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

net/publication/319251548

Quantum Algorithm for K-Nearest Neighbors Classification Based on the

Metric of Hamming Distance

Article in International Journal of Theoretical Physics · November 2017

DOI: 10.1007/s10773-017-3514-4

CITATIONS READS

112 4,823

5 authors, including:

Yue Ruan Heng Liu

Anhui University of Technology Anhui University of Technology
20 PUBLICATIONS 294 CITATIONS 51 PUBLICATIONS 707 CITATIONS

SEE PROFILE SEE PROFILE

Jianing Tan Xi Li
Southeast University Southeast University
9 PUBLICATIONS 155 CITATIONS 11 PUBLICATIONS 154 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Xi Li on 06 March 2018.

The user has requested enhancement of the downloaded file.

Int J Theor Phys
DOI 10.1007/s10773-017-3514-4

Quantum Algorithm for K-Nearest Neighbors

Classification Based on the Metric of Hamming Distance

Yue Ruan1,2 · Xiling Xue1 · Heng Liu1 · Jianing Tan3 ·

Xi Li3

Received: 12 May 2017 / Accepted: 16 August 2017

© Springer Science+Business Media, LLC 2017

Abstract K-nearest neighbors (KNN) algorithm is a common algorithm used for classifi-
cation, and also a sub-routine in various complicated machine learning tasks. In this paper,
we presented a quantum algorithm (QKNN) for implementing this algorithm based on the
metric of Hamming distance. We put forward a quantum circuit for computing Hamming
distance between testing sample and each feature vector in the training set. Taking advantage
of this method, we realized a good analog for classical KNN algorithm by setting a distance
threshold value t to select k − nearest neighbors. As a result, QKNN achieves O(n3 ) per-
formance which is only relevant to the dimension of feature vectors and high classification
accuracy, outperforms Llyod’s algorithm (Lloyd et al. 2013) and Wiebe’s algorithm (Wiebe
et al. 2014).

 Yue Ruan
yue ruan@163.com
Xiling Xue
stmxue@163.com
Heng Liu
hengliusky@aliyun.com
Jianing Tan
jn tan@163.com
Xi Li
230169107@seu.edu.cn

1 School of Computer Science and Technology, Anhui University of Technology, Maanshan 243005,
China
2 Key Laboratory of Computer Network and Information Integration, Southeast University,
Ministry of Education, Nanjing 210096, China
3 School of Computer Science and Engineering, Southeast University, Nanjing 210096, China
 Int J Theor Phys

Keywords Quantum algorithm · K-nearest neighbors algorithm · Hamming distance ·

Classification · Machine learning

1 Introduction

Classification (the broader understanding of it should include clustering), has been studied
as a core issue in machine learning for many decades. In recent years, the research of such
issues meets real challenges for the larger and larger dataset to be processed, which is known
as the term of Big Data. The main challenge faced is the computational inefficiency of
classical machine learning algorithms processing data of such huge volume. The common
method to respond to this challenge is to layout these algorithms in Cloud [1]. With the
aid of computing power of Cloud, such as Map-Reduce framework of google [2], these
problems have been resolved to some extent. But, the storage and management of data in
distributed heterogeneous networks have also brought about some new problems, such as
the modification of canonical algorithms applied for the cloud framework[3], the reliability
and error tolerance of cloud system [4], and the safety of privacy data in cloud computing
[5–8]. All these problems are not easy tackled and become hot topics in computer science
today.
Are there other approaches to deal with this challenge? Physicists present another per-
spective. By exploiting the superposition and entanglement properties of quantum state,
quantum information processing technology can compress the representation space of data
exponentially and speed up classical algorithms inherently[9]. The idea of doing machine
learning task fused into quantum computational properties seems to be promising. Bene-
fited from the works of Llyod [10–12], Wiebe [13, 14] and other pioneers [15–22] , this idea
attracts more attention and gradually develops into a crossover research field — quantum
machine learning.
In this paper, we present a quantum version for a concrete machine learning algorithm
— k-nearst neighbors algorithm (KNN). KNN algorithm is an important and basic algo-
rithm for classification, and also a sub-routine in various complicated machine learning
algorithm.
The following parts are organized as follows: In Section 2, we will depict some realted
works firstly, involving the classical KNN algorithm, the virtue of quantum machine learn-
ing, esp. on the high efficiency of distance computing, and the trick of computing Hamming
distance. And then, we will describe the proposed quantum KNN algorithm (QKNN) in
Section 3. In Section 4, we will discuss the time performance and classification accuracy of
QKNN compared to analogous works presented by Llyod [11] and Wiebe [13]. Finally, we
will draw the conclusion and give our insight of quantum machine learning.

2 Related Works

2.1 K-Nearest Neighbors Algorithm

KNN is a commonly used algorithm for supervised machine learning. The idea of its work-
ing mechanism is very simple: Give a testing sample, find its k nearest neighbors based on
some distance metric, and then determine its category according to the information of these
Int J Theor Phys

neighbors. As a general rule, the algorithm uses “majority voting” to the end. That is, the
testing sample is labeled as the leading category tag of its k nearest neighbors.
Figure 1 illustrates the principle of this algorithm. When k = 1, the testing sample
(indicated as a question mark) is labeled as the category “blue star”. When k = 3 and k = 5,
the testing sample is labeled as the majority category “red triangle”. Obviously, k is a very
important factor. The classification result is different as we set k to different value. We have
to note that if we set k = 1, KNN will be degenerated into the closest neighbor algorithm.
This simplified assumption does not work well in practice, esp. in the Big Data scenario.
Because in large data sets of the real world, unavoidable outliers (for example, polluted
data) may lead to faulty judgment. In contrast, “majority voting” tends to be effective from
a statistical view.

2.2 The Virtue of Quantum Machine Learning

The virtues of Machine Learning fused into quantum properties (Quantum Machine
Learning, QML) are embodied in the following two aspects:
One is the storage scale (represention space) can be reduced exponentially by exploiting
the superposition properity of quantum states. For example, a n qubit state |φ1 φ2 · · · φn  can
be written as
n −1
2 
|φ1 φ2 · · · φn  = ci |i st. |ci |2 = 1 (1)
i=0
This equation shows that, in the quantum computer, all the binary number from the set
{0, 1, ..., i, ..., 2n − 1} coexists in a n qubit quantum register with the corresponding proba-
bility |ci |2 . However, in a classical computer, a n bit register can just store one number from
the set {0, 1, ..., 2n − 1} exclusively.
The other virtue is the acceleration of algorithm execution speed by quantum parallelism.
Also inherited from the superposition property of quantum state, the unitary evolution of
the closed quantum systems can operate on each item of a quantum superposition state

Fig. 1 The illustration of KNN algorithm

 Int J Theor Phys

simultaneously. As shown in (2), supposing unitary operator Uf realizes the computing

function f (x), then Uf has accomplished all the computing task f (x) by inputting each
variable x from |00 · · · 0 to |11 · · · 1 simultaneously. In classical algorithm, the computing
of f (x) from the same scale of input variables x needs to be done by 2n cycles or to be done
by the parallel work of 2n CPUs.
2 −1 n 2 −1 n 2 −1 n
1  1  1 
Uf ( √ |x) = √ Uf |x = √ |f (x) (2)
2n x=0 2n x=0 2n x=0

Obviously, the reduction of data space and the acceleration of algorithm execution speed
appear to be especially important and meaningful in Big Data scenario. But we have to note,
in general, there is no effective manner to retrieve all the f (x) after one run of the quantum
algorithm. The only way to get useful information from a quantum state is measurement,
which will lead to the collapse of quantum state and loss of most execution results of the
quantum algorithm. If we want to get all the f (x), we must run the algorithm many times
and do measurement many times. So, how to retrieve useful information from a quantum
state with high performance is a key issue and needs real ingenuity.
Buhrman takes good advantage of quantum parallelism, proposes a quantum computing
trick of calculating the distance of two vectors with high performance [23]. Figure 2 illus-
trates this trick, where |0 is an auxiliary qubit, through left H gate, it will be changed to
√1 (|0 + |1). Then under the control of |1, the circuit will swap two vectors |x and |y,
2
i.e. |xy → |yx. Finally, we can get (3) at the right end of the circuit:
1 1
|0anc |x|y → |0anc  (|xy + |yx) + |1anc  (|xy − |yx) (3)
2 2

If we measure auxiliary qubit alone, then the probability of final state in the ground state
|0 is:
1 1
P (|0anc ) = + |x|y|2 (4)
2 2
|x|y| is named as fidelity in quantum information theory and also named as cosine dis-
tance in classical machine learning. Obviously, if |x and |y have maximum distance, i.e.
orthogonality, this probability is 12 ; if |x and |y has minimum distance, i.e. overlap, this
probability is 1. We should notice that the estimation of this probability is irrelevant to the
dimension of the vectors. The higher the dimension of the vector, the higher efficiency of
the quantum solution is. Lloyd points out even the cost of quantum state preparation were
considered, the performance of computing distance by this quantum trick would be more
efficient than classical manner [11].

Fig. 2 Swap test for distance

computing
Int J Theor Phys

Based on this trick, popular

√ distance metric Euclidean distance can be easliy computed
(Euclidean distance = (2 − 2|x|y|). And because distance metric is the core of sim-
ilarity measurement, this trick is used widely in quantum machine learning. Wiebe takes
advantage of this trick, and also uses two existing algorithm [24, 25] to implement a simpli-
fied version for KNN algorithm — nearest neighbor algorithm [13]. Lloyd also use this trick
to implement a variant KNN algorithm — nearest centroid algorithm [11]. In the quantum
support vector machine realized by Rebentrost et al., they also use this trick to transform
the computing of kernel matrix to computing the partial trace of density matrix [21].

2.3 Computing Hamming Distance

Inspired by the aforementioned trick and the work of Schuld [22], we design a quantum
KNN algorithm based on the metric of Hamming distance. In order to articulate our algo-
rithm, as the background, we review the definition of Hamming distance and introduce a
quantum circuit adopted to comupte this distance here.
Hamming distance is defined as counting the number of positions at which the cor-
responding symbols of two bit vectors of equal length are different. For example, the
Hamming distance: 0110 ↔ 0001 has a distance of 3; while the Hamming distance:
0110 ↔ 1110 has a distance of 1. This may look unnatural in determining the similarity of
two natural feature vectors. But in practice, Hamming distance is used widely in document
classification, image classification, etc. [26–28]. Mapping a natural vector to a bit vector by
well-defined hash function, simple KNN classifiers in Hamming space are competitive with
sophisticated discriminative classifiers, including SVMs and neural networks [27].
In quantum machine learning algorithm, if we map the features of object to the ground
quantum states in Hilbert space, then it is easier to select k nearest neighbors of the test-
ing sample by computing Hamming distance among these states, which will improve the
performance further by avoiding some time-consuming operations in manipulating general
quantum states such as tomography or phase estimation. This assertion will be externalized
after we describe the whole algorithm. Here, we first depict an incremental circuit, which is
adopted as the core module to compute Hamming distance in our algorithm.
The circuit is proposed by Kaye [29] (shown as Fig. 3). It realizes the incrementation
operation of a number a, i.e. a = a + 1. The number a is described as a[0..n-1]. And incre-
mentation by 1 means the flipping from the least significant qubit. If a[i] flips from 1 to 0,

Fig. 3 Quantum a + 1 circuit

 Int J Theor Phys

the addition would continue. If a[i] flips from 0 to 1, which means no carry qubit is pro-
duced, the addition should stop. The ancillary qubit in the circuit can be viewed as a “flag”
which signals the first time a qubit flips from 0 to 1. It should be reset to 1 for the next run
of addition. The work flow of this circuit can be depicted as the following pseudo-code:

i = 0;
Do
if a[i]==1 then {
a[i]: 1 → 0;
i++;
}
else
a[i]: 0 → 1;
Until (a[i]: 0 → 1)

3 Quantum K-Nearest Neighbors Algorithm

With the knowledge of the previous section, we can describe the QKNN algorithm now. This
algorithm is intended to determine a testing sample belongs to which class in the training
set.

3.1 Preliminary Steps

Step 1: Extract the feature of the training set and store them as bit vectors by exploiting
the method of Ref [27],
Step 2: Map and parepare these bit vectors to quantum ground state straightforwardly, i.e.
0 → |0 and 1 → |1.
After these two steps, the training set is indicated by N feature vectors
|v p , p = 1, ..., N and the corresponding class cp ∈ {1, ..., l}, which can be
written as {|v1 ...vn , cp } ∈ H⊗n
p p
2 ⊗ Hl .
Step 3: Construct a training set superposition.

1  p p p
|T  = √ |v1 ...vn , c  (5)
N p

3.2 Main Steps

Step 1: Input a testing sample, taking the same method to transform it to a quantum state
|x (a normalized n-dimensional feature vector |x1 ...xn  ).

Step 2: Prepare the unclassified quantum state |x1 ...xn  in the first register, prepare the
training set |T  in the second register, prepare an ancillary qubit |0 in the last
register. The result can be written as |φ0  .

1  p p
|φ0  = √ |x1 ...xn ; v1 ...vn , cp ; 0 (6)
N p
Int J Theor Phys

p p
Step 3: Record the difference between |x1 ...xn  and each |v1 ...vn  in the training set,
p p
store the result |d1 ...dn  in the first register and reverse the value. For example, if
p p p p
|x1 ...xn  = |0010, |v1 ...vn  = |1010, then the final result |d1 ...dn  = |0111.
The reason that we reverse the value is just a mathematical trick. Its function will
be demonstrated in the following steps.
 p 1  p p p p p
|φ2  = X(xk )CN OT (xk , vk )|φ1  = √ |d1 ...dn ; v1 ...vn , c ; 0 (7)
k
N p
Where, CNOT(a, b)-gate overwrites the first entry a with 0 if a = b and else with
1. X gate is used to reverse the value.
p p
Step 4: In this step, we will compute Hamming distance in terms of |d1 ...dn , and label
the k-nearest neighbors of the testing sample (modify the corresponding ancillary
qubit, |0 → |1) according to a distance threshold value t. The operation of this
step can be defined as a certain unitary operation U , which realizes:
⎛ ⎞
1 ⎝ p p p p p  p p p p
|φ3  = U |φ2  = √ |d1 ...dn ; v1 ...vn , c ; 1+ |d1 ...dn ;v1 ...vn , cp ; 0⎠
N p∈Ω p ∈Ω
(8)
Where set Ω contains certain indexes p which indicates the Hamming distance
between |x and the pth sample in the training set ≤ t.
p
In step 3, each di has recorded the difference
 p at the corresponding position i. Hence,
the Hamming distance can be computed as i di . As we recall, In Section 2, we introduce
p
the quantum circuit of a = a + 1. Then the key step for adding di can be realized by taking
advantage of this circuit as Fig 4a, which isreduced as Fig 4b.
p
Taking inCk as the core module, then i di can be gotten by invoking this module n
p
times. As we also recall, in step 3, the value of di is reversed, so H amming distance ≤ t
can be depicted as follows:  p
di ≥ n − t (9)
i
Suppose 2k−1 ≤ n ≤ 2k , if we set a variable l = 2k − n, then the condition
H amming distance ≤ t can be finally derived as:
 p  p
di + l ≥ n + l − t =⇒ di + l + t ≥ 2k (10)
i i
This equation means that if we set initial a = l + t, then the condition  p of
H amming distance ≤ t can be determined by whether the addition of i di + a
overflow or not. Once the addition is done, we select the logt most significant qubits
and use quantum OR gate to get the signal CON D p , which indicates the condition of
H amming distance ≤ t. The quantum OR gate is shown in Fig. 5 and the overall circuit
is shown in Fig. 6.
Step 6: Define a projection operator Γ = 1 ⊗ |11|, apply it on |φ3 , and renormalize the
result, we can get:
 p p p p 
|φ4  = Γ |φ3  = α |d1 ...dn ; v1 ...vn , cp ; 1 st. |α|2 = 1 (11)
p∈Ω

α is the renormalized amplitude of each component in |φ4 . Their values are same
because projection does not change the uniform superposition property of |φ3 .
 Int J Theor Phys

b
Fig. 4 The quantum circuit of a + di

Now, |φ4  is composed of |v p  whose distance are no more than t to the testing sample
|x. Measure cp alone, we can get the category to which |x belongs.

4 Discussion

4.1 Performance Analysis

This algorithm includes two parts: preliminary steps and main steps. In preliminary steps,
because each sample in the training set should be extracted features and mapped to quantum
state, the time performance of these steps is positively correlated with the number of samples
in the traning set, at least O(N ). But it should be noted the aim of these steps is to generate
the training set superpostion, which does not change in the main steps. That means we
just execute these steps once before the first run of this algorithm. So, the time cost of
preliminary steps can be ignored with the repeated execution of the algorithm.
In the main part, each step is executed in quantum parallelism, so the time cost is irrel-
evant to the training set size. It is only relevant to the dimension of the feature vectors n.
In these steps, the step 4 computing Hamming distance is the most time-consuming oper-
ation. The cost of the core module (Fig. 4) is depicted as the following equation, which is
measured by the number of “elementary gates” {N OT , CN OT , T off oli} [29].
⎧
⎨ 1, k=1
the cost of inCk = 10, k=2 (12)
⎩ 2
2k + k − 5, k ≥ 3
Int J Theor Phys

a
Fig. 5 The quantum OR circuit

In Fig. 6, the sub-circuit that generates the logt largest qubits invokes module inCk n
times. The sub-circuit that generates CON D p has logt + 1 NOT gates and logt − 1
Toffoli gates. Hence, the cost of the total circuit is:

n(2n2 + n − 5) + ( logt + 1) + ( logt − 1) (13)

The last step is to determine the category to which the testing sample belongs by measur-
ing the cp alone. As we know, |φ4  has k items (k-nearest neighbors) and k  N in general.
So the time cost can be ignored compared to big N . The total time cost of this algorithm is
∼ O(n3 ).
Compared to Lloyd algorithm [11] Wiebe’s algorithm [13], because they use similar pro-
cesses like quantum amplitude estimation or Grover algorithm to get the final result [24, 25,
30], the
√ time cost is relevant to the training set size N . The former is O(logN n), the latter
is O( N logN ) (require feature vectors are sparse). In Big Data scenario, the dimension n
is far less than the dataset size N . Hence, Our algorithm has a huge performance benefit.

Fig. 6 The quantum circuit to generate CON D p

 Int J Theor Phys

4.2 Classification Accuracy

Whether Lloyd algorithm [11] or Wiebe’s algorithm [13], they may both lead to inaccurate
classification for the simplification of the application scenario. Wiebe’s algorithm simplifies
the classification of the testing sample only in terms of its nearest neighbor. As Fig. 1 shows,
the nearest neighbor “blue star” is an outlier indeed. In such case, Wiebe’s algorithm will
lead to an error. Lloyd algorithm uses the distance to nearest-centroid (central vector of the
cluster) to determine the category, but nearest-centroid may distort the classification result.
Imagine that the cluster {A} is dense but cluster {B} is sparse. Then even if |x −mean(A)| ≤
|x − mean(B)|, it may be much more likely that the testing vector x should be assigned to
B because the probability of a large deviation from the centroid is much greater for {B} than
{A}.
Different from the aforementioned two algorithms, QKNN is a complete analog of the
original KNN algorithm. To illustrate the point, we reduce the final state |φ4  to a simple
form by ignoring other parts:

|φ4  = α1 |v 1 , c1  + α2 |v 2 , c2  + ... + αk |v k , ck  |αi |2 = 1 (14)

The remaining work is to determine which ci is the major category amongest  k nearest
neighbors. Measuring ci alone, the probabilty of getting category p is i∈{ci =p} |αi |2 .
Obviously, amongest k − nearest neighbors, more |v i  belongs to category p, bigger this
probability is. So, the measurement probability of getting the final classification result is a
good analog for “majority voting” of original KNN algorithm.
The above discussion shows that QKNN should be having a higher classification accu-
racy. This assertion is confirmed on the numerical experiments on real dataset MNIST http://
yann.lecun.com/exdb/mnist/. The MNIST digit database is a benchmark image dataset of
ten handwritten digits (0 to 9). To QKNN, we use binary hash function presented in Ref.
[27] to map raw MNIST images to 64-bit codes and set k = 20 (20 nearest neighbors) to
classify handwritten digits. Then, we select 10%, 20%, . . . , 90% images as training set and

MNIST digits database

1.1
QKNN
Centroid
1.05 QNN

0.95
Accuracy

0.9

0.85

0.8

0.75
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fraction of Training Data

Fig. 7 The classification accuracy of QKNN, Centroid and QNN

Int J Theor Phys

the rest images as testing images. From the results as Fig. 7 shows, QKNN outperforms
Centroid and QNN as we estimated.

5 Conclusion

In this paper, we presented a quantum algorithm for performing k-nearest neighbors

(QKNN) classification based on the metric of Hamming distance. First, we developed a
method for computing Hamming distance inspired by the quantum circuit presented by
Kaye. And owing to this method, the performance of QKNN can be highly improved in
Big Data scenario because the time cost is only relevant to the dimension of feature vectors
and is irrelevant of dataset size. Second, QKNN is a good analog for classical KNN algo-
rithm, which avoid the defect of the simplified assumption of the task of classification. In
summary, QKNN outperforms Centroid and QNN on time performance and classification
accuracy.
Quantum machine learning presents a new perspective, sheds light on the new possibility
of doing classification or other traditional machine learning tasks with high performance.
QKNN is a step toward this approach. Beyond computational speedups, a more interesting
question is whether quantum computation allows new classes of learning algorithms that
do not have natural classical analogs. The answer to this question may reveal potentials
and limitations of quantum physics on the ability to learn from data and bring wider/deeper
application of quantum machine learning.

Acknowledgements This work is supported by the National Natural Science Foundation of China (Grant
No. 61170321,61502101), Natural Science Foundation of Jiangsu Province, China (Grant No. BK20140651),
Natural Science Foundation of Anhui Province, China (Grant No. 1608085MF129, 1708085MF162), Foun-
dation for Natural Science Major Program of Education Bureau of Anhui Province (Grant No. KJ2015ZD09)
and the open fund of Key Laboratory of Computer Network and Information Integration in Southeast
University, Ministry of Education, China (Grant No. K93-9-2015-10C).

References

1. Armbrust, M., Fox, A., Griffith, R., et al.: A view of cloud computing. Commun. ACM. 53(4), 50–58
(2010)
2. Schölkopf, B., Platt, J., Hofmann, T.: Map-Reduce for machine learning on multicore. In: Proceedings of
the 2006 Conference Advances in Neural Information Processing Systems 19, pp. 281–288. MIT Press
(2007)
3. Low, Y., Bickson, D., Gonzalez, J., et al.: Distributed GraphLab: a framework for machine learning and
data mining in the cloud. Proceedings of the VLDB Endowment 5(8), 716–727 (2012)
4. Liu, Q., Cai, W., Shen, J., et al.: A speculative approach to spatial-temporal efficiency with multi-
objective optimization in a heterogeneous cloud environment. Security Commun. Netw. 9(17), 4002–
4012 (2016)
5. Xia, Z., Wang, X., Zhang, L., et al.: A privacy-preserving and copy-deterrence content-based image
retrieval scheme in cloud computing. IEEE Trans. Inf. Forensics Secur. 11(11), 2594–2608 (2016)
6. Fu, Z., Ren, K., Shu, J., et al.: Enabling personalized search over encrypted outsourced data with
efficiency improvement. IEEE Trans. Parallel Distrib. Syst. 27(9), 2546–2559 (2016)
7. Fu, Z., Sun, X., Liu, Q., et al.: Achieving efficient cloud search services: multi-keyword ranked search
over encrypted cloud data supporting parallel computing. IEICE Trans. Commun. E98-B(1), 190–200
(2015)
8. Xia, Z., Wang, X., Sun, X., et al.: A secure and dynamic multi-keyword ranked search scheme over
encrypted cloud data. IEEE Trans. Parallel Distrib. Syst. 27(2), 340–352 (2015)
 Int J Theor Phys

9. Nielsen, M.A., Chuang, I.L.: Quantum computation and quantum information [M]. Cambridge Univer-
sity Press, Cambridge (2000)
10. Lloyd, S., Mohseni, M., Rebentrost, P.: Quantum principal component analysis. Nature Physics. 10(9),
631–633 (2014)
11. Lloyd, S., Mohseni, M., Rebentrost, P.: Quantum algorithms for supervised and unsupervised machine
learning. arXiv:1307.0411 (2013)
12. Lloyd, S., Garnerone, S., Zanardi, P.: Quantum algorithms for topological and geometric analysis of data.
Nat. Commun. 7, 10138 (2016)
13. Wiebe, N., Kapoor, A., Svore, K.: Quantum nearestneighbor algorithms for machine learning.
arXiv:1401.2142 (2014)
14. Wiebe, N., Granade, C., Ferrie, C., et al.: Quantum Hamiltonian learning using imperfect quantum
resources. Phys. Rev. A 89(4), 042314 (2014)
15. Trugenberger, D.A.: Quantum pattern recognition. Quantum Inf. Process. 1(6), 471–493 (2002)
16. Aı̈meur, E., Brassard, G., Gambs, S.: Machine learning in a quantum world. In: Advances in Artificial
Intelligence Lecture Notes in Computer Science, vol. 4013, pp. 431–442 (2006)
17. Hentschel, A., Sanders, B.C.: Machine learning for precise quantum measurement. Phys. Rev. Lett.
104(6), 063603 (2010)
18. Gammelmark, S., Mølmer, K.: Quantum learning by measurement and feedback. New J. Phys. 11,
033017 (2009)
19. Bisio, A., Chiribella, G., D’Ariano, G.M., et al.: Optimal quantum learning of a unitary transformation.
Phys. Rev. A 81(3), 032324 (2010)
20. Lu, S., Braunstein, S.L.: Quantum decision tree classifier. Quantum Inf. Process. 13(3), 757–770 (2014)
21. Rebentrost, P., Mohseni, M., Lloyd, S.: Quantum support vector machine for big data classification.
Phys. Rev. Lett. 113(13), 130503 (2014)
22. Schuld, M., Sinayskiy, I., Petruccione, F.: Quantum computing for pattern classification. In: Proceedings
13th Pacific Rim International Conference on Artificial Intelligence (PRICAI), pp. 208–220 (2014)
23. Buhrman, H., Cleve, R., Watrous, J., et al.: Quantum fingerprinting. Phys. Rev. Lett. 87(16), 167902
(2001)
24. Brassard, G., Høyer, P., Mosca, M., et al.: Quantum amplitude amplification and estimation.
arXiv:0005055[quant-ph/] (2000)
25. Dürr C., Høyer, P.: A quantum algorithm for finding the minimum. arXiv:9607014[quant-ph/] (1996)
26. Manevitz, L.M., Yousef, M.: One-class SVMs for document classification. J. Mach. Learn. Res. 2(12),
139–154 (2001)
27. Norouzi, M., Fleet, D.J., Salakhutdinov R.R.: Hamming distance metric learning [C]. In: Advances in
Neural Information Processing Systems (NIPS). 061–1069 (2012)
28. Rai, H., Yadav, A.: Iris recognition using combined support vector machine and Hamming distance
approach. Expert Syst. Appl. 41(2), 588–593 (2014)
29. Kaye, P.: Reversible addition circuit using one ancillary bit with application to quantum computing.
arXiv:0408173v2[quant-ph] (2004)
30. Grover, L.K.: Quantum mechanics helps in searching for a needle in a haystack. Phys. Rev. Lett. 79(2),
325–328 (1997)

View publication stats