Speakers
Description
Distributed quantum machine learning (DQML) provides a promising route for training quantum and hybrid quantum-classical models across multiple clients while keeping data local. This setting is especially relevant when data are heterogeneous and distributed across institutions, devices, or quantum processors, and direct data sharing is restricted or inefficient. However, the performance of DQML depends on the local quantum models and on how client updates are aggregated at the server. A fixed aggregation rule, such as simple averaging, may perform well under balanced and clean settings, but it can become less effective when clients have non-IID data, different sample sizes, noisy updates, unequal circuit depths, hardware-dependent errors, unstable training behavior, or faulty and adversarial clients. These challenges are more pronounced in quantum learning, where shot noise, barren-plateau effects, circuit expressivity, and device noise can cause large variation in update quality across clients and communication rounds.
In this work, we propose an adaptive aggregation-selection framework for DQML based on reinforcement learning (RL) and quantum reinforcement learning (QRL). Instead of using the same aggregation strategy throughout training, we formulate aggregation as a sequential decision-making problem. At each communication round, an agent observes the current training state and selects the aggregation technique expected to produce the most reliable global update. The state representation can include client-level losses, validation accuracy, update divergence, parameter variation, client participation, gradient statistics, model confidence, and quantum resource indicators. Classical RL agents provide a stable baseline, while QRL agents use parameterized quantum circuits to represent the policy or value function.
| Topic | Hybrid quantum-classical approaches for learning and optimization |
|---|---|
| Additional Topic | Quantum variational circuits and their applications |