AutoQML: A Framework for Automated Quantum Machine Learning

TL;DR

AutoQML tackles the high entry barrier of quantum machine learning by adapting AutoML principles to automate end-to-end QML pipelines.It introduces a modular framework built on Ray Tune, Optuna, and the sQUlearn QML library to orchestrate data cleaning, preprocessing, model selection, and hyperparameter optimization across quantum and classical models.The authors benchmark AutoQML on four industrial supervised-learning use cases (time-series and image tasks, plus tabular and time-series regression), showing pipelines that are competitive with classical models and comparable to manually crafted quantum solutions.The work highlights AutoQML's potential to democratize QML, accelerate prototyping, and provide a benchmarking platform for QML research.

Abstract

Automated Machine Learning (AutoML) has significantly advanced the efficiency of ML-focused software development by automating hyperparameter optimization and pipeline construction, reducing the need for manual intervention. Quantum Machine Learning (QML) offers the potential to surpass classical machine learning (ML) capabilities by utilizing quantum computing. However, the complexity of QML presents substantial entry barriers. We introduce \emph{AutoQML}, a novel framework that adapts the AutoML approach to QML, providing a modular and unified programming interface to facilitate the development of QML pipelines. AutoQML leverages the QML library sQUlearn to support a variety of QML algorithms. The framework is capable of constructing end-to-end pipelines for supervised learning tasks, ensuring accessibility and efficacy. We evaluate AutoQML across four industrial use cases, demonstrating its ability to generate high-performing QML pipelines that are competitive with both classical ML models and manually crafted quantum solutions.

Figures (4)

• Figure 1: Architecture overview of the AutoQML framework. Data is supplied by the user. Using Ray and Optuna, AutoQML constructs a pipeline that is optimized over a preconfigured search space. A loss value $l_i$ is obtained for each configuration $\vec{\lambda}_i$, which consists of data cleaning, preprocessing, and a model with hyperparameters evaluated using a simulator or a real quantum computer (QC). After a given budget is exhausted, the best-performing pipeline is returned to the user. Optional pipeline steps are indicated as dashed boxes.
• Figure 2: Example code for fitting a tabular regression pipeline. Here, it is assumed that the training data is supplied as X_train with corresponding targets y_train. Within AutoQML, the data is split into a test and validation set. Options such as the time budget timedelta for the optimization or the backend for execution of the QML algorithms can be specified. In the example, the preset configuration "quantum_regression" is used to restrict the search space to quantum computing based regression algorithms only.
• Figure 3: Performance of AutoQML (boxes) for two different time budgets $T_1$ and $T_2$. Additionally, manual QML pipelines (red, dashed) and classical models (black, dotted) are depicted. (a) shows the balanced accuracy (higher is better) for the time series classification. (b) shows the accuracy (higher is better) for the image classification. (c) shows the mean absolute percentage error (MAPE, lower is better) for the tabular regression, and (d) shows the mean absolute scaled error (MASE, lower is better) for the time series forecasting. For the box plots, points that are outside $1.5\times$ the inter-quartile range are shown as circles, and the lines inside the boxes denote the sample median.
• Figure 4: Application of the best AutoQML pipelines (blue) from Fig \ref{['fig:autoqml-results']} on the respective use cases. The application of the manually created models is shown in red. (a) shows signal in the upper part of the figure. The bars bellow show the presence or absence of events which are classified by the models. (b) shows the two principal components with the larges singular values of the test set of the image classification use case. The points that have been missclassified by the AutoQML (cross, blue) or the manual pipeline (plus, red) are shown in addition. The classes 0 (no slat) and 1 (slat) are shown in different colors. Figure (c) and (d) depicts the prediction vs. the true target values of the tabular regression and time series forecasting, respectively.