level QML for some complex functions still face the problem of
of the regression line in conveying the variations of the target performance level in matching with the classical machine
variable sustained using the features as inputs. For instance, the learning algorithms. In the same way, this frat sorry demon-
achieved score of no more than 0.9756 indicates that the re- strates the fact that the field of QML is undergoing rapid
gressor successfully conveys as much as 97.56% of the variance evolution but is definitely in an initial stage, which requires
in the target variable. Such a wonderful achievement is the more research, adequate developments, and testing. QML is
noticeable fact that the regression model which we built has a another ML framework on the block that shows great promise
high level of accuracy as well as reliability. Therefore, this but is currently hindered by the gap with the traditional ML
model is recommended precisely for our research objectives. algorithms. Yet, it has to be mentioned that the fields of
Figure 10 shows the green line which is the regression line classical artificial intelligence which include computer vision,
overlaid to the blue dots, which are the scatter plot. Despite GANs, and NLP have been successfully implementable into
achieving a relatively high score of 97% on the displayed data, the quantum area as well. This convergence signals a solid way
 26328925, 2024, 4, Downloaded from [Link] Wiley Online Library on [05/01/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
KUMAR ET AL.
- 319

that QML could go, despite the fact that it will need extra as democratisation of quantum computing resources. One of
effort and advances to manifest its full potential. the most critical responsibilities of researchers and policy-
makers is the joint development of such frameworks and
regulations that will be used in the implementation of QML
4 | DISCUSSION technologies.
Future Prospects for Universal Quantum Computers
The discussion section delves deeper into the implications of Advancements in Error Correction Codes: Error correction
the results obtained and provides insights into the broader is crucial for building fault‐tolerant quantum computers.
context of QML research. Techniques such as the surface code and the colour code are
Model Performance and Limitations: The fact that the promising because they can correct arbitrary errors on qubits.
quantum AI model regarded the test results as primarily a Research is ongoing to make these codes more efficient and
success points out its strengths and weaknesses at the same practical for large‐scale quantum systems.
time. A model attains an outstanding 97% percent score of Fault‐Tolerant Quantum Computing: Combining
accuracy on simple data though stumbles in its performance quantum error correction with fault‐tolerant techniques en-
when confronted by more intricate functions. This margin sures that quantum computations can proceed reliably even in
shows us the challenges contained in applying the measure- the presence of errors. This involves encoding logical qubits
ments of quantum mechanical learning to ordinary world ap- into multiple physical qubits and using operations that correct
plications in places where the distribution of data is errors dynamically.
stereotypical. Scalability and Hardware Improvements: Significant
Comparison with Classical ML Algorithms: However, progress in hardware, including the development of more
classical algorithms which are being used in machine learning stable qubits, better control systems, and cryogenic technology,
are still considered superior in form, as they tend to outper- will be essential. Companies, like IBM, Google, and Rigetti, are
form their quantum equivalents in various scenarios. Classical making strides in these areas, aiming for scalable quantum
algorithms that have been long‐studied and improved so much processors.
over the many past decades are robust models that perform
well in different areas. Contrary to these facts, the quantum
analogue of machine learning has only just started and the 5 | CONCLUSION
algorithms in the early stages of creation are still under
refinement. Therefore, being able to recognise classically The research considered QML as one of the tools which involve
digestible designs as the basis for achieving parity with them is regression tasks into the quantum world. Our accuracy results
still a significant challenge. for simple datasets give evidence about a prospect for striking
Future Research Directions: The current study revealed similar results but they also bring to light difficulties with more
the limitations. The limitations found in this study will serve as elaborate functions. These revelations thus underline the vital
the basis for future research initiatives on the topic of QML. importance of another attempt of research and development in
Exploring the future of QNNs requires the development of the expanding operations. As we move forward, certainly, aug-
different ways of improving model expressivity, efficient menting the expressivity and the speed of our quantum circuits
optimisation methods, and the design of quantum circuits remains the biggest challenge ahead. With advanced tools such as
which consume less resources. The research direction should novel features maps, analysis, and optimisation techniques
be developed in terms of developing the hybrid approaches exactly made for the QML tasks in use, there might be an
that combine the strengths of both the classical and quantum emergence of more complicated and scalable solutions. Addi-
features of machine learning to outplay the unique short- tionally, classical and QML tech approaches can be integrated
comings of quantum computing hardware. through hybrid models and are expected to fulfil the purpose by
Practical Applications and Real‐World Impact: Sur- overpassing the current officers' limitations. When two tech-
mounting the present challenges, QML features profound niques used are classical optimisation techniques and the power
potential for resolving extreme real world problems. In areas of quantum computing, hybridised algorithms can unravel better
such as drug discovery, financial modelling, or optimisation, performance in many disparate domains and problems. As a
QML is a new kind of science that allows us to look at varied further step, applications of QML to real life issues like drug
industries in a quantitatively different way. Through the discovery, materials, and finances, introduces research that will
implementation of quantum computing, scientists can resolve be adorable, and unimaginably potent. Ensuring that issues
problems that are non‐workable under the classical approach. dealing with data privacy, algorithmic bias, and scalability are
The promise of quantum computing is bringing the solutions being addressed will help to harness QML to the limit, in order to
to these, which is leading to innovative advances in science and resolve the complex and high‐dimensional problems and propel
industry. the innovation forward. Future research should focus on
Ethical and Societal Implications: As with any emerging enhancing error correction techniques and developing more
technology, QML will be accompanied by accompanying robust quantum algorithms. Additionally, exploring QML ap-
ethical and social concerns, which must be addressed. They plications in more complex and real world datasets could
encompass topics such as data privacy, algorithmic bias, as well significantly broaden the scope of quantum computing.
 26328925, 2024, 4, Downloaded from [Link] Wiley Online Library on [05/01/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
320
- KUMAR ET AL.

AUT HO R CO NT R I BU T I ON S 14. Jerbi, S., et al.: Quantum machine learning beyond kernel methods. Nat.
Amit Kumar: Data curation; formal analysis; investigation; Commun. 14(1), 517 (2023). [Link]
software. Neha Sharma: Project administration; supervision; 36159‐y
15. Kumar, A., Sharma, N.: Review of artificial intelligence‐integrated
validation; visualization; writing – original draft. Nikhil blockchain for training autonomous vehicles. In: 2023 International
Kumar Marriwala: Conceptualization; formal analysis; Conference on Computing, Communication, and Intelligent Systems
methodology; project administration; resources; writing – (ICCCIS), Greater Noida, India, pp. 1147–1152 (2023). [Link]
original draft; writing – review & editing. Sunita Panda: 10.1109/ICCCIS60361.2023.10425488
Formal analysis; investigation; validation. M. Aruna: Re- 16. Cai, Z., et al.: Quantum error mitigation. Rev. Mod. Phys. 95(4), 045005
(2023). [Link]
sources; validation; writing – review & editing. Jeetendra 17. Google Colab. (n.d.). [Link]
Kumar: Formal analysis; resources. 1wMpSJDis8VukxU5qsb8a‐NpK3AJivZR9?usp=sharing
18. Kim, Y., et al.: Scalable error mitigation for noisy quantum circuits
CON FLIC T OF I N T ER E ST STAT E M EN T produces competitive expectation values. Nat. Phys. 19(5), 752–759
(2023). [Link]
The authors declare no conflicts of interest.
19. Kumar, A., et al.: Cryptocurrency price forecasting in a volatile landscape:
SARIMAX modeling and short‐term strategies. In: 2023 3rd Interna-
DATA AVA IL AB I LI T Y S TAT E M EN T tional Conference on Smart Generation Computing, Communication and
Data sharing not applicable to this article as no datasets were Networking (SMART GENCON), Bangalore, India, pp. 1–6 (2023).
generated or analysed during the current study. [Link]
20. Borras, K., et al.: Impact of quantum noise on the training of quantum
generative adversarial networks. Journal of Physics: Conference Series
O RC ID 2438(1), 012093 (2023). [Link]
Nikhil Kumar Marriwala [Link] 012093, IOP Publishing
1093-630X 21. Manzano, A., et al.: Parametrized quantum circuits and their approxi-
mation capacities in the context of quantum machine learning. arXiv
(2023). preprint arXiv:2307.14792
RE F ER EN CES 22. Kumar, A., et al.: Blockchain in healthcare: types, privacy challenges, and
1. Tychola, K.A., Kalampokas, T., Papakostas, G.A.: Quantum machine security resilience. In: 2023 3rd International Conference on Smart
learning—an overview. Electronics 12(11), 2379 (2023). [Link] Generation Computing, Communication and Networking (SMART
10.3390/electronics12112379 GENCON), Bangalore, India, pp. 1–6 (2023). [Link]
2. Shin, S., Teo, Y.S., Jeong, H.: Analyzing Quantum Machine Learning SMARTGENCON60755.2023.10442615
Using Tensor Network (2023) 23. Lamata, L.: Quantum machine learning implementations: proposals and
3. Lindsay, J.M., Zand, R.: A Novel Stochastic LSTM Model Inspired by experiments. Adv. Quan. Tech. 6(7), 2300059 (2023). [Link]
Quantum Machine Learning (2023). [Link] 1002/qute.202300059
ISQED57927.2023.10129344 24. Cherbal, S., et al.: Security in internet of things: a review on approaches
4. Sarkar, S.: Quantum machine learning: a review. Int. J. Sci. Technol. Eng. based on blockchain, machine learning, cryptography, and quantum
11(3), 352–354 (2023). [Link] computing. J. Supercomput. 80(3), 3738–3816 (2024). [Link]
5. A., C., V., M.: Research oriented reviewing of quantum machine learning. 10.1007/s11227‐023‐05616‐2
Int. Res. J. Adv. Sci. Hub 5(Issue 05S), 165–177 (2023). [Link] 25. Kumar, A., et al.: Evaluation of users' most discussed and trending topics
10.47392/irjash.2023.s022 in blockchain technologies and smart contracts. In: 2023 3rd Interna-
6. Tancara, D., et al.: Kernel‐based quantum regressor models learning non‐ tional Conference on Smart Generation Computing, Communication and
Markovianity. Phys. Rev. 107(2), 022402 (2023). [Link] Networking (SMART GENCON), Bangalore, India, pp. 1–6 (2023).
1103/physreva.107.022402 [Link]
7. Sedykh, A., Sagingalieva, A.B., Melnikov, A.O.: Quantum machine 26. Kaushik, K., Kumar, A.: Demystifying quantum blockchain for healthcare.
learning for image classification. [Link] (2023). [Link] Security and Privacy 6(3), e284 (2023). [Link]
48550/arXiv.2304.09224 27. Xu, M., et al.: When quantum information technologies meet blockchain
8. Liu, J.: Towards real‐world implementations of quantum machine learning. in web 3.0. IEEE Netw. 38(2), 255–263 (2023). [Link]
Quan. Views 7, 77 (2023). [Link] mnet.134.2200578
9. Manjunath, T.D., Bhowmik, B.: Quantum machine learning and recent 28. Kumar, A., et al.: Navigating the NFT twittersphere: insights from
advancements. In: 2023 International Conference on Artificial Intelli- geographic, temporal, and engagement analysis. In: 2023 3rd Interna-
gence and Smart Communication (AISC), pp. 206–211. IEEE (2023) tional Conference on Smart Generation Computing, Communication and
10. Kumar, A., et al.: Predicting cryptocurrency prices: an exploration of Networking (SMART GENCON), Bangalore, India, pp. 1–7 (2023).
blockchain's impact on traditional currency. In: 2023 3rd International [Link]
Conference on Smart Generation Computing, Communication and 29. Kumar, A., et al.: Unveiling DAO objectives: a textual analysis and topic
Networking (SMART GENCON), Bangalore, India, pp. 1–6 (2023). modeling approach. In: 2023 3rd International Conference on Smart
[Link] Generation Computing, Communication and Networking (SMART
11. Zeguendry, A., Jarir, Z., Quafafou, M.: Quantum machine learning: a GENCON), Bangalore, India, pp. 1–5 (2023). [Link]
review and case studies. Entropy 25(2), 287 (2023). [Link] SMARTGENCON60755.2023.10442501
3390/e25020287 30. Pulicharla, M.R.: Hybrid quantum‐classical machine learning models:
12. Kumar, A., Sharma, N.: An investigation of advancements in blockchain powering the future of AI. J. Sci. & Technol. 4(1), 40–65 (2023)
consensus algorithms & leading protocols. In: 2023 4th IEEE Global 31. Kumar, A., et al.: Blockchain's transcendent evolution: uncovering its
Conference for Advancement in Technology (GCAT), Bangalore, India, impact across industries and cutting‐edge applications. In: 2023 3rd In-
pp. 1–6 (2023). [Link] ternational Conference on Smart Generation Computing, Communica-
13. Peral‐García, D., Cruz‐Benito, J., García‐Peñalvo, F.J.: Systematic litera- tion and Networking (SMART GENCON), Bangalore, India, pp. 1–5
ture review: quantum machine learning and its applications. Comput. Sci. (2023). [Link]
Rev. 51, 100619 (2024). [Link] 10441804
 26328925, 2024, 4, Downloaded from [Link] Wiley Online Library on [05/01/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
KUMAR ET AL.
- 321

32. Gupta, S., Dave, M.: An overview of recommendation system: methods 2023 International Conference on Advances in Computation, Commu-
and techniques. In: Advances in Computing and Intelligent Systems: nication and Information Technology (ICAICCIT), pp. 1083–1087.
Proceedings of ICACM 2019, pp. 231–237 (2020) IEEE (2023)
33. Sandhu, A., et al.: Anti‐theft system for monitoring wheels of a vehicle. J. 40. Kumar, A., Sharma, N.: Review of artificial intelligence‐integrated
Comput. Theor. Nanosci. 16(10), 4299–4303 (2019). [Link] blockchain for training autonomous vehicles. In: 2023 International
1166/jctn.2019.8516 Conference on Computing, Communication, and Intelligent Systems
34. Lindsay, J.M., Zand, R.: A Novel Stochastic LSTM Model Inspired by (ICCCIS), pp. 1147–1152. IEEE (2023)
Quantum Machine Learning (2023). [Link] 41. Egon, K., Rosinski, J., Karl, L.: Quantum Machine Learning: The
ISQED57927.2023.10129344 Confluence of Quantum Computing and AI (2023)
35. Sharma, N., Batra, U.: A study on integrating crypto‐stego techniques to 42. Enad, H.G., Mohammed, M.A.: A Review on artificial intelligence and
minimize the distortion. In: Data Science and Analytics: 4th International quantum machine learning for heart disease diagnosis: current tech-
Conference on Recent Developments in Science, Engineering and niques, challenges, and issues, recent developments, and future directions.
Technology, REDSET 2017, Gurgaon, India, October 13‐14, 2017, Fusion Pract. Appl. 8, 25 (2023)
Revised Selected Papers 4, pp. 608–615. Springer, Singapore (2018) 43. Gonaygunta, H., et al.: Quantum machine learning: exploring quantum
36. Raj, A., Kumar, N., Sheikh, T.H.: Empirical Evaluation of Software algorithms for enhancing deep learning models. Int. J. Adv. Eng. Res. Sci.
Testing Techniques 11(05), 35–41 (2024). [Link]
37. Sharma, N., Batra, U.: Performance analysis of compression algorithms
for information security: a Review. EAI Endorsed Trans. Scalable Inf.
Syst. 7(27) (2020)
38. Kumar, A., et al.: Analyzing bitcoin blockchain data: insights, coefficients How to cite this article: Kumar, A., et al.: Quantum
and model adequacy. In: 2023 International Conference on Advances in machine learning with Qiskit: evaluating regression
Computation, Communication and Information Technology (ICAIC-
accuracy and noise impact. IET Quant. Comm. 5(4),
CIT), pp. 438–445. IEEE (2023)
39. Kumar, A., Sharma, N., Chauhan, R.: Implementation of a privately 310–321 (2024). [Link]
operated blockchain utilizing the multichain open source platform. In: