Automatic design of quantum feature maps

Sergio Altares-López 1 , Angela Ribeiro 1 , Juan José

Garcı́a-Ripoll 2
arXiv:2105.12626v1 [quant-ph] 26 May 2021

1
Centre for Automation and Robotics - CAR (CSIC-UPM), Ctra. Campo Real km.
0,200, 28500 Arganda, Spain.
2
Instituto de Fı́sica Fundamental IFF-CSIC, Calle Serrano 113b, 28006 Madrid,
Spain.
E-mail: [email protected]

Abstract. We propose a new technique for the automatic generation of optimal ad-
hoc ansätze for classification by using quantum support vector machine (QSVM). This
efficient method is based on NSGA-II multiobjective genetic algorithms which allow
both maximize the accuracy and minimize the ansatz size. It is demonstrated the
validity of the technique by a practical example with a non-linear dataset, interpreting
the resulting circuit and its outputs. We also show other application fields of the
technique that reinforce the validity of the method, and a comparison with classical
classifiers in order to understand the advantages of using quantum machine learning.

quantum machine learning, genetic algorithms, artificial intelligence, quantum

computing, optimization, automatic quantum classifier generation

1. Introduction

Quantum machine learning is an emerging field of research that bridges the progress
in quantum computing hardware and algorithms with ideas and problems coming from
artificial intelligence. The field is suffering steady and fast progress but already features a
large corpus of algorithms and applications [1, 2]. On the one hand, we find applications,
such as clustering [3], quantum anomaly detection [4], dimensionality reduction [5, 6, 7]
or support vector machines [8], which build on the HHL algorithm for matrix inversions
or related matrix-vector operations. On the other hand, we place algorithms that are
ready for near-term intermediate-scale quantum (NISQ) devices, and which are typically
based on parameterized quantum circuits [9] that implement autoencoders [10], support
vector machines and quantum classifiers [11, 12], or generative adversarial networks
[13, 14], among other applications.
Due to their simplicity and immediate experimental access, this second framework
has gained considerable interest. However, a key problem of parameterized circuit is
their design, both from the point of view of the structure of the circuit as well as
their parameterization. The structure and design condition the expressive power of the
Automatic design of quantum feature maps 2

circuit [15, 16], and its capacity to explore the Hilbert space and encode probability
distributions more efficiently than other generative models. However, too expressive
circuits can be subject to local minima and barren plateaus [17] that prevent reaching
the optimal parameterizations. Partial solutions to these challenges include adaptive
initialization strategies [18], pruning of circuits [19], simultaneous optimization of
parameters and rotation generators [20], or the implementation of global optimization
strategies such as genetic algorithms [21, 22, 23] that optimize gates or structure.
In this work, we focus on the problem of supervised learning using quantum feature
maps [11, 12] that are optimized with a genetic algorithm. As compared to earlier
variational works [23], we provide a comprehensive solution that automatically designs
both the structure and the parameterization of the feature map circuit. Our algorithm
uses an NSGA-II genetic algorithm with a multiobjective Pareto front that optimizes
the accuracy and generalization power of the map, while minimizing the circuit size. We
test this method against both synthetic and realistic benchmarks, finding remarkable
accuracy for all problems. Moreover, the Pareto strategy and our weighting of gates
seems to produce quantum feature maps that are largely uncorrelated. This hints at
the possibility of constructing hybrid quantum-inspired strategies for machine learning
based on these ideas.
The structure of this work is as follows. In section 2 we review the method of
quantum feature maps and quantum kernels for supervised learning with support vector
machines. We introduce a new function (6) that seems to exhibit good separation
properties and will be used in simulations. With this knowledge, section 3 introduces
the algorithm for genetically designed quantum kernels. After a brief review of genetic
algorithms in section 3.1, we introduce an encoding of quantum feature maps as binary
strings of genes. The genetic map from section 3.2.1 is a small example with only 5
bits, but exemplifies how to encode structure, types of gates, dependence on the input
parameters and numerical parameterization of the circuit. This contrasts with other
works [23] where structure and parameters were optimized separately, with different
methods. In section 3.2.2 we describe the fitness function, which is designed for a
multiobjective optimization of the accuracy, the capacity for generalization and the
simplicity of the quantum feature map. We also describe how a Pareto search and elitist
genetic operations can be designed to help in this optimization. Section 4 presents the
results of applying our algorithm to synthetic and realistic benchmarks, in sections 4.1
and 4.3, respectively. In these examples we also see how the optimization converges to
low-entanglement feature maps, while still having good accuracy and generalization.
Based on this, in section 4.2 we discuss how such circuits could be amenable for
interpretation. Finally, section 5 summarizes the main conclusions and possible avenues
for future exploration.
Automatic design of quantum feature maps 3

2. Quantum Kernel Method

In this work we will focus on the supervised training of binary classifiers. Given a
training dataset {(xi , yi )}Li=1 with normalized feature vectors xi ∈ Rd and binary classes
yi ∈ {+1, −1}, we can design a model f (x) that predicts the class of any other point,
either in this set or in unseen data. The support vector machine (SVM) is one of
the earliest binary classification techniques. Developed for linearly separable data, this
method constructs a hyperplane with normal w and displacement b such that the two
classes y = +1 and y = −1 lay on opposite sides of the hyperplane. The classifier has a
simple form, given by a sign function
f (x) = sign wT · x + b .


(1)
The hyperplane is constructed using support vectors from the training set
X
w= βi yi xi , (2)
i
in a way that maximizes the margin between those vectors and the hyperplane.
There are various techniques that turn SVM into a universal classifier, working data
that is not linearly separable. One is to construct additional features or variables out the
original vectors, enlarging the dimension of the classification space x̃i := Φ(xi ) ∈ Rr ,
with r  d. By raising the dimensionality, the so called feature map can transform the
problem into a linearly separable one. Interestingly, the classifier can be inferred from
a kernel function that encodes the scalar product between the new features
K(x, x0 ) = Φ(x)T Φ(x0 ). (3)
This can be seen from the expression of the hyperplane w in terms of the new features,
and how this all fits into the final classifier
!T
X X
f (x) = βi yi Φ(xi ) Φ(x) + b = βi yi Φ(xi )T Φ(x) + b (4)
i i
X
= βi yi K(xi , x) + b.
i

Moreover, by Mercer’s theorem [24], we do not need to know the form of the feature
map—which may even be an infinite-dimensional function—, but just a kernel function
K(x, x0 ) that has the right positivity properties to encode a scalar product.
When developing quantum classifiers for classical data, the usual approach is to
engineer a feature map from classical to quantum features |Φ(x)i [11, 12]. This map can
be trivial—e.g. encoding data into quantum register states or quantum amplitudes—,
but more generally it is a parameterized unitary transformation, built from quantum
gates that are depend on the input features |Φ(x)i := U(x; θ) |0in , and on some
additional controls θ. The feature map can be combined with further classification
circuits or measurements, to create the so called quantum neural networks. However, as
argued in Refs. [11, 12, 25], we could simply use those circuits to evaluate a kernel
2 2
K(x, x0 ) = |hΦ(x)|Φ(x0 )i| = h0n |U(x; θ)† U(x0 ; θ)|0n i , (5)
Automatic design of quantum feature maps 4

and derive the corresponding SVM classifier f (x; θ).

In this work we use a different type of quantum kernel
K(x, x0 ) = Re hΦ(x)|Φ(x0 )i = Re h0n |U(x; θ)† U(x0 ; θ)|0n i . (6)
By not squaring the scalar product between vectors, the function resembles more the
original motivation for K(x, x0 ). Moreover, as we have confirmed numerically, this
kernel allows for sharper separations and more easy convergence of the optimizer.
However, while this choice is neutral from a classical simulation point of view, it is
more complicated to evaluate in a quantum computer. Unlike Ref. [12], to estimate
K(x, x0 ) we would need to use an ancillary qubit, prepared in a quantum superposition,
and controlling the U operation, to estimate the scalar product as the result of an
interference process.
This quantum kernel method is usually combined with some kind of iterative update
of the parameters θ, to maximize a cost function that includes the accuracy of the model
and some other regularizations. Depending on the expressive power of the underlying
feature map, this approach can lead to barren plateaus and other obstacles that prevent
a good training. For that reason, in this work we explore a global optimization method
that trains both the parameters θ as well as the structure of the quantum circuit U, by
using evolutionary artificial intelligence techniques, also known as genetic algorithms.

3. Genetically Designed Quantum Kernels

3.1. Overview of Genetic Algorithms

Genetic algorithms are meta-heuristic optimization techniques based on the theory of
evolution. The algorithms perform a guided search in a space of solutions, evolving a
population of individuals with encoded the feature maps, through the application of
genetic operations. In each algorithm iteration or generation, the resulting offspring
is selected in order to improve one or more objetives. As a result of evolutionary
pressure, the collective is more likely to select the best suited individuals from a very
large configuration space, in a efficient way.
A very important ingredient in the genetic algorithm is the fitness function. This
function depends on some metric that we wish to maximize (or minimize), as well as
other regularizations. However, as we show in this work, it is possible for a genetic
algorithm to achieve more than one goal, performing a multiobjective optimization
process. In this case, the individuals in each generation that maximize (or minimize)
the fitness function are called the Pareto front. More precisely, a solution x is called
dominated by another solution y only if y is equal or better than x solution with
respect to all the objectives of the function. The Pareto front is the set of non-
dominated solutions which satisfy all objectives defined in the fitness function, and
which is progressively improved by the evolution algorithm.
A key ingredient for the success and utility of a genetic algorithm is the choice of
genetic operations that evolve the population of individuals, as seen in figure 1d. The
Automatic design of quantum feature maps 5

Figure 1. Technique scheme. a). The initial population is initialized. b). We

can see the decoding process based on five bits per quantum gate and angle. c). This
individual is used in the fitness function as feature map for QSVM in order to calculate
accuracy and number of gates. d). The genetic algorithm process continue with phases
already seen until early stopping condition so that the most optimized ansatz for this
dataset is provided.

selection operator chooses a subset of the existing population to create a new generation
using the crossover and mutation operators. The mutation operator randomly alters
the information of selected individuals to explore places far from the solution space.
The crossover is an stochastic operation that allows even more drastic explorations, by
allowing two individuals to exchange their genetic information. Note that, while the
probability of selection is proportional to the fitness of the individuals, the mutation
and crossover probabilities are fixed values that have been tuned for performance.
Finally, the genetic algorithm typically involves some stopping early conditions,
statements which determine whether the evolution process has achieved its goal. Some
possible strategies are checking the convergence or saturation of the fitness objetive, a
minimum accuracy threshold that keeps the process going for further steps or defining
a maximum number of generations.

3.2. Genetic quantum feature map

We will now describe a multiobjective genetic algorithm that automatically designs
and optimizes quantum classifiers based on quantum feature maps and support vector
machines. The algorithm explores a configuration space of parameterized quantum
Automatic design of quantum feature maps 6

circuits that potentially represent feature maps. It looks for those circuits that, once
trained in a QSVM method, maximize the accuracy with which they generalize to the
validation data set, while minimizing the complexity of the circuit, which is measured
in terms of circuit depth and difficulty of operations.
The complete algorithm is summarized in figure 1. The process starts with an
initial population of individuals, represented by bit strings of size M × N × 5. The
evaluation function decodes each individual, creating an associated quantum circuit
(see section 3.2.1). This circuit, together with the training dataset, is used to implement
a quantum kernel SVM algorithm, computing the fitness function (see section 3.2.2).
The best individuals have more probability to be elected and subjected to the different
genetic operations, such as mutation, crossover and selection, creating a new generation
of individuals or quantum circuits. The whole process is repeated until we meet the
convergence criteria.

3.2.1. Encoding The first step for engineering our algorithm is to design a map from
the genetic information to the quantum circuit that we wish to characterize. In our
model, the genes will be binary strings that encode local, entangling and parameterized
quantum gates. To create a minimalistic encoding that exemplifies all types of gates,
we use five bits per gene s0 s1 s2 s3 s4 , as shown in figure 1b. We aim to create a quantum
circuit acting on M qubits with a maximum of N layers and use M × N × 5 bits in
total. For simplicity, the order of actuation of the genes is sequential. Thus the i-th
gene operates on the j-th qubit of the quantum register and possibly depends on the
k-th a variable from the input data x ∈ Rd , with j = i mod M and k = i mod d.
The mapping from bits to gates is also very straightforward. The first three
bits s0 s1 s2 determine whether the gate is fixed—a Hadamard or a CNOT gate—, or
whether it is a local rotation parameterized by a value in the input data, Rα (θi xk ) =
exp(−iθi xk σkα ). When the gate is parameterized, the first bits s0 s1 s2 select the rotation
s s
axis, the last two bits select a proportionality parameter θi = π4 2−2 3 −4 4 . When the gate
is a CNOT, it acts on consequitive qubits, j and j + 1 mod M. All other combinations
of bits not reflected in figure 1b are taken to be just identity operations.
This gates selection has been chosen in such a way that all the elements found
in variational circuits used as feature maps are represented, being local parameterized,
non-parameterized and entanglement gates.

3.2.2. Fitness Function Our fitness function is designed to maximize the accuracy and
minimize the complexity of the variational circuit. To measure the latter, we introduce
a size metric, labeled SM, which assigns different costs to the number of local gates
Nlocal and the number of entangling gates NCNOT , weighted as follows
Nlocal + 2NCNOT
SizeMetric(SM) = . (7)
Nqubits
The second ingredient in the fitness function is the accuracy of the encoded circuit.
To compute this metric, we divide the data into a training set and a test set. We use
Automatic design of quantum feature maps 7

the quantum circuit and the training set to compute the classifier f (x) in the quantum
kernel SVM. We then estimate the accuracy of the model f (x) over the test set, as the
fraction of points that are properly classified.
We aim to maximize both quantities in a multiobjective optimization process,
creating a Pareto front that carefully balances the relative importance of both figures
of merit. A high weight in accuracy can produce a collapse into one individual loosing
the necessary genetic diversity to be able to minimize the quantum circuit size along
the evolution. On the other hand, a very small number of gates can hinder the power
of the quantum kernel to separate the features. In order to achieve a proper balance,
we engineer a fitness function that increases the relevance of the SM as the accuracy
approaches its limiting value 1, using the following multiobjective fitness function
WeightsControl(WC ) = SM + SM ∗ accuracy2 . (8)

3.2.3. Genetic Operators In the genetic algorithm we use selection, mutation and
crossover operators. The selection operator is a multiobjetive Non-Sorted Genetic
Algorithm II (NSGA-II). This algorithm decides which individuals survive to the next
generation based on Pareto-dominance and density-based metrics [26]. This algorithm
has a strong tendency to keep individuals with higher fitness, because it selects the
individuals after ordering the population by dominance. However, it also uses the non-
dominated individuals’ rank and density-distance to add diversity in each generation.
Since our feature maps are coded in binary format, we can use a bit flip mutation
operator (c.f. figure 2b). The value pmut indicates the probability of an individual to
be mutated. Once an individual is selected for mutation, each gene can be flipped
with a probability pind . As for the crossover operator, we implement a binary swap of
contiguous bit substrings (see figure 2c). The value pcross determines probability that
a crossover takes place, while the beginning and end of the swapped bits are randomly
chosen along the complete strings.
We increase the elitism of the algorithm with a Mu Plus Lambda (µ + λ) algorithm.
This strategy modifies how we create the next generation of individuals, establishing
a competition between the current population (size µ) and the offspring (λ) that is
obtained by genetic operators, as sketched in figure 2a. This competition ensures genetic
diversity while it also preserves the best individuals that have been obtained through
the evolutionary algorithm [27].
All hyperparameters previously mentioned —crossover and mutation probabilities
and population sizes—were optimized and tested, to achieve a good compromise
between convergence speed and optimal classification. After several tests, the optimal
hyperparameters were found to be 30% probability of crossover and 70% mutation,
with a 20% probability of bits being mutated. This is an interesting balance that allows
exploring drastic changes in the population through crossover, while maintaining a high
rate of small changes through mutation. While this may seem very aggressive, the
random component is kept in check by the high elitism of the competition between
children and parents in the Mu Plus Lambda strategy.
Automatic design of quantum feature maps 8

Figure 2. Genetic operators. (a) Genetic algorithm µ + λ strategy. The initial

population µ, after applying the genetic operators, produces an offspring of λ size. This
offspring competes with the parents and from this competition a number of individuals
equal to the size of the individual population µ is selected. This process is repeated
throughout the generations of evolution. (b) Mutation. A bit flip mutation affects a
fraction of the genes , giving rise to a different individual. (c) Crossover. Randomly
selected genes (marked in green) are exchanged between two individuals, keeping the
rest of the genes chain constant.

4. Results and Discussion

4.1. Toy model

In order to proof this new method, we use the Moons synthetic non-linear dataset with
two classes, shown in figure 3a, and generated using Scikit [28]. The 150 datapoints are
scaled between [-1,+1] as a preprocessing step, and randomly split into a training (70%)
and test (30%) sets. These sets are used to train quantum circuits with up to 6 qubits
and 6 layers, which are progressively optimized by the genetic algorithm. As illustrated
in figure 3b, we optimize the circuits over 5000 generations, using a population of 100
individuals.
The initial circuits make use of all available qubits and all layers, as shown in
figure 3c. Already in these circuits we observe the penalty associated to CNOT gates
decreasing the number of entangling unitaries, as compared to other ansätze in the
literature. More interestingly, the Pareto front combined with the elitist strategies is
capable of further realizing that no entanglement at all is required to fit this model.
Thus, after 5000 generations, the algorithm produces the simple uncorrelated circuit
Automatic design of quantum feature maps 9
a) 1.00 b) Variables Value
0
0.75 1
Generations 5000
0.50
Population (µ) 100
0.25
0.00 Offspring (λ) 15
x2

0.25 Qubits 6
0.50 Max. Layers 6
0.75 Crossover Prob. 0.3
1.00
1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00 Mutation Prob. 0.7
x1 Mutation Prob. Ind. 0.2
c) d)
q0

q1

q2

Figure 3. (a) Dataset composed by 150 points with a non-linear pattern and a binary
target. (b) Hyperparameters used to optimize the QSVM circuit with our genetic
algorithm. (c) Structure of circuits created in the first generations of the genetic
algorithm. (d) Final circuit with 1.0 accuracy.

1.00 0
a) 0.75 1 b)
0.50
0.25
0.00
x2

0.25
0.50
0.75
1.00
1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00
x1

Figure 4. (a) Validation dataset, together with the predictions and decision boundary
from the generated model. (b) Confusion matrix produced by the application of the
QSVM model onto the validation dataset.

from figure 3d, which fits the test set with perfect accuracy.
The fact that the generated model has perfect accuracy is useless, if it cannot
generalize to other data from the same distribution. We validate the utility of the
model using additional datapoints, a validation set, with 500 points generated by the
same synthetic algorithm. The same scaling preprocessing step [-1,+1] that is applied
to the training data is also applied to these validation datapoints. Figure 4a shows both
Automatic design of quantum feature maps 10
1.00
a)
0.75 Accuracy: 1.0
0
1
1.00
b) 0
1
0.75 Accuracy: 0.57
0.50 0.50
0.25 0.25
0.00 0.00
x2

x2
0.25 0.25
0.50 0.50
0.75 0.75
1.00 1.00
1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00 1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00
x1 x1
1.00
c)
0.75 Accuracy: 0.68
0
1
1.00
d) 0
1
0.75 Accuracy: 0.53
0.50 0.50
0.25 0.25
0.00 0.00
x2

0.25 x2 0.25
0.50 0.50
0.75 0.75
1.00 1.00
1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00 1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00
x1 x1

Figure 5. (a) Data points and prediction boundaries from the full quantum kernel
SVM. (b), (c) and (d) are the decision boundaries provided by the circuits on the first,
second and third qubit, respectively.

the validation dataset used and the predictions made by the quantum support vectorial,
defined by the decision boundary. Figure 4b also illustrates the confusion matrix of
this validation process, considering both real and predicted labels, and identifying the
incorrectly classified data. The confusion matrix allows us to conclude that the QSVM
extrapolates to unseen data with same distribution, because 473 out of 500 data in the
dataset have been correctly classified. In other words, a 94.6% of correct classified data
or 0.946 accuracy.

4.2. Interpretability
As we see above and will see in later examples, the strong penalty on entangling gates
makes the genetic algorithm prefer circuits that have smaller clusters of uncorrelated
qubits. Ideally, only the gates that are essential for the modelization are included.
The result is a circuit that can be decomposed as a tensor product of separate
unitaries, and a quantum kernel that is a scalar product of separate kernels, as in
K(x, x0 ) = m 0
Q
i=1 Ki (x, x ). We suggest to study the classification induced by each kernel
Ki separately and by their combination, as a strategy to provide interpretations of the
rules that the evolutionary strategy has produced.
An example of this study is performed in figure 5 for our synthetic model. Figure
5a illustrates the boundaries of the complete kernel, which has accuracy 1.0, while later
figures 5b-d show the boundaries induced by each separate kernel. As we can see, both
Automatic design of quantum feature maps 11

qubits one and two, provide linear hyperplanes, being qubit three the one that provides
a degree of non-linearity achieved by applying three rotations. Finally, the combination
of these qubits forms the desired non-linear pattern. Interestingly, the single-qubit
boundaries have a lower classification accuracy, of 0.57, 0.68 and 0.53, respectively, but
their nonlinear combination in the final kernel gives the right predictions.

4.3. Other Use cases

Table 1. Results from applying the genetic engineering of QSVM to other model
problems in supervised machine learning. For the most difficult model we also provide
the accuracy of other classical methods for supervised machine learning: a k-NN, a
linear support vector machine, and an SVM with a polynomial kernel with degree 2
(poly).
Parkinson [29] IoT irrigation [30] Drug classification [31]
Circuit Fig. 6a Fig. 6c Fig. 6b
Accuracy 1.0 1.0 1.0
Generations 5000 1000 500
# attributes 22 2 5
# classes 2 2 5
Max qubits 15 5 5
Max depth 8 5 5
Mutation probability (pmut ) 0.7 0.7 0.7
Mutation ind. prob. (pind ) 0.2 0.2 0.2
Crossover prob. (pcross ) 0.3 0.3 0.3
k-NN accuracy 0.82 1.0 0.70
SVM (linear) accuracy 0.89 1.0 0.87
SVM (poly - 2) accuracy 0.89 1.0 0.65

We have applied the method to other problems that are standard benchmarks
for classical supervised learning techniques. Table 1 lists three problems, with the
characteristics of the datasets, the hyperparameters of the genetic algorithm and the
resulting accuracy. As seen in our experiments, the technique performs well for datasets
with a high number of features as well as for datasets with more than two classes. The
comparison with non-quantum classification methods also shows and advantage of the
QSVM technique, specially in the highly complex multiclass classification of drugs.
Figure 6 illustrates the structure and parameterization of the quantum feature
maps that optimally classify these benchmarks. Interestingly, two of the circuits
are uncorrelated and have no CNOT gates, while the third one, for multiclass drug
classification, has just one entangler gate. This is relevant for several reasons. First,
it illustrates the power of individual qubits a quantum classifiers, a realization already
introduced in Ref. [32]. Second, the structures we have obtained, having little or no
Automatic design of quantum feature maps 12
a)

b)

c)
q0

q1

Figure 6. Circuits generated for the supervised learning problems in Table 1. (a)
Parkinson problem [29], (b) drug classification [31] and (c) IoT irrigation [30].

correlation, admit an efficient classical simulation which constitutes in itself a type of

quantum-inspired machine-learning technique.

5. Summary and Outlook

In this work we have explored the global optimization of quantum feature maps in
a quantum kernel SVM algorithm using evolutionary multiobjective algorithms. The
feature map is built as a parameterized quantum circuit that depends on the input
data. The genetic algorithm stored the structure of the circuit, the actual gates and the
functional dependence on the data as a string of binary-encoded genes. The algorithm
evolves a population of individuals with genetic operators that seek to maximize the
accuracy of these feature maps in modeling the data, while minimizing the complexity
of the circuit. This is implemented using a nonlinear fitness function that combines both
Automatic design of quantum feature maps 13

goals, and simultaneously applying a Pareto front selection strategy for the individuals.
We have applied this algorithm both to synthetic and to realistic benchmarks in
the field of supervised machine learning, both single- and multiclass classification. The
algorithm produces 100% accurate classifiers that can still generalize to unseen data
since this metric is obtained from test sets. Moreover, the classifiers have a simple
structure, with minimal or no correlations, which still capture the underlying nonlinear
patterns. We attribute the simplicity of these circuits to the classification power of
single qubits and single-qubit operations [32], which is enhanced by the combination of
multiple parallel circuits. We believe that the resulting circuits are amenable to further
interpretation strategies, in a simpler way than neural networks or other ML ansätze.
Moreover, our results suggest the power of product states as another quantum-inspired
variational strategy for supervised learning.
Our work leaves many avenues for exploration. The gene encoding that we have
implemented contains a minimalistic set of entangling, local and parameterized gates,
with sufficient precision for the problems we have explored. This can be extended in
various ways, such as enlarging the set of weights in the parameterization, changing the
order in which parameters appear in the circuit, including also more local and entangling
gates, including free parameters θi that can be optimized using SPSA or other strategies,
etc. If we focus on entanglement-free ansätze, we also find a rich avenue to explore the
implementation of these models as standalone tools for machine learning, or developing
a more clear strategy for the interpretation of the resulting classifiers—e.g. developing
a kind of rule-based explanation of the model.

Acknowledgments

The authors gratefully acknowledges the computer resources at Artemisa, funded by

the European Union ERDF and Comunitat Valenciana as well as the technical support
provided by the Instituto de Fı́sica Corpuscular, IFIC (CSIC-UV).
This work has been supported by Spanish project PGC2018-094792-B-
100 (MCIU/AEI/FEDER, EU), CAM/FEDER Project No. S2018/TCS-4342
(QUITEMAD-CM), and CSIC Platform PTI-001.

6. References

[1] Biamonte J, Wittek P, Pancotti N, Rebentrost P, Wiebe N and Lloyd S 2017 Nature 549 195–202
URL https://doi.org/10.1038/nature23474
[2] Schuld M and Petruccione F 2018 Supervised Learning with Quantum Computers (Springer
International Publishing) URL https://doi.org/10.1007/978-3-319-96424-9
[3] Kerenidis I, Landman J, Luongo A and Prakash A 2018 q-means: A quantum algorithm for
unsupervised machine learning (Preprint 1812.03584)
[4] Liu N and Rebentrost P 2018 Physical Review A 97 URL https://doi.org/10.1103/physreva.
97.042315
[5] Lloyd S, Mohseni M and Rebentrost P 2014 Nature Physics 10 631–633 URL https://doi.org/
10.1038/nphys3029
Automatic design of quantum feature maps 14

[6] Cong I and Duan L 2016 New Journal of Physics 18 073011 URL https://doi.org/10.1088/
1367-2630/18/7/073011
[7] Duan B, Yuan J, Xu J and Li D 2019 Physical Review A 99 URL https://doi.org/10.1103/
physreva.99.032311
[8] Rebentrost P, Mohseni M and Lloyd S 2014 Physical Review Letters 113 URL https://doi.org/
10.1103/physrevlett.113.130503
[9] Benedetti M, Lloyd E, Sack S and Fiorentini M 2019 Quantum Science and Technology 4 043001
URL https://doi.org/10.1088/2058-9565/ab4eb5
[10] Romero J, Olson J P and Aspuru-Guzik A 2017 Quantum Science and Technology 2 045001 URL
https://doi.org/10.1088/2058-9565/aa8072
[11] Schuld M and Killoran N 2019 Physical Review Letters 122 URL https://doi.org/10.1103/
physrevlett.122.040504
[12] Havlı́ček V, Córcoles A D, Temme K, Harrow A W, Kandala A, Chow J M and Gambetta J M
2019 Nature 567 209–212 URL https://doi.org/10.1038/s41586-019-0980-2
[13] Dallaire-Demers P L and Killoran N 2018 Physical Review A 98 URL https://doi.org/10.1103/
physreva.98.012324
[14] Lloyd S and Weedbrook C 2018 Physical Review Letters 121 URL https://doi.org/10.1103/
physrevlett.121.040502
[15] Du Y, Hsieh M H, Liu T and Tao D 2020 Physical Review Research 2 URL https://doi.org/
10.1103/physrevresearch.2.033125
[16] Sim S, Johnson P D and Aspuru-Guzik A 2019 Advanced Quantum Technologies 2 1900070 URL
https://doi.org/10.1002/qute.201900070
[17] McClean J R, Boixo S, Smelyanskiy V N, Babbush R and Neven H 2018 Nature Communications
9 URL https://doi.org/10.1038/s41467-018-07090-4
[18] Grant E, Wossnig L, Ostaszewski M and Benedetti M 2019 Quantum 3 214 ISSN 2521-327X URL
https://doi.org/10.22331/q-2019-12-09-214
[19] Sim S, Romero J, Gonthier J F and Kunitsa A A 2021 Quantum Science and Technology 6 025019
URL https://doi.org/10.1088/2058-9565/abe107
[20] Ostaszewski M, Grant E and Benedetti M 2021 Quantum 5 391 URL https://doi.org/10.
22331/q-2021-01-28-391
[21] Li R, Alvarez-Rodriguez U, Lamata L and Solano E 2017 Quantum Measurements and Quantum
Metrology 4 1–7 ISSN 2299-114X URL http://dx.doi.org/10.1515/qmetro-2017-0001
[22] Lamata L, Alvarez-Rodriguez U, Martı́n-Guerrero J D, Sanz M and Solano E 2018 Quantum
Science and Technology 4 014007 ISSN 2058-9565 URL http://dx.doi.org/10.1088/
2058-9565/aae22b
[23] Chivilikhin D, Samarin A, Ulyantsev V, Iorsh I, Oganov A R and Kyriienko O 2020 Mog-vqe:
Multiobjective genetic variational quantum eigensolver (Preprint 2007.04424)
[24] Géron A 2019 Hands-on machine learning with Scikit-Learn, Keras, and TensorFlow: Concepts,
tools, and techniques to build intelligent systems (O’Reilly Media)
[25] Schuld M 2021 Quantum machine learning models are kernel methods (Preprint 2101.11020)
[26] Barán B, Carballude A and Villagra M 2021 SN Computer Science 2 URL https://doi.org/10.
1007/s42979-020-00398-3
[27] Gutiérrez Reina D, Tapia Córdoba A and Rodrı́guez Del Nozal A 2020 Algoritmos Genéticos con
Python (ALFAOMEGA MARCOMBO)
[28] Thirion B, Varoquaux G, Gramfort A, Michel V, Grisel O, Louppe G and Nothman J scikit-
datasets (generate samples of synthetic data sets). URL https://github.com/scikit-learn/
scikit-learn
[29] Little M, McSharry P, Roberts S, Costello D and Moroz I 2007 BioMedical Engineering OnLine
2007 6–23
[30] Intelligent irrigation system (by using temperature and moisture data) URL https://www.
kaggle.com/harshilpatel355/autoirrigationdata
Automatic design of quantum feature maps 15

[31] Drug classification dataset URL https://www.kaggle.com/prathamtripathi/

drug-classification
[32] Pérez-Salinas A, López-Núñez D, Garcı́a-Sáez A, Forn-Dı́az P and Latorre J I 2021 One qubit as
a universal approximant (Preprint 2102.04032)