On the Importance of Fundamental Properties in Quantum-Classical Machine Learning Models

We present a systematic study of how quantum circuit design, specifically the depth of the variational ansatz and the choice of quantum feature mapping, affects the performance of hybrid quantum-classical neural networks on a causal classification task. The architecture combines a convolutional neural network for classical feature extraction with a parameterized quantum circuit acting as the quantum layer. We evaluate multiple ansatz depths and nine different feature maps. Results show that increasing the number of ansatz repetitions improves generalization and training stability, though benefits tend to plateau beyond a certain depth. The choice of feature mapping is even more critical: only encodings with multi-axis Pauli rotations enable successful learning, while simpler maps lead to underfitting or loss of class separability. Principal Component Analysis and silhouette scores reveal how data distributions evolve across network stages. These findings offer practical guidance for designing quantum circuits in hybrid models. All source codes and evaluation tools are publicly available.

💡 Research Summary

This paper presents a systematic investigation of two pivotal design choices in hybrid quantum‑classical neural networks (HQNNs): the depth of the variational ansatz and the structure of the quantum feature map. The authors construct a hybrid architecture that first processes 8×8 histogram images derived from a causal inference dataset using a three‑layer convolutional neural network (CNN). The CNN extracts hierarchical features, reduces them to a 64‑dimensional vector, and then projects this vector onto a three‑qubit space via a linear layer.

The quantum layer, built with Qiskit‑Machine‑Learning, consists of a data‑encoding circuit (feature map) followed by a TwoLocal variational ansatz. Nine distinct feature maps are examined, ranging from simple Z‑rotation encodings to more expressive multi‑axis rotations (Rx, Ry, Rz) and entangling ZZ gates. The ansatz employs alternating Ry‑Rz single‑qubit rotations and linear CNOT entanglement, with depth d varied from 1 to 5, yielding 2·d·n_q trainable parameters (n_q = 3). Gradients are obtained via the parameter‑shift rule, enabling end‑to‑end training with stochastic gradient descent.

Experiments are organized along two axes. First, the impact of ansatz depth on training loss, validation accuracy, and stability is measured. Increasing depth from 1 to 3 dramatically reduces loss and improves accuracy by roughly 5 percentage points, indicating enhanced expressivity. Beyond depth = 3, gains plateau; deeper circuits exhibit slower convergence and higher variance, suggesting the onset of barren‑plateau effects. Second, each of the nine feature maps is evaluated at a representative depth (d = 3). Only maps that incorporate multi‑axis Pauli rotations achieve high performance (≈92 % accuracy, silhouette ≈ 0.68). Simpler Z‑only encodings underperform (≈68 % accuracy) and suffer from dimensional collapse, as confirmed by principal component analysis (PCA) where the leading components explain less variance and class separation is poor.

The authors also analyze gradient statistics: shallow ansatz (d = 3) maintain sufficient gradient variance for effective SGD, whereas deeper circuits (d ≥ 4) show near‑zero variance, corroborating the barren‑plateau hypothesis.

Based on these findings, the paper proposes practical design guidelines for HQNNs: (1) limit ansatz depth to 3–4 layers to balance expressivity and trainability; (2) employ feature maps with multi‑axis rotations and at least one entangling operation to ensure rich Hilbert‑space embeddings; (3) use adjacent CNOT entanglement to capture inter‑qubit correlations; (4) mitigate barren‑plateau risks in deeper circuits through careful parameter initialization and possible layer‑wise regularization.

In conclusion, the study demonstrates that while deeper variational circuits can improve generalization up to a point, the choice of feature mapping exerts a more decisive influence on learning success. The results provide concrete, reproducible guidance for constructing effective hybrid quantum‑classical models, and all source code and evaluation tools are publicly released.