Speaker
Description
Integrating quantum computation into deep neural networks remains a central challenge in quantum machine learning, because quantum measurements are probabilistic and do not naturally fit standard gradient-based optimization pipelines. This work investigates trainable quantum layers as modular representation-learning components that can be inserted into pre-trained classical neural networks without requiring end-to-end quantum backpropagation.
We introduce two architectures. The first is a circuit-based quantum layer in which a parameterized quantum circuit acts as a latent-space transformer. Its architecture is optimized automatically using Progressive Widening Monte Carlo Tree Search, guided by task-specific cost functions. The second is a trainable Open Quantum System layer, where learning arises from a compact system–environment interaction rather than from a deep variational quantum circuit.
Both layers are evaluated in AutoEncoder-based anomaly detection pipelines, where the quantum layer is inserted at or after the bottleneck and the transformed latent representations are assessed using an Isolation Forest or an OQS-based classifier. Results on benchmark datasets show that quantum transformations can improve latent-space separability and anomaly detection performance compared with classical baselines.
| Topic | Hybrid quantum-classical approaches for learning and optimization |
|---|