0% found this document useful (0 votes)
15 views6 pages

Quantum Based Support Vector Machine Identical To Classical Model

Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
15 views6 pages

Quantum Based Support Vector Machine Identical To Classical Model

Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 6

IEEE - 49239

Quantum based Support Vector Machine Identical to


Classical Model

Disha Uke, Kapil Kumar Soni and Akhtar Rasool


Computer Science Engineering
Maulana Azad National Institute of Technology
Bhopal, India

Abstract—The paper presents an influence of machine the supervised learning. Whereas in an unsupervised learning,
learning techniques that can be efficiently solved using quantum the labels are missing, k-means clustering, Gaussian mixture
computer. Machine learning is considered as a core concept of model, hierarchical clustering, recommender system and low-
modeling and mining large dimensionality data and visualizes dimensional embedding belong to this category.
data instances as a vector. Support vector is one of the trending
methods that use supervised learning for classifying the data Machine learning has a lot to adapt from quantum
efficiently irrespective of linearity or non-linearity in data. As the mechanics; superposition, and sensitivity to context,
quantum computations already proved to be inherently parallel entanglement and the linearity of evolution proves to be useful
and can become one of the key factors in order to reduce the in many scenarios [1]. Many fields have achieved great
complexities while dealing with the classical machine learning advantage from its equivalent quantum version and many are
model. The quantum machine learning takes advantage of still on progress. Quantum version of particle swarm
parallelism and can explore the number of vectors and their optimization [2] finds the global optimum for the optimization
dimensions in logarithmic time, and hence it can achieve the problem. The great advantage in dynamic quantum clustering
computational speedup. Now, the main contribution of the paper is that the methods can be computed with the simple linear
is to discuss quantum fundamental, quantum algorithm and then algebra operations. Quantum neural network [3] exploits the
to review over classical equivalent quantum support vector superposition of quantum states to accommodate gradual
machine along with their training algorithm and the comparative
increase of data instances. Quantum version has gain higher
analysis between algorithmic growths. At last, the paper justifies
accuracy and effectiveness over their existential version.
the quantum support vector as a highly efficient model that gains
the computational speedup and also shows the possibility of In this paper, we have taken support vector machine which
further improvement. is a well known and effective method for regression and
pattern classification which often leads to outstanding
Keywords— supervised learning, support vector machine performance in real world applications. Support vector
(SVM), superposition, quantum support vector machine (QSVM). machine (SVM) minimizes the error classification by using
the optimization theory to classify that is based on stastical
I. INTRODUCTION learning. SVM is better applicable in pattern recognition,
regression estimation, time-series forecasting, marketing,
Machine learning is a field of artificial intelligence that production efficiency estimation, text classification, face
seek pattern from a large set of data that means the approach is recognition using image, recognition of hand writing and
more of a data driven rather than model driven. Examples like medical diagnosis in comparison with any other learning
ranking of web pages face detection, controlling robots and technique. SVM starts with optimal margin hyperplanes that
many more works on the core concept of machine learning. separate two classes expecting the higher performance than
Machine Learning can be mainly classified under categories any other. Hard margin that clearly separate the two classes
like supervised learning and unsupervised learning. In and the soft margin which allow noise that is in case where
supervised learning, each data point in a feature space comes two classes are not separable. With the use of kernel function,
with a label also known as output or a response. A supervised non-linearity is addressed, allowing the data to be embedded
learner predicts the label of instances after training on a into the higher dimensions where they can be linearly
sample of labeled examples that is the training set. The two separable.
main categories of supervised learning are classification and
regression. In regression, label has a continuous numerical Equivalent to the classical model quantum version support
range. Therefore, regression mainly predicts the problem that vector machine uses Grover’s search to replace sequential
is based on sizes and quantities. In classification, labels are minimum optimization. Least-square SVM translate an
mostly finite set of numerical values like yes/no, 0/1, +1/-1 optimization problem into a set of linear equations. We
that is binary classification or more than two labels also compare various algorithms that assist the quantum SVM that
known as multiclass classification. Algorithms like linear results in achieving higher accuracy rate. Linear equation
regression, logistic regression, decision tree, Naïve Bayes, requires quick calculation of the kernel matrix which is one
Support vector machine, random forest, adaboost falls under source of the speedup and the efficient solution of the linear

11th ICCCNT 2020


Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on May 08,2024 at 18:11:50 UTC from IEEE Xplore. Restrictions apply.
July 1-3, 2020 - IIT- Kharagpur
Kharagpur, India
IEEE - 49239

equations on the quantum hardware being another source of Each of 𝑢ₘ is p-dimensional real vector.
speedup in the quantum version. The classical SVM where we
have considered N number of features that represent a single The equation that describes the hyperplanes is
data instance and M number of training instances has at least
𝑂(𝑀 2 (𝑀 + 𝑁)) complexity. The quantum version yields an 𝑊. 𝑢 + 𝑏 = 0 (2.1)
exponential speedup leading to overall complexity of Where,
𝑂(log 𝑀𝑁). W is p-dimensional vector perpendicular to hyperplane.
b is a scalar that increases the margin.
In this paper, section 2 describes the classical support
vector machine. Section 3 introduces the quantum
fundamental concepts followed by section 4 which discuss the The two parallel hyperplanes can be described by the equation
quantum version of support vector machine along with the 𝑊. 𝑢 + 𝑏 = 1.
different training algorithms. Section 5 concludes the 𝑊. 𝑢 + 𝑏 = −1
computational analysis between the two approaches and at last
section 7 which provide the conclusion. By solving the above two equations, distance between two
2
planes is‖ ‖.
𝑊

Therefore, to maximize the distance between the hyperplanes


we need to minimize‖𝑊‖ . Considering every data instances
the equation of hyperplanes becomes
𝑊. 𝑢ₛ + 𝑏 ≥ 1
𝑊. 𝑢ₛ + 𝑏 ≤ −1
Samples along the hyperplane are called support vector.

Fig.2. Hyperplanes for a support vector machine trained with samples of two
Fig.1. Organization of the paper classes.

For all the data, the optimal hyperplane should satisfy the
II. CLASSICAL SUPPORT VECTOR MACHINE following constraints
Support vector machine (SVM) introduced by Boser, vₛ. [𝑊 𝑇 . 𝑢ₛ + 𝑏] ≥ 1 (2.2)
Guyon and Vapnik in COLT-92 is a supervised learning Where
method used for both classification and regression [4]. 𝑢ₛ is the sth training example.
Applications like hand writing analysis, face analysis and
Thus, a learning machine should minimize ‖𝑊‖2 subject to,
pattern classification has achieved many great results by the
SVM. The main goal of SVM is to construct an optimal vₛ . [𝑊 𝑇 . 𝑢ₛ + 𝑏] ≥ 1 s= 1, 2, 3……m
hyperplane that separates the classes with the help of two Where
parallel hyperplanes and to maximize the distance between m is the training data points.
them. Let us consider the data instances of the form:- The optimization is subject to these constraints and seeks the
{(𝑢₁, 𝑣₁)(𝑢₂, 𝑣₂)(𝑢₃, 𝑣₃) … … … … (𝑢ₘ, 𝑣ₘ)} optimal decision hyperplane
1
𝐿ₚ = ‖𝑊‖2
2
Where, We more often refer to the dual formulation of the problem.
𝑣ₘ = −1/1, binary classification labeling. We introduce Lagrange multiplier 𝛼ₛ.
𝑢ₘ = 𝑚 , Number of data instances where each 𝑢 ∈ 𝑅𝑁 .
𝐿ₚ = 𝐿(𝑊, 𝑏, 𝛼)

11th ICCCNT 2020


Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on May 08,2024 at 18:11:50 UTC from IEEE Xplore. Restrictions apply.
July 1-3, 2020 - IIT- Kharagpur
Kharagpur, India
IEEE - 49239

1
𝐿ₚ = ( ‖𝑊‖2 − ∑𝑚 𝑇 (2.3) ∂L
= 0 ; vₛ[𝑊 𝑇 . 𝑢ₛ + 𝑏] − 1 + ᶓₛ = 0 (2.12)
2 𝑠=1 𝛼 ₛ [vₛ[𝑊 . 𝑢ₛ + 𝑏] − 1]) ∂ᶓₒ
Partially differentiating (2.2) and (2.3), we get saddle points Substituting (2.9) (2.10) (2.11) on (2.8), we get
(w0 , b0 ,α0). 1
Lₔ(α) = ∑m
s=1 αₛ − ∑s,t αₛαₜvₛvₜuₛuₜ
∂L
= 0 ; Wₒ = αₛvₛuₛ (2.4) 2
∂Wₒ Subject to
∂L
=0 ; ∑𝑚 𝛼ₛvₛ = 0 (2.5) 0≤ 𝛼₃ ≤ C s= 1……m
∂bₒ 𝑠=1
∑𝑚𝑠=1 𝛼ₛvₛ = 0
Substituting (2.4) and (2.5) in (2.3) we get the dual form
1
(2.6) Note2: The only difference is one parameter C constraining
Lₔ(α) = ∑m
s=1 αₛ − ∑s,t αₛαₜvₛvₜuₛuₜ − 𝑏 ∑m
s=1 αₛvₛ the size of 𝛼ₛ and ᶓₛ also known as slack variables that will
2

Subject to 𝛼ₛ ≥ 0 penalize wrongly classified samples. Farther the distance of


wrongly classified sample from the hyperplane, larger will be
Now, according to (2.5) last term will become zero.(2.6) ᶓₛ value.
becomes
1
Lₔ(α) = ∑m s=1 αₛ − ∑s,t αₛαₜvₛvₜuₛuₜ
(2.7)
2 B. Kernel function
Note1: On solving (2.7) with some training data points, we get Support vector machines become even more powerful when
the values of 𝛼ₛ 𝑎𝑛𝑑 𝛼ₜ. After calculating W and b from (2.4) they are not only applicable for linear separable data instances
and (2.6) respectively and substituting in (2.1) we get our but for non linear data instances too. Here, a non linear kernel
optimal hyperplane. maps the data into a higher dimensional space which can be
A. Soft Margin then linearly separable. The idea is to replace the inner product
by a function which is non linear and this function is called
It is not likely to get an exact separation of a data with a line; a kernel.
few instances of one class may present with the elements of The embedding function ∅ that classifies point as:-
the other class. Therefore, not all classification problems are
linearly separable. Here we introduce the slack variable ᶓ₃ vₛ. [𝑊 𝑇 . ∅(𝑢ₛ) + 𝑏] = 1 − ᶓₛ ᶓₛ ≥ 0 ; s = 1, 2, 3 … … m
which measures the error. For the solution of the optimization problem, the dual
To maximize the separating margin and to minimize the formulation becomes the following quadratic problem
training error ᶓ₃. Lₔ(α) = ∑m
1 (2.13)
s=1 αₛ − ∑s,t αₛαₜvₛvₜK(uₛ, uₜ)
The primal form of the optimization is 2

1 Subject to
𝐿ₚ = ‖𝑊‖2 + 𝐶 ∑ ᶓ ₛ 0 ≤ 𝛼ₛ ≤ C; s = 1 … … m
2
Subject to: ∑𝑚𝑠=1 𝛼ₛvₛ = 0.
vₛ. [𝑊 𝑇 . 𝑢ₛ + 𝑏] ≥ 1 − ᶓₛ ;s=1,2,……m 𝑘(𝑢ₛ, 𝑢ₜ) = ∅(𝑢ₛ)𝑇 ∅(𝑢ₜ) is the kernel function.

ᶓₛ ≥ 0 ; s = 1,2,3 … … m (2.8)
Note3: for solving (2.13) we only need to know about the
Where C also known as a penalty or cost parameter balances inner product. Therefore, after choosing a kernel K and data
the tradeoff between maximizing the margin and controlling points, we map to kernel function, we calculate the inner
instances. We might have a huge slack variables thus to product of the transformed points.
normalize it we introduce 𝛼 as a Lagrangian parameter to
allow more data to lie on the wrong side.
Based on the Karush–Kuhn–Tucker KKT [5] , the equivalent C. Least-Square Formulation
dual optimization problem By solving a set of linear equations instead of quadratic
1 programming, least-square approach is implemented.
𝐿ₚ = ( ‖𝑊‖2 + 𝐶 ∑ ᶓ ₛ − ∑𝑚 𝑇
𝑠=1 𝛼 ₛ [vₛ[𝑊 . 𝑢ₛ + 𝑏] − 1 + 1 𝑐
2
𝐿ₚ = ‖𝑊‖2 + ∑ ᶓₛ2
2 2
ᶓₛ] − ∑ 𝛽ₛᶓₛ)
Subject to
𝛼ₛ𝛽ₛ ≥ 0 (2.9) vₛ[𝑊 𝑇 . ∅(𝑢ₛ) + 𝑏] = 1 − ᶓₛ; ᶓₛ ≥ 0 ; s = 1, 2, 3……m
Partially differentiating (2.8), we get saddle points (w 0, b0, α0, set of linear equations (instead of QP)
ᶓ0) [0 1𝑇 𝑏 0 (2.14)
−1 ] [𝛼 ] = [ ⃗ ]
∂L (2.10) 1 𝐾+𝐶 𝐼 𝑌
= 0 ; Wₒ = αₛvₛuₛ
∂Wₒ Where,
∂L
= 0 ; ∑𝑚 (2.11) 𝐾 = 𝑍𝑍 𝑇
∂bₒ 𝑠=1 𝛼ₛvₛ = 0

11th ICCCNT 2020


Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on May 08,2024 at 18:11:50 UTC from IEEE Xplore. Restrictions apply.
July 1-3, 2020 - IIT- Kharagpur
Kharagpur, India
IEEE - 49239

𝑍 = [∅(𝑢ₛ)𝑇 . 𝑣ₛ … … … . ∅(𝑢ₘ)𝑇 . 𝑣ₘ] 1


|𝑢ₛ = |𝑢ₛ| ∑𝑛𝑘=1(𝑢ₛ)ₖ| 𝐾〉 (4.1)
𝑌 = [𝑣₁, 𝑣₂, 𝑣₃ … … … . 𝑣ₘ]
Where the norms |𝑢ₛ| and their labels 𝑣ₛ for each 𝑠 ∈
⃗1 = [1,1,1 … … … .1] {1,2 … 𝑚} are given.
𝛼 = [𝛼₁, 𝛼₂, 𝛼₃ … … … . 𝛼ₘ]
On applying Mercer’s condition, we get
𝐾 = 𝑣ₛ𝑣ₜ∅(𝑢ₛ)∅(𝑢ₜ)
𝐾 = vₛvₜK(uₛ, uₜ)
Note4: The only difference is we need to solve only the
linear equations (2.14) in LS-SVM rather than solving
complex quadratic equations in other variant. Moreover,
solving linear equations are more efficient in quantum
computing. Fig.3 Feature map circuit

The major drawback of support vector machine is the One of the most efficient way to store and to access these
problem of overfitting. The classical approach does not data in parallel is QRAM [11]. The two phase involve in
perform well when there is large set of data moreover; it is QSVM is training and classification.
difficult to choose kernel function which can become complex
sometimes. Therefore, the quantum approach is explored to 1) Training phase:
overcome the limitation and results in achieving better Quantum SVM training phase allows to construct the kernel
performance. matrix K [10].
̂= 𝑘
Normalized kernel function 𝐾
III. QUANTUM FUNDAMENTALS 𝑡𝑟(k)

Our normal computers can understand classical bit that is it For the construction of kernel function K, first call all the
can be 0 or 1 but in quantum computing we have quantum bits training input (4.1) and then prepares the state which is
or qubits they can be 0 or 1 at the same time which is known
1
as superposition [6]. This factor gives quantum computers |𝑈〉 = ∑𝑚
𝑠=1|𝑢ₛ| |𝑠〉|𝑢ₛ〉
√𝑁ₓ
superior computing power. Many tangible objects can be used
as qubits like a single photon, nucleus and electrons.
Where,
Superposition is a phenomenon in which an object can be a 𝑁ₓ = ∑𝑚 2
(4.2)
𝑠=1|𝑢ₛ|
present in more than one state at once. The other important
principle is quantum entanglement. Suppose there are ̂
The kernel matrix 𝐾 plays an important role both in the dual
entangled particles then irrespective of distance between them formulation (2.13) and the least-square formulation (2.14). A
if we measure the particular property in one particle then at the least-square formulation of the SVM allows the use of phase
same time exactly opposite results will be shown by the other estimation and quantum matrix inversion [8][9].
particle. The goal is to classify a state by generating a quantum
A qubit or a system of qubit can change its state by unitary state|𝑏, ⃗⃗⃗⃗
𝛼〉 that describes the hyperplane with matrix inversion
transformations. Classical logic gates uses AND, NOT gate algorithm. From (2.14) we got,
similarly in quantum these unitary transformation are carried
𝑏
𝑀 ( ) = [0 1𝑇 𝑏 0
out by using quantum gates [7] which are used to build up the −1 ] [𝛼 ] = [ ⃗ ]
quantum circuit interconnected by quantum wires. These 𝛼 1 𝐾+𝐶 𝐼 𝑌
transformation made by the gates are always reversible. In order to get the quantum state we need to solve normalized
IV. QUANTUM SUPPORT VECTOR MACHINE ̂ = |𝑏, ⃗⃗⃗⃗
𝑀 𝛼〉 = |𝑣 〉
A. Least-Square Quantum Support Vector Machine Where,
𝑀
Quantum support vector machine gives the exponential ̂=
𝑀
𝑡𝑟(M)
speedup which arises due to the fast computation of the inner
product. Solving linear system is a central to many 𝑀 = 𝐺 + 𝐾 + 𝐶 −1 𝐼
departments. HHL algorithm proposed by Harrow, Hassidim, ⃗𝑇
𝐺 = [0 1 ]
and Lloyd is used to solve linear equations with exponential ⃗1 0
speedup compared to the classical algorithm. The process
involved in the HHL algorithm uses the phase estimation
We need to invert ̂ 𝑀. By using lie product formula, we get the
algorithm [8]. Phase estimation is a quantum subroutine that
helps to find out the eigenvalues of the unitary matrix which exponential as
can be applied to a quantum register as a controlled gate. 𝑒 −𝑠𝑀̂∆𝑡 = 𝑒 −𝑠 ∆𝑡𝐼/𝑡𝑟(𝑀) 𝑒 −𝑠𝐺∆𝑡/𝑡𝑟(𝑀) 𝑒 −𝑠 𝐾∆𝑡/𝑡𝑟(𝑀) + (4.3)
The input of a quantum support vector machine is M 𝑂(∆𝑡 2 )
known quantum states:-

11th ICCCNT 2020


Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on May 08,2024 at 18:11:50 UTC from IEEE Xplore. Restrictions apply.
July 1-3, 2020 - IIT- Kharagpur
Kharagpur, India
IEEE - 49239

Now the matrix G has two Eigen values +√𝑃 and -√𝑃 and that, classification can begin by the classical computer to
the corresponding Eigen states which can be useful to solve predict the labels for the test dataset. Here we use quantum
(4.3) and the result of which can be used to invert kernel computer to estimate kernel function.
̂ . Performing the controlled rotation and uncomputing
matrix 𝐾 The efficiency of SVM depends on kernel selection. Based
the Eigen value register, we get the final state. on [13] , M-band wavelet kernels within both classical and
The expansion coefficients of these states are the desired quantum versions are constructed for corresponding Support
SVM parameters: Vector Machine which turns out to give 𝑂(𝑁. log 𝑀) on
considering M training points with N dimensions.
1
|𝑏, ⃗⃗⃗⃗
𝛼〉 = (𝑏|0〉 + ∑𝑚
𝑘=1 𝛼ₖ|𝐾〉) D. Other Quantum Approaches
√𝑅
Where, Support vector machine is still facing the problem of
𝑅 = 𝑏 2 + ∑𝑚
𝑘=1 𝛼ₖ|𝐾〉 (4.4) complexity and overfitting due to the large dataset which
makes the machine non-efficient. From [14] a quantum-
inspired classical SVM algorithm inspired by quantum SVM,
2) Classification phase:
Tang’s methods [15] and Andras Gilyen et al.’s work [16] has
Now we can construct the state by using the parameters we got
been proposed. The main idea is to solve the classification
in (4.4) and also the query state. For the classification we
problem through a linear equation
perform the swap test. After that we calculate the success
probability Pr. If Pr < ½ we classify the state as +1, otherwise 𝐾𝛼 = 𝑣
-1. Where,
B. Quantum Variational Method 𝐾 = 𝑈 𝑇 . 𝑈 is the kernel matrix
From [12] we use variational circuits. The overall quantum U is the data matrix.
classifier will consist of two concatenated circuits. The first is We cannot solve the above equation directly through the
the “feature map” circuit (Fig.3). The other, combined with proposed quantum-Inspired matrix inverse algorithm [16]. The
the measurement, a hyper-plane classification in the feature row sampling and column sampling is done on data matrix U
spaces. The parameter of this second circuit is the training by using theorems in [14] which is called as the indirect
parameters of the model we consider. This second circuit is a sampling technique that is used in many algorithms
variational circuit, also referred to as short-depth quantum [15][16].This indirect sampling implements the pseudo-
circuit. A variational circuit is a hybrid quantum-classical inverse by finding an approximate decomposition of K via
circuit which consists of parameterized gates that depends on sampling on U, then inverting the singular values K.
set of parameters θ as well as a learning algorithm and
objective function. In our case, we will train the circuit to Based on [17] a hybrid quantum-classical version of LS-
correctly label the data-points. If we use some circuit SVM which proposes a simplified quantum algorithm as well
parameters to feed inputs into the circuit and compute the as sparse solution, a simplified algorithm that uses the matrix
results (outputs), we can see that a variational circuit has inversion with two continuous variables [18] and then we have
similar intuition as other supervised learning models. sparse solution for quantum LS-SVM. With the help of
Following steps summarizes the procedure: quantum matrix toolbox, including quantum principal
component analysis (QPCA)[22] and the quantum singular
• We combine both variational and feature map circuits. value threshold algorithm (QSVT)[23][24] are used to
By using the feature map function, the data input 𝑢
⃗ is compare the large dataset (encoded into quantum state) to a
mapped to a quantum state. smaller dataset.
• A short depth quantum circuit 𝑊(𝜃) is applied to the The other QSVM [19] which uses quantum annealing is
feature state. proposed where the support vectors are determined by ranking
data points based on different training instances around them.
• Measure the circuit for a binary measurement𝑍ₓ. The One of the clustering algorithm DBSCAN [20] and quantum
outcome of which will be assigned as a class label. algorithm Deutsch-Jozsa algorithm [21] is applied.
After running the same circuit on loop, we derive the
probabilities of assigning the class. V. COMPARATIVE ANALYSIS BETWEEN C-SVM AND Q-SVM
• Using SPSA optimizer, minimize the circuit. In the classical support vector machine where we have
consider N number of features that represent a single data
• Evaluate overall performance of the training set and instance and M number of training instances the calculation
update θ accordingly. are majorly dominated via kernel matrix, the dot product 𝑢ₛ𝑢ₜ
C. Quantum kernel based algorithm. in the kernel matrix takes 𝑂(𝑁) time then calculating the
whole kernel matrix has 𝑂(𝑀2 𝑁)time complexity and then
Using [12] which is only applicable to binary classification finding the optimal 𝛼ₛ values by quadratic programming
problems; a kernel matrix is estimated with the quantum takes 𝑂(𝑀3 ) Therefore, the classical support vector machine
computer by using all the training data. Then, the classical takes at least 𝑂(𝑀2 (𝑀 + 𝑁)).
computer takes this quantum kernel matrix of training data to
calculate the support vectors with the classical computer. After

11th ICCCNT 2020


Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on May 08,2024 at 18:11:50 UTC from IEEE Xplore. Restrictions apply.
July 1-3, 2020 - IIT- Kharagpur
Kharagpur, India
IEEE - 49239

However in the quantum SVM algorithm, the first [5] R. Fletcher, Practical Methods of Optimization: Volume 2 Constrained
̂ . To Optimization. New York: Wiley, 1981.
improvement done is calculating the kernel matrix 𝐾
[6] Amiri, P.K.:Quantum computers. In: IEEE Potentials, vol. 21, no. 5, pp.
calculate the entire kernel matrix we need to evaluate 𝑂(𝑀2 ) 6-9. (2003).
inner products which would give K in time 𝑂(𝑀2 log 𝑁)
[7] David P. DiVincenzo.: Quantum gates and circuits. IBM Research
using quantum inner product. Because quantum RAM allows Division, Thomas J. Watson Research Center, Yorktown Heights, NY
for 𝑂(log 𝑀𝑁) access time, K takes𝑂(log 𝑀𝑁). Instead of 10598 USA. arXiv:quant-ph/9705009.
using QP as in the classical algorithm we evaluate inverse of [8] A. W. Harrow, A. Hassidim, and S. Lloyd, Phys. Rev. Lett. 103,150502
M and using the lie product formula we got (4.3). G and 𝐶 −1 (2009).
are sparse and k is not, we must have two different techniques [9] N. Wiebe, D. Braun, and S. Lloyd, Phys. Rev. Lett. 109, 050505 (2012).
called quantum self analysis which takes 𝑂(log 𝑀𝑁) On [10] Seth Lloyd, Masoud Mohseni, and Patrick Rebentrost. Quantum
combining, we get total 𝑂(log 𝑀𝑁) to solve the least-square algorithms for supervised and unsupervised machine learning. arXiv
support vector machine. preprint arXiv: 1307.0411, 2013.
[11] Vittorio Giovannetti, Seth Lloyd, and Lorenzo Maccone. Quantum
Predicting using the classical model takes 𝑂(𝑁) times by random access memory. 100:160501, 04, 2008..
direct calculation. In the quantum model, classification [12] Havenstein, Christopher; Thomas, Damarcus; and Chandrasekaran,
requires iterating 𝑂(𝑃(𝑝 − 1)/𝜖 2 ) times where P is the Swami "Comparisons of Performance between Quantum and Classical
success probability and accuracy 𝜖. In general classification, Machine Learning," SMU Data Science Review: Vol. 1 : No. 4 , Article
11,2018.
the quantum SVM is faster when n is large but classical SVM
[13] Runlang Ma, Tong Liu, Shangyi Li, Dr. Xiaodi Wang,”M-band
may be preferable if accuracy is important. Wavelet Kernels for Classical and Quantum SVM”, IEEE International
Conference of Safety Produce Informatization,2018.
VI. CONCLUSION [14] Chen Ding, Tian-Yi Bao, and He-Liang Huang,” Quantum-Inspired
The article describes the classical and different quantum Support Vector Machine “,arXiv:1906.08902v2 [cs.LG] , 2019.
approaches of support vector machine. A least-square [15] E. Tang, “A quantum-inspired classical algorithm for recommendation
formulation of the support vector machine uses the phase system”,51st Annual ACM SIGST Symposium Thoery Computing
vol.25, New York,NY,USA:ACM, June 2019,pp.217-228.
estimation and the quantum matrix inversion algorithm and
realized that for the data classification the support vector [16] A. Gilyen, S. Llyod, and E. Tang, “Quantum-inspired low-rank
stochastic regression with logarithmic dependence on the
machine can be implemented with logarithmic time compared dimension”,2018.
to the classical algorithms. Quantum variational method that [17] Jie Lin, Dan-Bo Zhang,, Shuo Zhang, Xiang Wang, Tan Li, and
uses two circuits which on combined speedup the Wansu Bao ,” Quantum-enhanced least-square support vector machine:
classification performance, Kernel approach in which the simplified quantum algorithm and sparse solutions,
kernel matrix is calculated by using quantum computer. Many arXiv:1908.01462v1,2019.
training algorithms like quantum-inspired SVM that is [18] S. Lloyd, in Quantum information with continuous variables, pp. 37–
implemented using indirect sampling technique, a hybrid 45,Springer(2003).
quantum-classical version which also gives the sparse solution [19] Koudai Shiba, Katsuyoshi Sakamoto, Tomah Sogabe,” Variational
Quantum Support Vector Machine based on Deutsch-Jozsa Ranking”,
of LS-SVM and many more quantum approaches that are yet ISSN 2186-5140 Volume 9, Number 1, pages 63–68, January 2020.
to explore which can be useful for training and classification. [20] Martin Ester, Hans-Peter Kriegel, Jörg Sander and Xiaowei Xu, “A
By further exploring the algorithms and the error correction density-based algorithm for discovering clusters in large spatial
processes we can focus on reducing the contents in the data databases with noise,” 2nd International Conference on Knowledge
matrix which can contribute in reducing the time complexity Discovery and Data Mining, pp. 226-231, 1996.
when implemented using quantum fundamentals and [21] David Deutsch and Richard Jozsa, “Rapid solution of problems by
algorithms. quantum computation,” Proceedings of the Royal Society of London A,
439, 553, 1992.
[22] S. Lloyd, M. Mohseni, and P. Rebentrost, Nature Physics 10, 631(2014).
References
[23] B. Duan, J. Yuan, Y. Liu, and D. Li, Physical Review A 96, 032301
[1] Peter Wittek “Quantum mechanics and Quantum computing” in (2017).
Quantum Machine learning, 1st ed.,pp. 25-52.
[24] B. Duan, J. Yuan, Y. Liu, and D. Li, Physical Review A 98, 012308
[2] Shuyung Yang, Min Wang, Licheng jiao, A Quantum Particle Swarm
(2018).
Optimization, IEEE (2004).
[3] Kerstin Beer, Dmytro Bondarenko, Terry Farrelly, Tobias J. Osborne,
Robert Salzmann,and Ramona Wolf, “Training deep Quantum Neural
Network”,Institute of Theoretische Physic, Leibniz University at
Hannover,Germany
[4] V. Vapnik. The Nature of Statistical Learning Theory. NY: Springer-
Verlag. (1995), 2nd ed, pp. 156-162.

11th ICCCNT 2020


Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on May 08,2024 at 18:11:50 UTC from IEEE Xplore. Restrictions apply.
July 1-3, 2020 - IIT- Kharagpur
Kharagpur, India

You might also like