Sarayut-HQCNN: A Hybrid Quantum–Classical Neural Network with Semantic Stability Operator for Robust Classification Under Noise

Abstract Hybrid quantum–classical neural networks have emerged as a promising computational paradigm for learning complex data representations under the constraints of noisy intermediate-scale quantum devices. However, practical deployment remains limited by instability, noise sensitivity, uncertain decision boundaries, and semantic drift between input perturbations and model outputs. This paper proposes Sarayut-HQCNN, a Hybrid Quantum–Classical Neural Network equipped with a Semantic Stability Operator (SSO) for robust classification under noisy conditions. The proposed architecture combines classical feature extraction, variational quantum encoding, hybrid decision layers, and an explicit semantic stability gate. The SSO estimates output variance under controlled input perturbations and rejects unstable predictions before final classification. A triple-guard decision mechanism based on confidence, class margin, and semantic sigma is introduced to improve decision reliability. Experimental protocols are designed for biomedical signal classification and financial time-series classification under multiple noise regimes. The proposed framework aims to reduce false confidence, improve robustness, and provide a practical rejection mechanism for safety-sensitive machine learning applications. Index Terms—Hybrid quantum–classical neural network, quantum machine learning, semantic stability operator, noisy classification, robust artificial intelligence, variational quantum circuit, uncertainty rejection. I. Introduction Quantum machine learning has attracted significant attention as a possible extension of classical learning systems, particularly in problems involving high-dimensional data, non-linear feature spaces, and complex probabilistic structures. Variational quantum circuits provide a trainable interface between classical data and quantum state representations, enabling hybrid architectures that combine classical optimization with quantum transformations. Despite this promise, current quantum machine learning systems face several limitations. First, noisy intermediate-scale quantum devices are sensitive to decoherence, gate noise, measurement uncertainty, and limited circuit depth. Second, hybrid neural systems may produce confident predictions even when the internal representation is unstable. Third, many models lack an explicit mechanism to detect whether a prediction remains semantically consistent under small perturbations of the input. To address these issues, this paper introduces Sarayut-HQCNN, a hybrid quantum–classical neural architecture with a built-in Semantic Stability Operator (SSO). The central idea is simple but important: a prediction should not be accepted only because its confidence score is high; it should also be stable under small, controlled perturbations. If a model’s output changes significantly when the input is slightly disturbed, the decision is considered semantically unstable and should be rejected or passed to a mindful review mode. The main contributions of this paper are as follows: A hybrid quantum–classical neural architecture combining classical feature extraction and variational quantum processing. A Semantic Stability Operator that estimates local semantic drift under perturbation. A triple-guard decision rule using confidence, margin, and semantic sigma. A reproducible experimental protocol for testing robustness under noisy biomedical and financial datasets. A rejection-aware classification framework suitable for safety-sensitive applications. II. Related Work Hybrid quantum–classical neural networks are typically constructed by combining classical preprocessing layers with parameterized quantum circuits. Earlier works on quantum machine learning demonstrated that quantum circuits can be used as feature maps, classifiers, and trainable variational models. Variational quantum algorithms have also been studied as practical tools for near-term quantum devices. However, several challenges remain. Barren plateaus may reduce trainability in deep parameterized quantum circuits. Noise in quantum gates and measurement processes can degrade model reliability. In classical machine learning, confidence calibration and adversarial robustness have been studied extensively, but many approaches still focus on output probability rather than semantic stability. The proposed Sarayut-HQCNN differs from standard hybrid models by adding an explicit semantic stability layer after the hybrid inference stage. Instead of accepting the highest-probability class directly, the model evaluates whether the prediction is stable under perturbation. This introduces a rejection-aware mechanism that separates confident and stable predictions from confident but unstable predictions. III. Proposed Architecture The proposed Sarayut-HQCNN architecture consists of five major components: Input Layer Classical Feature Extractor Quantum Encoding Layer Hybrid Decision Layer Semantic Stability Operator Given an input sample , the model first transforms it into a classical feature representation: h = F_{\phi}(x) where denotes the classical feature extractor with trainable parameters . The extracted feature vector is then encoded into a quantum circuit: |\psi(x)\rangle = U(x, \theta)|0\rangle^{\otimes n} where is a parameterized quantum circuit, represents quantum trainable parameters, and is the number of qubits. The quantum circuit output is measured as expectation values: z_q = \left[\langle Z_1 \rangle, \langle Z_2 \rangle, ..., \langle Z_n \rangle \right] The quantum output is then combined with classical features and passed to the final classifier: \hat{y} = \text{softmax}(W[z_q;h] + b) where denotes feature concatenation. IV. Semantic Stability Operator The Semantic Stability Operator (SSO) is designed to estimate the instability of a model prediction under small input perturbations. For each input , a set of perturbed samples is generated: x_k = x + \epsilon_k, \quad k = 1,2,\dots,K where represents bounded perturbation noise. The semantic sigma is defined as: \sigma(x) = \frac{1}{K}\sum_{k=1}^{K} \left\| f(x+\epsilon_k) - f(x) \right\|^2 where is the model output distribution. A low value of indicates that the prediction is stable under perturbation, while a high value indicates semantic drift. The final decision is governed by a triple-guard rule: D(x)= \begin{cases} \text{Accept}, & C(x) \geq \tau_c,\ M(x) \geq \tau_m,\ \sigma(x) \leq \tau_s \\ \text{Reject}, & \text{otherwise} \end{cases} where: is the prediction confidence, is the class margin, is the semantic instability score, , , and are decision thresholds. A typical configuration used in the proposed framework is: \tau_c = 0.55,\quad \tau_m = 0.08,\quad \tau_s = 0.0042 These values should be validated experimentally for each dataset. V. Training Objective The total training loss combines classification loss and semantic stability regularization: \mathcal{L}_{total} = \mathcal{L}_{CE} + \lambda \mathcal{L}_{SSO} where is the cross-entropy loss and penalizes unstable predictions: \mathcal{L}_{SSO} = \frac{1}{N}\sum_{i=1}^{N}\sigma(x_i) The coefficient controls the strength of stability regularization. This encourages the model not only to classify correctly but also to maintain output consistency under small perturbations. VI. Experimental Setup A. Biomedical Signal Classification The first experimental setting evaluates the model on biomedical time-series classification, such as ECG arrhythmia detection. Input signals are segmented into fixed-length windows and normalized before feature extraction. Noise is added at different signal-to-noise ratios to evaluate model robustness. Suggested noise levels: Noise Condition SNR Level Low noise 20 dB Medium noise 10 dB High noise 5 dB Evaluation metrics include: Accuracy F1-score Precision Recall Semantic Sigma Stability Rejection Rate Accepted Prediction Accuracy B. Financial Time-Series Classification The second experimental setting evaluates the model on financial market classification. Market data are transformed into fixed-length windows containing price, volume, and technical indicators. The prediction task is formulated as a three-class classification problem: y \in \{\text{Buy}, \text{Sell}, \text{Hold}\} The SSO module is used to filter unstable trading signals. Only predictions satisfying the triple-guard condition are accepted for decision execution. C. Baseline Models The proposed model should be compared against the following baselines: Model Description CNN Classical convolutional neural network LSTM Classical recurrent model CNN-LSTM Classical hybrid sequence model HQCNN without SSO Hybrid quantum–classical model without stability gate Sarayut-HQCNN with SSO Proposed method VII. Results and Discussion The expected outcome is that Sarayut-HQCNN with SSO may reject a portion of uncertain predictions while improving the reliability of accepted predictions. Unlike conventional models that always return a class label, the proposed framework distinguishes between stable and unstable decisions. Example result table format: Model Accuracy F1-score Sigma ↓ Rejection Rate Accepted Accuracy CNN [insert] [insert] [insert] — — LSTM [insert] [insert] [insert] — — HQCNN [insert] [insert] [insert] — — Sarayut-HQCNN + SSO [insert] [insert] [insert] [insert] [insert] The key advantage of the SSO mechanism is not merely higher raw accuracy, but improved decision trustworthiness. In safety-sensitive environments, an unstable prediction should not be treated the same as a stable one. Therefore, the rejection mechanism provides an operational safeguard against overconfident but fragile classifications. In noisy environments, the model is expected to show three important behaviors: Lower semantic sigma for accepted predictions. Higher accuracy on accepted samples compared with all predictions. Better separation between stable and unstable decision regions. These properties make the proposed architecture suitable for applications where the cost of false confidence is high. VIII. Limitations The proposed framework has several limitations. First, the SSO module introduces additional computational overhead because multiple perturbed samples must be evaluated for each input. Second, threshold values require dataset-specific calibration. Third, quantum circuit simulation may not fully represent the behavior of real quantum hardware. Fourth, the practical advantage of quantum components must be carefully validated against strong classical baselines. Therefore, the current work should be interpreted as a hybrid robustness framework rather than a definitive claim of quantum advantage. IX. Conclusion This paper presented Sarayut-HQCNN, a Hybrid Quantum–Classical Neural Network with a Semantic Stability Operator for robust classification under noise. The proposed SSO mechanism estimates semantic drift by measuring output variation under controlled perturbations. A triple-guard decision rule based on confidence, margin, and semantic sigma enables the model to reject unstable predictions. This design shifts the goal from simple classification accuracy toward stable and trustworthy decision-making. Future work will include hardware-based quantum experiments, adaptive threshold learning, larger benchmark validation, and deployment in real-time biomedical and financial monitoring systems. References ต้องตรวจสอบรูปแบบ/รายละเอียดอีกครั้งก่อนส่งจริงกับ IEEE [1] J. Preskill, “Quantum Computing in the NISQ era and beyond,” Quantum, vol. 2, p. 79, 2018. [2] J. Biamonte et al., “Quantum machine learning,” Nature, vol. 549, pp. 195–202, 2017. [3] M. Schuld and N. Killoran, “Quantum machine learning in feature Hilbert spaces,” Physical Review Letters, vol. 122, no. 4, 2019. [4] V. Havlíček et al., “Supervised learning with quantum-enhanced feature spaces,” Nature, vol. 567, pp. 209–212, 2019. [5] K. Mitarai, M. Negoro, M. Kitagawa, and K. Fujii, “Quantum circuit learning,” Physical Review A, vol. 98, no. 3, 2018. [6] J. R. McClean et al., “Barren plateaus in quantum neural network training landscapes,” Nature Communications, vol. 9, 2018. [7] M. Cerezo et al., “Cost function dependent barren plateaus in shallow parametrized quantum circuits,” Nature Communications, vol. 12, 2021. [8] I. Goodfellow, J. Shlens, and C. Szegedy, “Explaining and harnessing adversarial examples,” in International Conference on Learning Representations, 2015. [9] C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger, “On calibration of modern neural networks,” in International Conference on Machine Learning, 2017. [10] G. B. Moody and R. G. Mark, “The impact of the MIT-BIH Arrhythmia Database,” IEEE Engineering in Medicine and Biology Magazine, vol. 20, no. 3, pp. 45–50, 2001.