Towards Federated Learning on the Quantum Internet

Leo Sünkel, Michael Kölle, Tobias Rohe, Thomas Gabor

TL;DR

This paper explores the feasibility of Quantum Federated Learning (QFL) as a primary application for the future Quantum Internet. By simulating decentralized training across networks of heterogeneous quantum devices, the authors demonstrate that QFL is a viable alternative to centralized training, though the specific network topology and weight-handling methods significantly impact model stability and performance.

What Problem Does This Paper Address?

As we move from monolithic quantum computers toward a "Quantum Internet," we face a major question: What will we actually use it for? While distributed quantum computing and quantum key distribution are often cited, the field of Quantum Machine Learning (QML) offers another promising avenue.

However, a real-world quantum network won't be a collection of identical machines. It will likely consist of heterogeneous nodes—devices with different architectures and varying numbers of qubits. This paper addresses the challenge of training machine learning models across such a diverse, decentralized network without compromising data privacy, a paradigm known as Federated Learning. The authors specifically investigate how network constraints and different communication topologies (how nodes are connected) affect the learning process.

Methods

The authors employ Variational Quantum Circuits (VQCs) as their underlying machine learning models. VQCs are hybrid classical-quantum algorithms where a quantum circuit's parameters are optimized using a classical optimizer.

To test federated learning, they compare two primary network architectures:

  1. Star Topology: A central "global model" server aggregates weights from multiple clients, updates them, and redistributes them.
  2. Ring Topology: Clients are arranged in a circle. Each client trains on its local data and passes its weights directly to the next client in the sequence.

A key technical hurdle addressed is heterogeneity. If Client A has 4 qubits and Client B has 6, their weight vectors are different sizes. The authors implemented an ad-hoc mapping strategy: when moving to a smaller device, superfluous weights are discarded; when moving to a larger one, missing weights are filled in using the device's previous parameters.

Datasets

The study evaluates QFL across three distinct types of data, focusing on binary classification:

  • Moons: A synthetic 2D dataset from scikit-learn used for basic non-linear classification.
  • FashionMNIST: Grayscale images of clothing (e.g., Trousers vs. Pullovers).
  • PneumoniaMNIST: A medical dataset consisting of chest X-ray images.

Images were resized and preprocessed to fit the varying qubit capacities of the simulated quantum nodes (ranging from 4 to 10 qubits).

Key Insights and Ideas

  • QFL is a "Valid Alternative": The experiments show that decentralized quantum training can achieve performance nearly identical to centralized quantum training (the "Quantum Baseline").
  • The Topology Trade-off: The choice of network structure matters deeply. The Star topology tended to reach higher training accuracy but was significantly more "unstable" (volatile) during the process. Conversely, the Ring topology was more stable and actually performed better on unseen test data.
  • Heterogeneity is Manageable: Even with the "ad-hoc" method of truncating or padding weight vectors to account for different qubit counts, the models were able to learn effectively. This suggests that a future Quantum Internet doesn't require perfectly uniform hardware to be useful for ML.
  • Classical vs. Quantum Baselines: While classical models still outperform these quantum models (largely due to having vastly more parameters), the QFL models held their own against their centralized quantum counterparts, proving the federated approach doesn't inherently "break" the quantum advantage.

Key Results

  • MCSD (Single Dataset): In the "Moons" experiment, the QFL approach converged quickly and matched the quantum baseline's accuracy almost perfectly (approx. 85-90% accuracy).
  • MCMD (Multiple Datasets): On image data, the results were more nuanced. While the classical baseline was superior, the QFL approaches produced "acceptable" results.
  • Stability: In the Star topology, training accuracy fluctuated significantly between epochs, whereas the Ring topology showed a much smoother convergence curve.

Limitations and Future Work

The authors identify several areas for deeper investigation:

  • Simulation vs. Reality: These experiments were conducted via simulation. Real-world quantum noise and decoherence on actual hardware will introduce significant new challenges.
  • Classical Communication: In this study, model weights were exchanged over classical channels. A "true" Quantum Internet application might involve exchanging weights via quantum teleportation, which requires further study regarding entanglement fidelity and resource costs.
  • Weight Mapping: The method for resizing weight vectors between different-sized QPUs was relatively simple. The authors suggest that more sophisticated "aggregation" methods are needed to optimize how knowledge is transferred between devices of different capacities.
  • Scalability: The study used a small number of nodes (3-4). How these topologies behave when scaled to hundreds or thousands of nodes remains an open question.

Summary updated

Abstract

While the majority of focus in quantum computing has so far been on monolithic quantum systems, quantum communication networks and the quantum internet in particular are increasingly receiving attention from researchers and industry alike. The quantum internet may allow a plethora of applications such as distributed or blind quantum computing, though research still is at an early stage, both for its physical implementation as well as algorithms; thus suitable applications are an open research question. We evaluate a potential application for the quantum internet, namely quantum federated learning. We run experiments under different settings in various scenarios (e.g. network constraints) using several datasets from different domains and show that (1) quantum federated learning is a valid alternative for regular training and (2) network topology and nature of training are crucial considerations as they may drastically influence the models performance. The results indicate that more comprehensive research is required to optimally deploy quantum federated learning on a potential quantum internet.
BibTeX
@inproceedings{sunkel2024towards,
  author = {Leo Sünkel and Michael Kölle and Tobias Rohe and Thomas Gabor},
  title = {{Towards Federated Learning on the Quantum Internet}},
  booktitle = {ICCS},
  year = {2024},
  url = {https://openreview.net/forum?id=oH8N5SweEQ},
}

Loading related content...

Figure