A Quantum Convolutional Neural Network on NISQ Devices

I. INTRODUCTION

• CNNs exploit data-correlation information and have applications including image recognition and precision medicine.Source: §I · p. 1
• Quantum information processing uses superposition and entanglement and can offer higher computational speed for certain problems; quantum machine learning has often focused on the evaluation component.The passage cites quantum information processing as 1,2, quantum speedups as 3–5, developments in quantum machine learning as 6–15, and evaluation-focused quantum algorithms as 16–21.Source: §I · p. 1
• CNNs combine convolution, pooling, and fully connected layers; their filter-mask and fully connected weights are optimized through training on large datasets.Convolution uses neighborhood pixels, pooling reduces feature-map size, and the fully connected layer computes output from the remaining pixels.Source: §I · p. 1
• The QCNN implements convolution through linear combinations of unitary operators, pooling by abandoning qubits, and a fully connected layer through Hamiltonian expectation measurements and nonlinear post-processing.The authors preview numerical demonstrations with two examples and a discussion of computational complexity.Source: §I · p. 1
• A quantum image can be encoded in at least n qubits and transformed by unitary evolution, whereas a classical M × M image uses at least 2n bits and matrix computation.The figure defines n = ⌈log2(M2)⌉ for the classical encoding.Source: §I · p. 2, Figure 1 · p. 2

A. Quantum Convolution Layer

• Pixel positions are encoded in computational-basis states and pixel values in amplitudes, forming a pure state with n = ⌈log2(ML)⌉ qubits.The amplitudes are normalized pixel values; basis states beyond the ML pixels receive zero amplitude.Source: A. Quantum Convolution Layer · p. 2
• A filter mask W maps an input image state to an output image state, with each interior output pixel computed from a weighted neighborhood.For a 3 × 3 mask, the text gives the weighted sum over neighboring pixels; it says the construction generalizes to arbitrary m × m masks.Source: A. Quantum Convolution Layer · p. 2
• Because the linear filtering operator is generally non-unitary and direct implementation scales exponentially with system dimensions, the authors adjust its boundary matrix elements to reduce gate complexity.The boundary changes affect the first and last rows and columns of three matrices and are described as allowable for image processing.Source: A. Quantum Convolution Layer · p. 3
• The ancilla-controlled unitary construction can realize a linear combination of 16 filter masks, with each mask associated with a distinct ancilla state.The circuit prepares an ancilla superposition, applies controlled operations, and uncomputes the ancilla registers.Source: A. Quantum Convolution Layer · p. 4
• For a single desired mask, measuring the ancilla in state |0000⟩ yields a state proportional to the expected filtered result.The processed state retains unchanged edge elements, while evaluated interior pixel values match those of the output under the original operator.Source: A. Quantum Convolution Layer · pp. 4–5
• Four ancillary qubits implement the general adjusted filter operator as a linear combination of nine unitary operators, independent of image dimensions.The paper also notes that a nine-energy-level ancillary system can be used.Source: A. Quantum Convolution Layer · pp. 4–5

B. Quantum Pooling Layer

• Average pooling reduces representation size and parameter count; 2×2 pooling with stride 2 is realized by ignoring the last and m-th qubits.The layer averages each feature-map patch.Source: B. Quantum Pooling Layer · p. 5

C. Quantum Fully Connected Layer

• The quantum fully connected layer is a parameterized Hamiltonian consisting of identity operators and Pauli σz operators acting on qubits.Source: C. Quantum Fully Connected Layer · p. 5
• The network’s final output is the expectation value f(p) = ⟨p|H|p⟩, which is nonlinearly mapped by an active function to R(f(p)).Hamiltonian parameters are updated by gradient descent using ∂f(p).Source: C. Quantum Fully Connected Layer · p. 6
• Parameters in the S matrix can be trained through classical backpropagation.Source: C. Quantum Fully Connected Layer · p. 6

A. Image Processing: Edge Detection, Image Smoothing and Sharpening

• The simulated quantum and classical convolutional layers produce exactly the same output images for the spatial filtering tasks (Fig. 3).The simulation covers edge detection, image smoothing, and sharpening.Source: A. Image Processing: Edge Detection,… · p. 6
• For spatial filtering, one specific filter mask is used, and the task succeeds when measurement of the ancillary register yields |0⟩.The three demonstrated tasks use fixed masks Wde, Wsm, and Wsh for edge detection, smoothing, and sharpening, respectively.Source: A. Image Processing: Edge Detection,… · p. 6

B. Handwritten Number Recognition

• The simulations use a complete QCNN with convolutional, pooling, and fully connected layers for MNIST recognition of ‘1’ versus ‘8’ and ‘0’–‘9’.The 28 × 28 images are zero-padded to 32 × 32; QCNN uses a 10-qubit work register and 4 ancillary qubits, with pooling by abandoning the 4th and 9th work qubits.Source: B. Handwritten Number Recognition · pp. 6–7
• For two-class recognition, QCNN has 37 learnable Hamiltonian parameters versus 256 CNN fully connected parameters; for ten classes, the counts are 10 × 37 and 10 × 256.Both models use 9 learnable convolutional parameters; CNN applies average pooling, while QCNN measures Hamiltonian expectations on the remaining work qubits for classification.Source: B. Handwritten Number Recognition · p. 7
• Across 100 repeated runs, QCNN and CNN provide similar performance, while QCNN involves fewer parameters and has a smaller accuracy fluctuation range.The reported averages and accuracy ranges come from repeated noisy-QCNN, noise-free-QCNN, and CNN runs under the same construction.Source: B. Handwritten Number Recognition · p. 7
• The two-class task uses 5,000 training and 2,100 test images, while the ten-class task uses 60,000 training and 10,000 test images.Each training epoch randomly selects 100 images, and training runs for 50 epochs, so accuracy fluctuates.Source: B. Handwritten Number Recognition · p. 7

IV. ALGORITHM COMPLEXITY

• O((log_2 M)^6) basic gates implement the controlled operations for convolution, yielding an exponential gate-complexity speedup over the O(M^2) classical convolutional layer.State preparation takes O(poly(log_2(M^2))) steps when amplitudes and probabilities are classically efficiently calculable; alternatively, qRAM preparation takes O(log_2(M^2)) after the quantum memory cell is established.Source: §IV · pp. 7–8
• O(e) measurements are required in fully connected layers, where e is the number of parameters in the Hamiltonian.Source: §IV · p. 8
• Ancillary qubits scale as O(log_2(m^2)) and work qubits as O(log_2(M^2)), where m is the filter-mask dimension.Source: §IV · p. 8

• The authors report exponential speed-ups in gate complexity and similar handwritten-number recognition performance with fewer parameters than classical algorithms.They argue the approach has significant advantages for large data.Source: §V · p. 8
• The paper demonstrates the method through image processing and handwritten-number recognition, and maps specific classical convolutional kernels to quantum circuits.The mapping is presented as a bridge between QCNN and CNN.Source: §V · p. 8
• The authors describe the algorithm as general and implementable on programmable quantum computers, including superconducting, trapped-ion, and photonic platforms.They suggest it has potential to outperform classical counterparts for large data.Source: §V · p. 8

Appendix A: Adjusted operator U′ can provide enough information to remap the output imagine.

• The adjusted operator U′ differs from U in edge elements, yet its evolution provides enough information to remap the output image.Source: Appendix A · p. 9
• After remapping, the output image G′ matches G at pixels with 2 ≤ i, j ≤ M − 1; differences are confined to the edges.The elements affected by the differences between U′ and U are identified as edge elements.Source: Appendix A · p. 9

Appendix B: Decomposing operator Q into basic gates

• The filter operator U′ consists of nine operators Q1 through Q9, each formed as a tensor product of two of three operators.The supplied passage does not identify the three operators.Source: Appendix B · p. 9
• E2 is an M × M identity matrix and does not need further decomposition.Source: Appendix B · p. 9
• The n-qubit operator E1 is expressed using O(n^3) CNOT gates and Pauli X gates, while E3 is implemented by the inverse combination because E3 = E1†.The passage defines n = log2(M^2).Source: Appendix B · p. 9
• Because E2 needs no further decomposition, the Qk decomposition uses the E1 and E3 gate constructions.The passage states that Qk can be implemented using no more than O(n^6) basic gates.Source: Appendix B · p. 9
• No more than O((log2 M)^6) basic gates implement the controlled Qk operation, ignoring a constant number.The passage also expresses this bound as O(n^6).Source: Appendix B · pp. 9–10