A Quantum Convolutional Neural Network for Image Classification

Milind Upadhyay

Abstract
From self-driving cars to medical diagnoses, machine learning (ML) has revolutionized our world
in the last few decades. Additionally, Quantum Computing has considerable promise for the
future, with superposition and entanglement making quantum algorithms much more efficient
and effective than their classical counterparts. Quantum ML aims to apply these quantum
principles to the groundbreaking field of ML, such as for image classification. Convolutional
Neural Networks (CNNs) are very effective for classification, so we study a Quantum
Convolutional Neural Network (QCNN) that is trained to distinguish between handwritten
numbers in the MNIST dataset and compared to a similar-sized classical network. After tuning
the QCNN and quantum encoding, the QCNN achieved comparable accuracy to the classical
network.

Introduction
Quantum ML can be applied to many classification problems just like classical ML, using
quantum neural networks. Quantum networks involve layers of rotation gates by parameterized
angles. The optimal angle parameters for the given problem are solved for by a process such as
gradient descent, similar to how weights and biases are solved for in classical networks. Farhi et
al. [4] apply a simple, pure quantum network to classifying handwritten digits. Quantum
networks can also be done as "hybrids", with some classical component of ML or feature
detection, as done for medical image classification by Mathur et al. [7]

Classical CNN’s are widely used for image classification. QCNN’s are special quantum networks
that are somewhat similar to classical ones in the sense that they have layers of convolution and
pooling. Tensorflow Quantum [2] has applied a QCNN to classify if a cluster state of qubits [8] -
a state with a highly-entangled group of qubits - is excited or not. This QCNN is a simpler
version of the one presented by Cong et al. [3]

Classical inputs can be encoded into qubits in a multitude of ways, as described by LaRose et
al. [5] Angle encoding involves using classical inputs as parameters for rotation gates on qubits,
and simple versions of it use one qubit per classical input value. Angle Encoding is very simple,
requiring only one gate to encode a value.

We apply Tensorflow Quantum’s [2] QCNN, to distinguish handwritten digits in the same dataset
as Farhi et al. [4] We entangle all the input qubits in a cluster state before plugging them into the
QCNN.

1
Method
The input images are first downsized and stored in qubits with angle encoding. These qubits are
then plugged into a QCNN which outputs the image class.

Dataset
We use the MNIST dataset [6], which contains grayscale images of handwritten digits. Due to
the limits on current quantum hardware, we heavily downsample this dataset. We filter the
dataset into a binary classification between 3's and 6's, which results in 12,049 train images and
1968 test images.

We also downsample the images from 28 by 28 to 4 by 4 to avoid using too many qubits. This is
around the minimum size to still be able to distinguish between the numbers.

Due to heavy downsampling, some test images become identical to train ones. We removed
these 317 test images, resulting in 1597 testing images in the final dataset.

Encoding
We then encode the downsized images into the clusterstate qubits with Angle Encoding. For
each pixel’s grayscale value, we rotate a corresponding qubit around the X axis, using the pixel
value as an angle. In other words, for a grayscale pixel at row 𝑖 and column 𝑗, 𝑝𝑖𝑗, the
corresponding qubit has a 𝑅𝑋(𝑝𝑖𝑗) gate performed on it.

The simple quantum network [4] applied to the MNIST dataset, binarized the pixel values, which
means converting each grayscale pixel value (0 to 255) to either 0 or 1 based on if the value is
greater than a certain threshold. We instead use the actual grayscale value (0 to 255) as a
rotation, to make this extendable to complex images that would lose critical information when
binarized.

QCNN
Just like a classical CNN, a QCNN has layers of applying convolution on groups of inputs
(qubits) and then pooling them together. We follow Tensorflow Quantum’s [2] QCNN structure.

The convolution step involves performing parameterized two-qubit gates on consecutive qubits.
The two qubit gates involve a set of parameterized gates done separately on each qubit, then a
set of parameterized gates done on the two qubits together, and finally another set of separate
single-qubit gates rotations. The parameterized single-qubit gates are 𝑋, 𝑌, 𝑍 raised to the

2
power of the parameters, and the two-qubit gates are 𝑍𝑍, 𝑌𝑌, 𝑋𝑋 raised to powers which are the
parameters. The parameters will ideally become those that cause the rotations to check for
important features in the qubits, which can then get pooled together.

The pooling operation takes in two qubits and reduces them into one, transferring a "source"
qubit into a "sink" one. First, it involves another set of one-qubit parameterized gates on both
qubits, the same ones as the parameterized X, Y, Z gates in the convolution step. It then
performs a CX gate from the source to the sink one, and performs the inverse of the same
parameterized one-qubit gates on the sink qubit. This is different from the classical
pooling which would normally find the maximum output of the convolution kernel, but both
reduce the size of inputs by combining results.

The overall QCNN circuit involves layers of convolution and pooling. Since the number of qubits
is halved after each layer due to pooling, the number of qubits has to be a power of 2. The
number is indeed a power of 2 because the 4 by 4 images result in 16 qubits.
For each layer, the convolution is run on the qubits, and then the first half of qubits are pooled
into the second half. After this, the first half of qubits are thrown away, and the process is
repeated with the second half until only 1 qubit is left. This is then read out with the 𝑍 operator to
determine whether the input was a 3 or a 6.

Classical Network
The classical network is a simple non-convolutional network. After flattening the input into 16
pixel values, it has a dense layer of 5 neurons and then an output layer of 1 neuron. This results
in 91 parameters, which is slightly over the quantum one which has 84 parameters. It is
therefore a decent comparison to evaluate the quantum network. We did not use a classical
convolutional network because it requires more parameters and higher-dimensional inputs.

3
Results
Using Tensorflow [1], we trained the quantum and classical networks on the MNIST dataset and
compared their performances.

QCNN
We achieved the best performance for the QCNN with 5 epochs and a batch size of 50. Even
though this only uses a fraction of all training images, this was the best because the network
started overfitting with more epochs or a higher batch size. The QCNN achieved an accuracy of
around 87%. When we binarized the images (mentioned in the Encoding section) it went up to
90%, however we did not pursue that strategy further for reasons mentioned in that section. The
accuracy for our QCNN is slightly above the accuracy of the simple quantum network [4] which
performed the same task.

Comparison to Classical Network

Conclusion
Ultimately, given the comparable performance of this QCNN on the MNIST dataset to a
similar-sized classical network, a QCNN combined with angle encoding is a viable strategy for
image classification. In the future, Quantum Computers will have many more qubits and
encoding pixel values for large images may be viable. However, for current Quantum
Computers, other techniques such as encoding features or keypoints may need to be used.

References
1. Abadi, M. et al. TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems.
Software available from [Link] (2015).
[Link]

2. Broughton, M. et al. TensorFlow Quantum: A Software Framework for Quantum Machine

3. Cong, I., Choi, S. & Lukin, M. Quantum convolutional neural networks. Nature Physics (2019).
[Link]

4
4. Farhi, E. & Neven, H. Classification with Quantum Neural Networks on Near Term Processors
(2018).
[Link]

5. LaRose, R. & Coyle, B. Robust data encodings for quantum classifiers (2020).
[Link]

6. LeCun, Y. & Cortes, C. MNIST handwritten digit database (2010).

7. Mathur, N. et al. Medical image classification via quantum neural networks (2021).
[Link]

8. Nielsen, M. CLUSTER-STATE QUANTUM COMPUTATION (2005).