Novel quantum residual neural networks overcome the limitations of existing models and provide a path to practical quantum machine learning. Amena Khatun and colleagues at the University of Melbourne demonstrate a hardware-efficient architecture that implements residual connectivity without relying on post-selection, an important advance in the field. The model achieves comparable accuracy of 99% for binary classification and 80% for multiclass classification on benchmark datasets such as MNIST, CIFAR, and SARFish, and requires 10 times fewer quantum gates than standard variational models. This reduction in complexity is essential for implementation on quantum processors in the near future, and the model further exhibits promising adversarial robustness, addressing key requirements for reliable quantum machine learning applications.
Quantum residual networks can reduce gate complexity by a factor of 10 for image classification
The number of gates for quantum image classification has been reduced by a factor of 10 using a new quantum residual neural network architecture. This advance is important because previous variational quantum classifiers required gates that increased exponentially with circuit depth, hindering their implementation on quantum processors in the near future. Researchers at the University of Melbourne have developed a model that achieves 99% accuracy in binary image classification on datasets including MNIST, CIFAR and SARFish, and 80% accuracy in multi-class tasks. At the same time, we also addressed the problem of sterile plateaus, which is a common limitation in training quantum machine learning models.
Using just 200 quantum gates, we achieved 80% accuracy on SARFish images, a remote sensing application, and observed consistent learning dynamics across diverse datasets. Analysis of the full-scale MNIST dataset reveals that combining 30 layers of a standard variational quantum classifier with five residual blocks increases accuracy to 99%, matching with a reduced number of gates while maintaining previously achieved accuracy. This performance was achieved with 10 times fewer gates than comparable deep variational circuits, increasing hardware efficiency for resource-constrained quantum processors.
The model also demonstrated adversarial durability and maintained accuracy when tested with attacks transferred from classical machine learning models. Quantum machine learning has potential for image classification and other applications, but to make this a reality, current quantum hardware requires circuits small enough. This model shows a reliance on amplitude encoding to transform classical data into quantum form. Although effective, this method can become limited as dataset size increases and more complex data types are used.
Although these results were obtained through simulations and do not demonstrate performance on real quantum hardware, where qubit limitations and noise remain major obstacles, this model establishes a path toward scalable quantum machine learning by addressing key limitations of existing variational models. The deterministic residual connection implemented by a combination of identity unitary and variational unitary allows fully differentiable training without stochastic post-selection. This efficiency is critical because current quantum computers have limited processing power. In other words, fewer gates means more complex problems can be solved with existing hardware. Achieving comparable accuracy to standard models for image classification tasks involving difficult datasets like SARFish, while requiring 10 times fewer quantum gates, represents an important hardware efficiency improvement for near-term quantum processors.
Researchers have developed a quantum residual neural network that achieves 99% accuracy on binary image classification and 80% on multiclass tasks using datasets such as MNIST, CIFAR, and SARFish. This model requires 10 times fewer quantum gates than the equivalent standard variational model while maintaining similar accuracy, making it more hardware efficient. This architecture alleviates a common problem in quantum learning known as the barren plateau and also demonstrates robustness against adversarial attacks. This work presents a new approach to building trainable and efficient quantum machine learning models suitable for short-term quantum processors.
