Qiboml: towards the orchestration of quantum-classical machine learning

TL;DR

Qiboml presents an open-source framework for orchestrating quantum and classical components in hybrid machine learning, built on the Qibo ecosystem and integrable with TensorFlow and PyTorch. The paper details a modular software design with model-building blocks, framework interfaces, differentiable backends, and support utilities, enabling end-to-end QML workflows from circuit construction to hardware execution. Through regression, VQE, real hardware experiments, and tensor-network scaling, Qiboml demonstrates versatile experimentation, real-time error mitigation, and calibration-aware training within a single open-source stack. Benchmarking against PennyLane shows competitive performance across differentiation strategies, highlighting Qiboml’s potential as a practical platform for research and industrial-scale hybrid quantum-classical ML tasks.

Abstract

We present Qiboml, an open-source software library for orchestrating quantum and classical components in hybrid machine learning workflows. Building on Qibo's quantum computing capabilities and integrating with popular machine learning frameworks such as TensorFlow and PyTorch, Qiboml enables the construction of quantum and hybrid models that can run on a broad range of backends: (i) multi-threaded CPUs, GPUs, and multi-GPU systems for simulation with statevector or tensor network methods; (ii) quantum processing units, both on-premise and through cloud providers. In this paper, we showcase its functionalities, including diverse simulation options, noise-aware simulations, and real-time error mitigation and calibration.

Figures (13)

• Figure 1: Schematic representation of a quantum machine learning pipeline with Qiboml.
• Figure 2: Custom differentiation example using Keras as interface, the parameter-shift rule Schuld2019 to calculate gradients, and the Qibolab hardware backend to execute circuits.
• Figure 3: Schematic representation of a real-time quantum error mitigation procedure. The mitigation map is periodically updated during the training, and it is used to mitigate the expectation values calculated for predictions and gradients. Those values are then utilized within the hybrid machine learning procedure.
• Figure 4: Parametric circuit composed of $L$ layers of rotations. $R_x$ gates are used to encode the data (Qiboml's encoders), while $R_y$ and $R_z$ gates are used as trainable gates.
• Figure 5: Four trainings are performed with the same initial configuration shown in Table \ref{['tab:init_config_train']}, each following a different strategy: noiseless and exact simulation (green), noiseless with shot-noisy simulation (blue), noisy with shot-noisy simulation (red), and noisy, shot-noisy with real-time mitigation (yellow). The approximations are compared with the target theoretical function introduced in Eq. \ref{['eq:target']}. Solid curves and uncertainty intervals are obtained from the median and median absolute deviation of twenty repetitions, each starting from a different random seed.