Deep learning models are powerful but opaque. As quantum machine learning matures, the field faces a defining choice: build quantum models that are equally opaque, or exploit the mathematical structure of quantum mechanics to make them inherently interpretable. We show that the latter is possible. By tracking quantum mutual information~(MI), entanglement entropy, and state fidelity through the layers of a Quantum Transformer Block (\qtb{}), a fully-coherent variational circuit with quantum analogues of both attention and feedforward, we gain direct insight into how the model processes information: which tokens it attends to, when correlations form, and why predictions fail. On four tasks with known dependency structure we show that (i)~learned MI matrices align with ground-truth task structure (AUC$\,{=}\,0.69$ on lookup), (ii)~disabling entangling gates collapses accuracy from 100\% to 15\% while MI$\to 0$, proving entanglement is the mechanism, (iii)~accuracy and MI co-evolve during training ($ρ\,{=}\,0.92$ on lookup), and (iv)~per-sample MI predicts prediction correctness on the conditional task with ROC AUC$\,{=}\,0.84$. All results are validated on IBM Quantum hardware (ibm\_kingston, Heron~r2): the circuit's reasoning process, from product state through structured entanglement, is directly observable on a superconducting processor. These proof-of-concept results, obtained on small synthetic tasks, suggest that the physics of quantum computation can provide intrinsic interpretability signals with no direct classical counterpart, motivating study of whether this advantage persists at scale.
AI Summary
The paper asks whether intermediate quantum-state correlations can reveal how a variational quantum sequence model processes information. It studies a fully coherent “Quantum Transformer Block” with inter-register attention gates and intra-register processing, using four three-qubit position registers and two layers. Mutual information (MI), register entropy, participation ratio, and inter-layer fidelity are evaluated on synthetic copy, lookup, conditional-routing, and identity tasks.
Across three training seeds, sequence accuracies are $100\%$, $82.7\%$, $88.6\%$, and $100\%$, respectively. MI identifies task-labelled relevant register pairs with AUC $0.65$ on copy and $0.69$ on lookup, but $0.19$ on conditional routing. Lookup MI and validation accuracy co-evolve with Spearman correlation $0.92\pm0.05$; correlations on other tasks vary substantially across seeds. On an illustrative conditional-task seed, mean MI predicts correctness with ROC AUC $0.84$. In a separate trainable-readout configuration, removing inter-register gates reduces copy accuracy from $100\%$ to $15\%$ and conditional accuracy from $99.8\%$ to $39.8\%$, while MI vanishes.
IBM Heron experiments use 12-qubit circuits with transpiled depths approximately $125$–$229$ and $4096$ shots. Computational-basis Shannon-MI fingerprints correlate with matched shot simulations above $0.88$ on the three routing tasks. These measurements are not quantum MI or direct entanglement certification. Stated limitations include small synthetic tasks, costly exact-state diagnostics, noise-compressed hardware signals, imperfect lookup performance, and readout-dependent MI unrelated to necessary routing on the identity task. A scalable interpretability advantage remains unestablished; gate-level construction details are withheld for intellectual-property reasons.
Central claim may not hold: Disabling all inter-register gates removes communication as well as entanglement generation, so the ablation does not isolate entanglement as the necessary computational resource.
The abstract overstates hardware validation: hardware experiments measure classical outcome correlations and do not reproduce the accuracy-collapse, training-correlation or correctness-prediction results.
The conditional-task MI-quartile accuracies average to 61%, incompatible with the reported three-seed accuracy of 88.6% ± 5.2% unless an unexplained different evaluation was used.
MI is not established as a general attention or routing diagnostic, given below-chance conditional-task alignment and substantial MI on the communication-free repeat task.
The claimed configuration-independent layerwise formation trajectory is unsupported because the compared models form approximately 83% versus 5% of final MI after the first layer.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
Deep learning models are powerful but opaque. As quantum machine learning matures, the field faces a defining choice: build quantum models that are equally opaque, or exploit the mathematical structure of quantum mechanics to make them inherently interpretable. We show that the latter is possible. By tracking quantum mutual information~(MI), entanglement entropy, and state fidelity through the layers of a Quantum Transformer Block (\qtb{}), a fully-coherent variational circuit with quantum analogues of both attention and feedforward, we gain direct insight into how the model processes information: which tokens it attends to, when correlations form, and why predictions fail. On four tasks with known dependency structure we show that (i)~learned MI matrices align with ground-truth task structure (AUC=0.69 on lookup), (ii)~disabling entangling gates collapses accuracy from 100% to 15% while MI→0, proving entanglement is the mechanism, (iii)~accuracy and MI co-evolve during training (ρ=0.92 on lookup), and (iv)per-sample MI predicts prediction correctness on the conditional task with ROC AUC=0.84. All results are validated on IBM Quantum hardware (ibm_kingston, Heronr2): the circuit's reasoning process, from product state through structured entanglement, is directly observable on a superconducting processor. These proof-of-concept results, obtained on small synthetic tasks, suggest that the physics of quantum computation can provide intrinsic interpretability signals with no direct classical counterpart, motivating study of whether this advantage persists at scale.
The paper asks whether intermediate quantum-state correlations can reveal how a variational quantum sequence model processes information. It studies a fully coherent “Quantum Transformer Block” with inter-register attention gates and intra-register processing, using four three-qubit position registers and two layers. Mutual information (MI), register entropy, participation ratio, and inter-layer fidelity are evaluated on synthetic copy, lookup, conditional-routing, and identity tasks.
Across three training seeds, sequence accuracies are 100%, 82.7%, 88.6%, and 100%, respectively. MI identifies task-labelled relevant register pairs with AUC 0.65 on copy and 0.69 on lookup, but 0.19 on conditional routing. Lookup MI and validation accuracy co-evolve with Spearman correlation 0.92±0.05; correlations on other tasks vary substantially across seeds. On an illustrative conditional-task seed, mean MI predicts correctness with ROC AUC 0.84. In a separate trainable-readout configuration, removing inter-register gates reduces copy accuracy from 100% to 15% and conditional accuracy from 99.8% to 39.8%, while MI vanishes.
IBM Heron experiments use 12-qubit circuits with transpiled depths approximately 125–229 and 4096 shots. Computational-basis Shannon-MI fingerprints correlate with matched shot simulations above 0.88 on the three routing tasks. These measurements are not quantum MI or direct entanglement certification. Stated limitations include small synthetic tasks, costly exact-state diagnostics, noise-compressed hardware signals, imperfect lookup performance, and readout-dependent MI unrelated to necessary routing on the identity task. A scalable interpretability advantage remains unestablished; gate-level construction details are withheld for intellectual-property reasons.
Abstract Unsupervised machine learning models build an internal representation of their training data without the need for explicit human guidance or feature engineering. This learned representation provides insights into which features of the data are relevant for the task at hand. In the context of quantum physics, training models to describe quantum states without human intervention offers a promising approach to gaining insight into how machines represent complex quantum states. The ability to interpret the learned representation may offer a new perspective on non-trivial features of quantum systems and their efficient representation. We train a generative model on two-qubit density matrices generated by a parameterized quantum circuit. In a series of computational experiments, we investigate the learned representation of the model and its internal understanding of the data. We observe that the model learns an interpretable representation which relates the quantum states to their underlying entanglement characteristics. In particular, our results demonstrate that the latent representation of the model is directly correlated with the entanglement measure concurrence. The insights from this study represent proof of concept toward interpretable machine learning of quantum states. Our approach offers insight into how machines learn to represent small-scale quantum systems autonomously.
Quantum computing has entered a stage of increasing practicality. Many quantum hardware vendors such as IBM, Rigetti, Honeywell, and IonQ now enable experiments on real devices in the Noisy Intermediate-Scale Quantum (NISQ) era. These platforms show computational advantages in domains such as optimization, machine learning, and materials science. However, they remain limited by hardware noise and the absence of human-interpretable information. Existing visual metaphors, such as the Bloch Sphere for single-qubit states or circuit schematics for algorithm design, struggle to convey multi-qubit entanglement or measurement probabilities in ways accessible to human reasoning. Likewise, the rise of variational quantum circuits and quantum neural networks (QNNs) introduces new layers of opacity: how data are encoded into quantum states and how parameters evolve through training remain largely under-explored. These challenges motivate a need for visual explanation frameworks that unveil the invisible processes of quantum computing visible, interpretable, and reasonable.To address these challenges, this dissertation presents a coherent sequence of visual analytics systems that progressively enhance explainability across the hardware, state, circuit, and model layers of quantum computing. First, VACSEN supports noise-aware execution by visually correlating hardware fidelity metrics with compiled quantum circuits, allowing users to identify unstable qubits, track temporal noise evolution, and select optimal execution strategies. Second, VENUS introduces a geometric representation for quantum state visualization, combining right triangles and semicircles to jointly encode amplitude and probability information, thereby enabling intuitive exploration of superposition and entanglement beyond the traditional Bloch Sphere. Third, QuantumEyes extends interpretability to quantum circuit behavior, integrating coordinated views to reveal how quantum gates transform amplitudes and probabilities during algorithm execution (e.g., Grover’s search or the Quantum Fourier Transform). Finally, VIOLET applies visual explainability to quantum neural networks, bridging concepts from explainable AI and quantum computing. It visualizes the encoding of classical data into quantum states, tracks parameter dynamics during variational training, and exposes learned features and measurement distributions, establishing the first end-to-end system for understanding QNNs.Each system was designed through iterative co-design with quantum computing experts and evaluated through case studies, task analyses, and in-depth interviews involving researchers and practitioners from academia and industry. Across evaluations, participants reported that visualization significantly enhanced their ability to reason about noise effects, understand quantum-state evolution, and interpret the learning behavior of QNNs. These findings demonstrate that visualization not only accelerates analysis and debugging but also strengthens conceptual understanding and user trust in quantum computing. To sum up, this dissertation establishes visualization
Abstract A recurring weakness in quantum machine learning (QML) is that reported ‘quantum advantages’ are seldom tested against a capacity-matched classical control, leaving it unclear whether a gain comes from the quantum substrate or from the architectural change that accompanies it. Our primary contribution is methodological: a protocol for attributing such gains honestly—a capacity-matched classical bottleneck of identical parameter budget, transparent reporting of where quantum does not help, and validation on real quantum hardware—which we develop and apply through a concrete case study. That case study is Quantum Adaptive Self-Attention (QASA), a hybrid Transformer that replaces the value projection of a single encoder layer with a 36-parameter parameterized quantum circuit (PQC), keeping all other layers classical. Across nine synthetic benchmarks and the real-world ETTh1 dataset, QASA improves on a full-capacity classical Transformer for chaotic and trend-dominated signals. To ask whether this is a genuinely quantum effect, we introduce a control rarely applied in QML—a capacity-matched classical bottleneck with the same parameter budget—and find that it matches the PQC on the error metrics. The gain is therefore attributable to the low-rank value-projection bottleneck (an architectural parsimony principle), not to quantumness; adding further quantum layers only degrades performance and trainability. We accordingly position the quantum layer not as a source of accuracy advantage but as a competitive instantiation of this principle: its low-rank compression onto the signal’s intrinsic dimensionality is matched by a classical bottleneck, so the gain is architectural rather than quantum. The quantum layer’s distinguishing features are physical—a high circuit entanglement (Meyer–Wallach Q = 0.981 with only 27 CNOTs, which we report as a circuit-level property) and, most concretely, noisy intermediate-scale quantum deployability, which we verify by executing the trained model on a real IBM Quantum processor (one-step prediction within 7.5 % of noiseless simulation on the quantum-favored task and 25 % , with overlapping error bars, on a classical-favored control; no error mitigation). Relative to other quantum sequence models (quantum long short-term memory (QLSTM), QnnFormer), QASA is moreover the most resource-efficient of the three: it reaches competitive accuracy with the fewest quantum parameters (36 vs 90–128) and the strongest entanglement-per-gate ( Q = 0.981 with 27 CNOTs vs QLSTM’s 56), and retains trainable gradients that deeper circuits lose. We argue that capacity-matched baselines and honest reporting of where quantum does not help are prerequisites for credible quantum-machine-learning claims.
Quantum machine learning offers a compelling representational perspective: angle-encoded states inhabit Hilbert spaces in which periodic similarities and interactions can be expressed naturally. Realizing this perspective for visual recognition remains difficult, however, because present quantum neural networks are constrained by limited qubit counts, costly circuit simulation and measurement, noise, and unstable optimization on noisy intermediate-scale quantum devices. We investigate whether useful structural ideas from quantum models can instead be realized as scalable classical Transformer operations. We introduce QiT, a Quantum-inspired Transformer for vision tasks with three components: (i) angle-inspired encoding that maps image tokens to learned trigonometric Hilbert-space features analogous to quantum rotation-based state encoding; (ii) self-attention over these periodic features, inducing a classical cosine kernel approximated to quantum fidelity kernels; and (iii) gated multiplicative emulation, a trainable classical surrogate for interaction terms found in variational circuits. All components are differentiable tensor operations, so QiT claims neither quantum computation nor quantum speedup and retains the O(N2D) attention complexity of a standard Vision Transformer. Across image-classification benchmarks, QiT is competitive with a matched classical Transformer while avoiding the severe runtime cost observed for a small simulated quantum Transformer. QiT-B reaches 78.3\% ImageNet-1K top-1 accuracy with 45.7M parameters and 11.5 GFLOPs. These results position QiT as a scalable baseline for isolating and evaluating quantum-motivated inductive biases in visual recognition.
Variational quantum circuits (VQCs) are a leading approach to quantum machine learning on near-term devices, yet it remains unclear which circuit architecture yields the best accuracy-parameter trade-off on classical tabular data. We present a systematic empirical comparison of four VQC families -- multi-layer fully-connected (FC-VQC), residual (ResNet-VQC), hybrid quantum-classical transformer (QT), and fully quantum transformer (FQT) -- across five regression and classification benchmarks. Our key findings are: \textbf{(i)}~FC-VQCs achieve 90-96\% of the R2 of attention-based VQCs while using 40-50\% fewer parameters, and consistently outperform equal-capacity MLPs (mean R2=0.829 vs.\ MLP720's 0.753 on Boston Housing, 3-seed average); \textbf{(ii)}~FC-VQC's Type~4 inter-block connectivity provides partial cross-token mixing that approximates the role of attention -- explicit quantum self-attention yields only marginal gains on most datasets while significantly increasing parameter count; \textbf{(iii)}~expressibility saturates at circuit depth~≈3, explaining why shallow VQCs already cover the Hilbert space effectively; \textbf{(iv)}~LayerNorm on the fully quantum transformer improves classification accuracy, suggesting normalization is important when all operations are quantum; \textbf{(v)}~in our noise study on Boston Housing, FQT degrades gracefully under depolarizing noise while QT collapses. All results are validated across three random seeds. These findings provide practical architectural guidance for deploying VQCs on near-term quantum hardware.
Abstract Transformer-based neural-network quantum states (NNQS) have shown great promise in representing quantum many-body ground states, offering high flexibility and accuracy. However, the interpretability of such models remains limited, especially in terms of connecting network components to physically meaningful quantities. We propose that the attention mechanism — a central module in transformer architectures — explicitly models the conditional information flow between orbitals. Intuitively, as the transformer learns to predict orbital configurations by optimizing an energy functional, it approximates the conditional probability distribution p ( x n | x 1 ,…, x n –1 ), implicitly encoding conditional mutual information (CMI) among orbitals. This suggests a natural correspondence between attention maps and CMI structures in quantum systems. To probe this idea, we compare weighted attention scores from trained transformer wavefunction ansatze with CMI matrices across several representative small molecules. In most cases, we observe a positive rank-level correlation (Kendall’s tau) between attention and CMI, suggesting that the learned attention can reflect physically relevant orbital dependencies. This study provides a quantitative link between transformer attention and conditional mutual information in the NNQS setting. Our results provide a step toward explainable deep learning in quantum chemistry, pointing to opportunities in interpreting attention as a proxy for physical correlations.