SPATE: Spiking-Phase Adaptive Temporal Encoding for Quantum Machine Learning

1. Problem Statement

Current Quantum Machine Learning (QML) pipelines predominantly rely on static encoding methods (e.g., Angle Encoding and Amplitude Encoding) to map classical data into quantum states. These methods face significant limitations:

• Loss of Temporal Structure: Static encodings treat data points as independent vectors, failing to capture temporal dynamics or spike-based information inherent in many real-world datasets.
• Suboptimal Geometry: Under constrained resources (limited qubits and shallow circuits), static encodings often produce embeddings where classes are not well-separated or aligned. This forces the trainable Quantum Neural Network (QNN) to learn complex decision boundaries from a poor initial representation, limiting generalization and training efficiency.
• Representation Collapse: Static methods can lead to weak label alignment or "collapse," where distinct samples map to similar quantum states, reducing the information richness of the Hilbert space.

The paper posits that improving the initial state preparation (encoding) stage is critical for QML success, particularly when using hybrid quantum-classical models with limited capacity.

2. Methodology: SPATE

The authors propose SPATE (Spiking-Phase Adaptive Temporal Encoding), a novel encoding mechanism that converts real-valued tabular features into quantum states using Leaky Integrate-and-Fire (LIF) spiking dynamics.

A. Core Mechanism

SPATE transforms input data into a quantum state through three main stages:

1. Input Normalization: Features are standardized and scaled to $[0, 1]$.
2. LIF Spike Generation: Each normalized feature $x_i$ drives a discrete-time LIF neuron simulation over a time horizon $T$.
   • The membrane potential $v_i(t)$ evolves based on the input, a time constant $\tau$, and noise.
   • When $v_i$ exceeds a threshold $v_{th}$, a "spike" is recorded, and the potential resets.
   • This generates a spike train $S_i$ for each feature.
3. Parameter Extraction: From the spike trains, SPATE extracts three distinct parameters for each feature:
   • Spike Rate ($\alpha_i$): The total count of spikes, mapped to a rotation angle (amplitude-like).
   • Spike Timing Phase ($\phi_i$): The circular mean of spike times, mapped to a phase rotation.
   • Temporal Bins ($B$): A histogram of spikes across time bins, centered to remove global offsets.

B. Quantum State Preparation

The encoding uses a hybrid register of Feature Qubits ($d$) and Time Qubits ($n_t$):

• Feature Qubits: Receive single-qubit rotations $R_X(\alpha_i)$ and $R_Z(\phi_i)$ to encode rate and timing.
• Time Qubits: Initialized in the $|+\rangle$ state.
• Coupling: Controlled phase rotations ($CRZ$) are applied between feature qubits and time qubits. The rotation angle is determined by the temporal bin weights ($\beta_{i,k}$) scaled by a factor $\lambda$. This explicitly injects coarse temporal structure into the entangled state.

C. Evaluation Protocol

To assess encoding quality independent of the classifier, the authors introduce a rigorous encoding-centric evaluation protocol using stratified cross-validation:

• Label Alignment: Centered Kernel-Target Alignment (CKTA).
• Separability: Fisher Ratio, Inter/Intra-class distance ratios, and Silhouette Score.
• Information Richness: Normalized Shannon Entropy ($H_{norm}$) and Pairwise Total Variation ($TV_{pair}$) to detect representation collapse.
• Downstream Performance: Training a fixed-architecture Hybrid QNN (StronglyEntanglingLayers) to measure Accuracy, Precision, Recall, and AUC.

3. Key Contributions

1. SPATE Architecture: Introduction of a spike-driven encoding method that integrates rate, timing phase, and temporal binning into a shallow quantum circuit using a compact register of time qubits.
2. Encoder-Level Diagnostics: A comprehensive evaluation framework that isolates the quality of the quantum embedding from the learning capability of the QNN, utilizing metrics like CKTA and Fisher scores.
3. Empirical Validation: Demonstration that SPATE produces superior embedding geometries compared to standard Angle and Amplitude encodings, leading to significant performance gains in hybrid QNNs on specific datasets.
4. Boundary Analysis: Identification of regimes where SPATE excels (temporal/clustered data) versus where static encodings remain superior (periodic/ring-shaped data like "Circles").

4. Results

The paper evaluates SPATE against Angle and Amplitude encodings across seven datasets (Iris, Wine, Cancer, Digits, Moons, Circles, Blobs).

A. Representation Quality (Embedding Geometry)

• Moons Dataset: SPATE dramatically improved the embedding. CKTA rose from ~0.015 (Angle/Amplitude) to 0.506, and the Silhouette score shifted from negative values to 0.355, indicating a transition from unstructured noise to clear cluster separation.
• Blobs Dataset: SPATE achieved near-perfect separability with a CKTA of 0.966 and a Fisher score of 7.37, significantly outperforming Angle (CKTA 0.632, Fisher 0.70).
• Information Richness: SPATE consistently showed higher normalized entropy ($H_{norm}$), indicating richer state distributions per qubit and less representation collapse.

B. Downstream Hybrid-QNN Performance

Improved embeddings translated directly into better classification metrics:

• Wine: SPATE achieved 82.6% accuracy and 0.978 AUC, surpassing Amplitude (68.5% acc) and Angle (39.9% acc).
• Moons: SPATE reached 84.0% accuracy and 0.923 AUC, compared to ~63% and ~72% for baselines.
• Cancer: SPATE improved accuracy to 83.7% with high recall (98.3%), whereas baselines hovered around 60-62%.
• Exceptions:
  • Circles: Angle encoding remained superior (79.7% vs. 64.7% for SPATE), confirming that simple rotation-based mappings are better suited for ring-shaped decision boundaries.
  • Digits: In a high-class (10-class) capacity-limited setting, accuracy gains were marginal, though SPATE achieved the highest AUC (0.799), suggesting better ranking behavior.

5. Significance and Conclusion

• Practical Interface: SPATE provides a practical "spike-to-phase" interface, allowing QML models to leverage temporal and spike-based data representations without requiring full Quantum Spiking Neural Networks (QSNNs) or complex pre-training.
• Resource Efficiency: It demonstrates that under strict qubit budgets, optimizing the encoding stage is as critical as optimizing the trainable circuit. Better initial geometry reduces the burden on the QNN to learn complex feature transformations.
• Modularity: SPATE is a general-purpose encoding prefix compatible with various QML pipelines (variational circuits, kernel methods) and can operate in "embedding mode" for classical downstream tasks.
• Limitations & Future Work: The method relies on hyperparameters (LIF time constants, thresholds) that require tuning. Future work will focus on systematic parameter selection and robustness to hardware noise.

In summary, SPATE bridges neuromorphic computing and quantum machine learning, proving that spike-based temporal encoding creates more informative and separable quantum representations, thereby enhancing the performance of hybrid quantum classifiers on constrained hardware.