Hybrid Classical-Quantum architecture for vectorised image classification of hand-written sketches

TL;DR

The paper investigates a vector-based sketch recognition task using a hybrid classical-quantum model (QuantumDraw) to evaluate QML in the NISQ era. It combines a sequential-stroke encoder with a five-qubit Hardware-Efficient Ansatz, followed by a classical classifier, and compares against a raster-based HQNN-Parallel and a simple classical baseline. While the QuantumDraw model achieves competitive accuracy, ablations (QD-Frozen, QD-Sep) suggest that entanglement may not be essential for performance, underscoring that this work aims to explore synergy rather than establish quantum advantage. The vector-sketch framework provides a compact, temporally structured benchmark for advancing QML research, guiding future work on quantum data representations and memory-enabled processing.

Abstract

Quantum machine learning (QML) investigates how quantum phenomena can be exploited in order to learn data in an alternative way, \textit{e.g.} by means of a quantum computer. While recent results evidence that QML models can potentially surpass their classical counterparts' performance in specific tasks, quantum technology hardware is still unready to reach quantum advantage in tasks of significant relevance to the broad scope of the computer science community. Recent advances indicate that hybrid classical-quantum models can readily attain competitive performances at low architecture complexities. Such investigations are often carried out for image-processing tasks, and are notably constrained to modelling \textit{raster images}, represented as a grid of two-dimensional pixels. Here, we introduce vector-based representation of sketch drawings as a test-bed for QML models. Such a lower-dimensional data structure results handful to benchmark model's performance, particularly in current transition times, where classical simulations of quantum circuits are naturally limited in the number of qubits, and quantum hardware is not readily available to perform large-scale experiments. We report some encouraging results for primitive hybrid classical-quantum architectures, in a canonical sketch recognition problem.

Figures (5)

• Figure 1: Sketch Samples from QuickDraw Dataset: This figure displays a curated selection of sketch drawings from the QuickDraw dataset. Each row represents a distinct category: cameras in the top row, calculators in the middle row, and cellphones in the bottom row. These examples illustrate the variety of drawing styles and levels of detail present in the dataset.
• Figure 2: Overview of QuantumDraw Framework: The vectorised drawing include the Bezier point coordinates and a flag for end-of-stroke. Sequential Stroke Data Handler is constituted by two LSTM units. Data is half-sized through max-pooling, and flattened to a vector feeding the Quantum Processing Module, with three fully connected layers, ending in five cells to feed the quantum circuit. A final fully connected layer provides the classification.
• Figure 3: Comparative Analysis of the QuantumDraw Model: We show training and validation performance on (left) baseline model consisting on two LSTM layers plus feed-forward post-processing, (center) raster-image classification carried out by the HQNN-Parallel circuit (see main body), and (right) vectorized-image classificaton carried out by our QuantumDraw model, explained in Sec. \ref{['ssec:qdraw']}.
• Figure 4: Average QuantumDraw performance. We show learning curves of our hybrid model, initialized over 50 different seeds. We show training/validation (left/right) loss per epoch, along with minimum, maximum, and mean training losses. The validation dataset consists of a set of data points not seen by the model during training, used to evaluate the model's ability to generalize to new, unseen data.
• Figure 5: Quantumness testing. We show learning curves for two models that test the quantumness/classicality of the QuantumDraw model. (Left): QD-frozen quantum parameters associated to rotation-angles are randomly initialized and not modified during training, while the classical part of the model needs to adapt its weights to optimize the cost value. (Right:) we show the QD-Sep model, in which CNOTs are removed from the quantum circuit layout.