Quantum computation is a new computing approach based on the laws of quantum mechanics, which is a very hot topic at the frontiers of computing world today. By carefully exploiting the unique features of quantum states, quantum computers can efficiently solve some problems that are believed to be hard for classical computers.

There are two well-known quantum algorithms that demonstrate the advantage of quantum computation over classical approach. Shor’s algorithm for finding the prime factors of a number runs in polynomial time while any classical similar algorithms run in exponential times. Grover devised a quantum algorithm for the problem of unstructured search which achieves a quadratic speed up over the classical algorithms. Not available in classical computing, superposition, entanglement and interference of quantum states employed in quantum computing are generally considered as resources for this speed up. Furthermore, quantum computers can perform simulations of another quantum system such as protein folding that is beyond the capability of any classical computers.

IBM recently released the Quantum Experience to allow users to access a five qubit quantum processor, with services including circuit design, simulation, testing and actual computation on a physical device [1] . It offers researchers great opportunities to test their quantum theories and algorithms on actual quantum hardware. In today’s data rich society, the growth rate of data generated by different devices goes up way faster than that of the computer processors’ speed. As such quantum computation provides a new computing paradigm to process the enormous data and information we are facing today.

It is true that a quantum computer using quantum laws can process many quantum states simultaneously via quantum parallelism, say it can compute f ( x ) through evaluating the function for multiple values of x simultaneously. But when we need to measure the results of the computation, we can only get one result and sometimes this result is even nondeterministic. Therefore, to get a classical result, we have to run a quantum algorithm multiple times in some cases. In this sense, to produce the same classical result, an algorithm that requires less number of runs the better. In particularly, for a quantum classifier, we are not only interested in the classification accuracy, but also less number of runs (or shots in IBM’s term) of the algorithm.

Machine learning is generally categorized into three classes: supervised learning, unsupervised learning, and reinforcement learning. In supervised learning, there are classification and regression problems. Quantum advantage has long been known for problems such as factoring, search, and simulation. Only recently have quantum algorithms begun to emerge that demonstrate their quantum speedups in machine learning. These quantum machine learning algorithms have to address the basic techniques such as how to compute the distance among the data points and inner product of two vectors. These computing tasks are easy for classical computes because they can measure the computed terms anytime during the whole process, but quantum computers usually can only take direct measurements at the end of the process. It is no easy task for a quantum algorithm to be creative in exploiting measurement in a non-trivial way during the computing process.

The work in [2] was to design a quantum circuit for a distance based binary classifier runnable on IBM’s 5Q quantum computer. The quantum advantage of this classifier is the ability to compute the distance of the test data point to all the training data points simultaneously. In contrast, a major drawback of classical distance based classifier is that it can be computationally expensive. Other quantum machine learning algorithms can be seen in [3] - [13] .

This classifier was tested on iris dataset and a circles data set. The numerical results were summarized in two tables [2] . We thought it might shed more light on the nature of this quantum classifier if some visualization of its working can be seen. More specifically how the distance of the data points influences the prediction probabilities caused by the stochastic nature of quantum computing. The second thought is how the swap operation in the circuit (Figure 1) has a toll on the prediction accuracy as this operation is necessary in quantum circuit design to meet certain hardware constraints. Our final thought is how to extend the quantum binary classifier to a multiclass classifier, a further step to illustrate the potential of quantum machine learning. The purposes of our study entail the use of IBM’s quantum simulator instead of its real quantum device as we need to use more than five qubits.

This paper reports our empirical analysis of a quantum classifier implemented on IBM’s 5Q computer. We will briefly present the distance based quantum classifier created in [2] . Quantum computing requires the data be encoded in quantum states for storage, transformation, and processing. In the context of the analysis of classical data, one method is to encode the coordinates of a classical data point as amplitudes of the quantum state, adopted by [2] . Another straight forward encoding method is to encode one classical bit of information into a qubit. The advantage of amplitude encoding is its compactness to represent real numbers.

Let N = 2 n and X = ( x 0 , x 1 ⋯ , x N − 1 ) ∈ R N is a unit vector. A dataset of size M can be denoted by D = { ( X 0 , y 0 ) , ( X 1 , y 1 ) , ( X 2 , y 2 ) , ⋯ , ( X M − 1 , y M − 1 ) } , where X m ∈ R N represents a data point and y m represents its class label. For a binary classification, y m ∈ { 0 , 1 } . The amplitude encoding of X is | ψ X 〉 = ∑ i = 0 N − 1 x i | i 〉 . The first step of the quantum classifier is, through some quantum operations (see the quantum circuit in Figure 1), to transform the dataset D together with a test point as:

| D 〉 = 1 Z 1 ∑ m = 0 M − 1 | m 〉 ( | 0 〉 | ψ X ˜ 〉 + | 1 〉 | ψ X m 〉 ) | y m 〉 (1)

where X ˜ is a test point, X m are training points, m is an index for the data point X m , and Z 1 is a normalization constant to make the whole quantum state of length one. After applying the Hadamard gate to the data qubit that contains X ˜ and X m , the state becomes

| D 1 〉 = 1 Z 1 ∑ m = 0 M − 1 | m 〉 ( | 0 〉 | ψ X ˜ + X m 〉 + | 1 〉 | ψ X ˜ − X m 〉 ) | y m 〉 (2)

here ψ X ˜ ± X m = | ψ X ˜ 〉 ± | ψ X m 〉 . The key observation is

1 Z 2 ∑ y m = 0 | X ˜ + X m | 2 = 1 − 1 Z 2 ∑ y m = 0 | X ˜ − X m | 2

where Z 2 is another normalization constant, which implies that the closer the test point X ˜ to the training point X m , the larger the amplitudes in front of y m and therefore the more likely to be observed at the last step of measurements. In another word, the probability of measuring the class qubit y m in state | 0 〉 is:

P ( y ˜ = 0 ) = 1 Z 2 ∑ y m = 0 | X ˜ + X m | 2 = 1 − 1 Z 2 ∑ y m = 0 | X ˜ − X m | 2 (3)

By exploiting superposition, entanglement, and interference, this algorithm is able to evaluate the distance of the test point to all the training points at once. And more importantly, through clever mathematical manipulation, the probability of the test point belonging to class 0 or 1 can be determined at the end of the quantum circuit. However, when measuring the ancilla qubit (the qubit containing | 0 〉 and | 1 〉 in equations (1) and (2)), we may see | 1 〉 instead of | 0 〉 . Therefore the final tally of the counts for P(0) or P(1) has to exclude the results of seeing | 1 〉 in the ancilla qubit, which is way different from classical machine learning algorithms.

In this circuit in Figure 1, the first two Hadamard gates are employed to super position the qubits q[0] and q[1]. Entanglement of the qubits is seen among the controlled gates while the interference of the qubits is seen at the last Hadamard gate. According to [2], q[0] is an ancilla qubit, q[1] is index for each training data point, q[2] is data, and q[3] is for class label. q[4] is an temporary qubit that we use to build the ccu3 gate. Note that a swap operation is used at the end of this circuit to satisfy the star topology employed in IBM’s 5Q quantum chip, see Figure 1.

A typical quantum circuit is composed of initial states, followed by several unitary operators, and terminated by a few measurements according to the task of the algorithm. During execution of the algorithm, the super positioned quantum states have to maintain their coherent superposition. As a result, the length of a quantum circuit depends directly on the amount of time that these quantum states can remain coherent [14] .

The quantum classifier in [2] uses amplitude encoding technique, so the classical data points that we will use in our experiments are all of the form ( x 1 , x 2 ) = ( sin θ , cos θ ) , to make the notation easier to follow we simply use θ to represent a point. Say, we can use θ = π 4 to represent the point of ( sin θ , cos θ ) = ( 1 2 , 1 2 ) . Since the classifier is distance based, we like to investigate experimentally how the distance of the data points affect its performance. The classifier is run several times with 1000 or 8192 shots, where 8192 is the maximal number of shots allowed on Quantum Experience. We use IBM’s quantum simulator in this work rather than the real 5Q quantum computer because part of our goal is to extend the binary classifier in [2] to a multiclass classifier that uses more than five qubits. We summarize our findings in the following three subsections.

Quantum machine learning, though in its initial stage, has demonstrated its

potential to speed up some of the costly machine learning calculations when compared to the existing classical approaches. This study presents an empirical analysis of the distance based quantum classifier created in [2] . Our work consists of three components. The first is to visualize the probabilities of observing a test data point being in class 0 or 1 in a binary classification as it moves among the training points. Although it is obvious the predication depends on the distance of the data points implied by the nature of this kind of algorithms. Since it is a quantum classifier so it might be helpful if some experiments can be designed to reveal the quantum nature of this classifier. Our analysis shows that when the training points are close enough, the classifier has hard time to

differentiate the training points. The values of P(0) and P(1) are very close to 0.5 and oscillating about 0.5. Because of the stochastic nature of quantum algorithms, this phenomena makes prediction more difficult than the corresponding classical algorithms.

The second is to see if the swap operation that is commonly utilized in quantum circuit design to meet the hardware constraints may have an impact on the performance of this classifier. A typical circuit of a quantum machine learning algorithm employs superposition, entailment, and interference of qubits. Because of this, any swap of two qubits in the circuit may influence the final outcome of the algorithm. Our experiments suggest that the swap operation does not make a difference in the cases that we have tested.

Our final goal is to extend the binary classifier in [2] to a multiclass classifier to further demonstrate the capability of this classifier. The four-class classifier that we create in this work shows its ability to distinguish the four different training points with four peaks of the curves of P(0), P(1), P(2), and P(3) occurring at the right test points.

It is known that some quantum machine learning algorithms can outperform their classical counterparts, but the full scope of their potential needs to be further investigated. Our work adds new knowledge to this field by helping learn the working of this quantum classifier, fulfilling part of the big goal of understanding the value of quantum computing to machine learning and AI.

Hu, W. (2018) Empirical Analysis of a Quantum Classifier Implemented on IBM’s 5Q Quantum Computer. Journal of Quantum Information Science, 8, 1-11. https://doi.org/10.4236/jqis.2018.81001