Title: Error-Corrected Quantum Convolutional Neural Networks
Department: Electrical Engineering and Computer Science, Ohio University
Duration: August 2024 - Present
Informally: We research methods to make quantum computers more reliable. Specifically, we focus on improving the accuracy of quantum systems against errors and noise from quantum phenomena and its environment. Our work integrates simulations and experiments, allowing us to develop more practical and scalable quantum technologies.
Formally: Our research focuses on embedding Quantum Convolutional Neural Networks (QCNNs) within state-of-the-art Quantum Error Correction (QEC) codes to evaluate their performance on noisy intermediate-scale quantum (NISQ) devices. The goal is to leverage modern QEC techniques to correct for quantum error and therefore improving QCNN learning convergence and reliability under realistic hardware conditions. We test the effects of various QEC codes through a combination of large-scale simulations on university high-performance computing resources. Our first results from this line of work are now available as a preprint, Low-Overhead Error-Corrected QCNNs Using Bivariate Bicycle Codes, summarized below.
Name: Low-Overhead Error-Corrected QCNNs Using Bivariate Bicycle Codes
[arXiv]
Authors: Alejandro J. Rosales and Animesh Yadav
The question: Can you error-correct quantum machine learning (QML) models without paying the qubit tax of
the surface code?
Informally: Quantum computers make mistakes, and the standard way of fixing those mistakes costs an enormous number of extra qubits to protect even a small machine learning model such that it is out of reach on today's hardware. We ask whether a cheaper family of error-correcting codes can keep a quantum machine learning model learning while noise is present, and we find that it can, using 29 qubits where the conventional approach would realistically need tens of thousands or more.
Formally: In this research, we investigate whether bivariate bicycle (BB) codes, a family of Quantum Low-Density Parity Check (qLDPC) codes, can help lower the training loss of QML circuits facing quantum noise. Our demonstration case is the Quantum Convolutional Neural Network (QCNN), which is a deep parameterized circuit full of non-Clifford rotations. We pair a distance-4 [[18,4,4]] BB code with a feed-forward neural decoder interleaved between the QCNN's layers, implementing both an 11-qubit transversal encoding for error mitigation and a full 29-qubit version with error correction. (The 11 and 29 qubit count comes from mapping the 4 logical qubits onto a single 11-qubit data register rather than two separate 9-qubit registers.) The error-mitigated design alone already helps: at p=0.001 simulated noise, the unprotected 4-qubit QCNN stops learning (loss 0.632 to 0.652) while the transversal version keeps training and settles near 0.348. For comparison, a surface code at d=4 would need roughly 64 physical qubits for the same 4 logical qubits, plus 400-1,000 more per magic state distillation factory to support the non-Clifford rotations, requiring realistically 10⁴-10⁶ physical qubits in total. Ours does it in 29, with the neural decoder standing in for the factory, at the cost of a lower threshold (~0.3% vs ~0.7%). The decoder still needs joint training with the QCNN, as the current method of training causes the decoder to overcorrect. This matters for scaling, given qLDPC codes hold qubit overhead close to constant rate as the model grows. We expect the strategy to extend to other deep gate-based QML circuits. Thanks to Jorma Kilpi and Olli Apilo of the Technical Research Centre of Finland (VTT) for the fruitful discussions! Feedback is always welcome :)
Acknowledgements: Thanks to Jorma Kilpi and Olli Apilo of the Technical Research Centre of Finland (VTT) for the fruitful discussions. Feedback is always welcome!
Name: Final Paper for Hardware for Deep Learning: Quantum Algorithms
for Deep Convolutional Neural Networks
[PDF]
Preview:
Computer scientists utilize principles of quantum mechanics, mathematics, and computer science in quantum
computing. By borrowing concepts from each field scientists can rigorously define both a broad and narrow
theoretical model of a quantum computer and later apply it to the real world. These theoretical models, such as
the result...
Name: Quantum Algorithms for Deep Convolutional Neural Networks
[PDF]
Abstract:
Current problem: it is difficult to implement non linearities with quantum
unitaries
Suggested solution: a new quantum tomography algorithm with norm guarantees,
and new applications of probabilistic sampling in the context of information
processing
Goal: The QCNN is particularly interesting for deep networks and could
allow new frontiers in image recognition, by using more or larger convolution
kernels, larger or deeper inputs
Name: Quantum Enhanced Feature Space [PDF]
Abstract:
Current problem: limitations on successful solution for problems when feature
space becomes large high-dimensional
Suggested solution: utilize controlled entanglement and interference to exploit
exponentially growing quantum state space
Goal: present new class of tools for exploring the applications of noisy
intermediate scale quantum computers to machine learning for improved
computational power and efficacy
Name: Quantum Error Code Correction Background
[PDF]
Abstract:
This presentation introduces the essential codes of quantum error correction, like the 3-qubit repetition code
and 9-qubit Shor code.
Name: Quantum Information Background
[PDF]
Abstract:
This presentation introduces the essential preliminaries of quantum information processing, from the basics of
qubits and quantum gates to the critical role of quantum error correction in making quantum computing practical.
Name: Modular-Network of Adaptive Topologies [Website]
Description: Modular-Network of Adaptive Topologies MAT is a biologically grounded spiking neural
network simulator that models the brain as seven anatomically motivated regions, each populated by neurons whose
intrinsic properties are determined by their neurotransmitter identity. Six neurotransmitters are represented:
glutamate, GABA, acetylcholine, dopamine, serotonin, and norepinephrine.
Name: Credit Card Fraud Detection using a Quantum Support Vector Machine [Code]
Description: Applying a Quantum Support Vector Machine for credit card fraud detection in Qiskit. The
applied
QSVM makes use of quantum enhanced feature space optimization based on the research paper, Supervised
Learning
with Quantum Enhanced Feature Spaces. The classifier used is a Variational Quantum Classifier with
29-dimensional ZZ feature mapping for a 2-qubit quantum kernel.
Name: My Cloud A.I. Controlling a Local Robot Body [Website]
Description: A robot I built that communicates with a remote server running artifical intelligence
software I designed and programmed from scratch.
Meaning, you can talk to the exact same Jarcey that is running in this robot, either on this website context
switching, on jarcey's own website, or any device with an internet connection. This is because all
communication, data, and programs are transimitted to, from, and processed on the same one remote server.