Speaker
Description
Residual learning transformed the training of deep classical neural networks, yet its adoption in quantum machine learning has remained limited because existing quantum residual architectures rely on post-selection, which prevents fully differentiable optimisation and makes execution increasingly inefficient as circuit depth grows. We introduce a new quantum machine learning architecture, the post-selection-free Quantum Residual Neural Network (QResNet), which implements differentiable residual learning through deterministic ancilla-controlled residual channels. Unlike previous QResNet formulations, the proposed architecture eliminates the sample-and-discard step while retaining the optimisation benefits of residual learning, enabling end-to-end gradient-based optimisation of both variational circuit parameters and residual strengths. We further characterise the trainability of the architecture and show that its optimisation behaviour is governed jointly by the residual depth and the depth of the variational blocks, identifying a polylogarithmic regime that avoids barren plateaus while remaining beyond the regime of provably classically simulable shallow circuits. On the ten-class MNIST benchmark, the proposed architecture achieves comparable classification accuracy using approximately six times fewer quantum gates than a 200-layer quantum variational classifier (QVC); adding only five residual blocks to a 30-layer variational backbone improves accuracy from 65% to 80%, while also demonstrating improved resistance to black-box transfer attacks. These results establish differentiable quantum residual learning as a practical architectural primitive for trainable, hardware-efficient, and robust quantum machine learning.
| Topic | Quantum variational circuits and their applications |
|---|---|
| Additional Topic | Theoretical foundations of quantum learning |