Low resource entanglement classification from neural network interpretability

Entanglement is a central resource in quantum information and quantum technologies, yet its characterization remains challenging due to both theoretical complexity and measurement requirements. Machine learning has emerged as a promising alternative, enabling entanglement characterization from incomplete measurement data, however model interpretability remains a challenge. In this work, we introduce a unified and interpretable framework for SLOCC entanglement classification of two- and three-qubit states, encompassing both pure and mixed states. We train dense and convolutional neural networks on Pauli-measurement outcomes, provide design guidelines for each architecture, and systematically compare their performance across types of states. To interpret the models, we compute Shapley values to quantify the contribution of each measurement, analyze measurement-importance patterns across different systems, and use these insights to guide a measurement-reduction scheme. Accuracy-versus-measurement curves and comparisons with analytical entanglement criteria demonstrate the minimal resources required for reliable classification and highlight both the capabilities and limitations of Shapley-based interpretability when using machine learning models for entanglement detection and classification.

💡 Research Summary

Entanglement is a cornerstone resource for quantum computation, communication, and metrology, yet its reliable detection and classification remain difficult because full state tomography scales exponentially with the number of qubits. Recent works have shown that machine‑learning models, especially neural networks, can infer entanglement properties from incomplete measurement data, but the black‑box nature of these models limits physical insight and trust. In this paper the authors present a unified, interpretable framework for classifying stochastic‑local‑operations‑and‑classical‑communication (SLOCC) entanglement classes of both two‑ and three‑qubit systems, covering pure and mixed states, using only Pauli‑tomography outcomes as input features.

The dataset is generated by sampling random local invertible operators to produce a uniform spread of states across all SLOCC classes. Pure states are obtained by applying random local unitaries to canonical representatives (e.g., Bell, GHZ, W), while mixed states are constructed as convex combinations of pure‑state samples. Each quantum state is encoded as a vector of 4ⁿ‑1 Pauli expectation values (the full correlation tensor without the identity component), providing a compact yet information‑rich representation.

Two neural‑network architectures are trained and compared. A fully‑connected (FC) multilayer perceptron treats the Pauli vector as a flat feature set, guaranteeing permutation invariance with respect to feature ordering. A convolutional neural network (CNN) reshapes the same vector into a 2‑D grid that reflects the tensor product structure of the Pauli operators and applies small convolutional kernels to capture local correlation patterns. Hyper‑parameters (layer widths, learning rates, regularisation) are optimised via cross‑validation on an 80/10/10 train/validation/test split.

Performance results show that CNNs excel on pure‑state classification, especially for three‑qubit GHZ/W discrimination, achieving >98 % accuracy, whereas FC networks are more robust on mixed‑state data, delivering >95 % accuracy with stable feature importance. To open the black box, the authors compute Shapley values using the model‑agnostic SHAP library (KernelSHAP). For each class, the Shapley analysis yields a ranking of individual Pauli observables according to their contribution to the network’s output. Notably, bipartite entangled states rely heavily on the three two‑qubit correlators ⟨σ_x⊗σ_x⟩, ⟨σ_y⊗σ_y⟩, ⟨σ_z⊗σ_z⟩, while separable states are driven by single‑qubit expectations. In the three‑qubit case, GHZ classification depends on global correlators such as ⟨σ_x⊗σ_x⊗σ_x⟩, whereas W classification emphasizes a different set of three‑body terms.

Guided by the Shapley rankings, the authors perform a systematic measurement‑reduction study: they iteratively remove the least important Pauli observables and re‑evaluate classification accuracy, producing accuracy‑versus‑measurement curves. For pure states, the Shapley‑selected minimal set does not coincide with the mathematically minimal sufficient statistics, indicating redundancy in the learned representations. Conversely, for mixed states the Shapley‑identified subsets match the true minimal measurement sets, and classification performance remains high even after discarding up to 60 % of the original measurements. Quantitatively, two‑qubit mixed‑state classification retains >90 % accuracy with only six of the fifteen possible Pauli correlators, while three‑qubit mixed‑state classification needs only twelve of sixty‑three correlators for comparable performance.

The machine‑learning approach is benchmarked against analytical entanglement criteria. For bipartite systems the Positive Partial Transpose (PPT) test is the gold standard; the neural networks achieve comparable detection rates while requiring far fewer measurements. For three‑qubit systems the authors compare against GHZ/W entanglement witnesses; the learned classifiers outperform the witnesses in the low‑measurement regime, demonstrating higher sensitivity and specificity.

A key contribution of the work is the explicit link between network architecture and interpretability reliability. CNNs, while powerful for capturing complex correlations, produce Shapley values that depend on the arbitrary ordering of input features, reducing their physical interpretability. FC networks, by contrast, yield permutation‑invariant Shapley values, making the importance rankings more trustworthy, especially for mixed‑state data where feature redundancy is lower.

In summary, the paper delivers (i) a comprehensive protocol for SLOCC entanglement classification using only partial Pauli data, (ii) a rigorous Shapley‑based interpretability analysis that informs measurement‑reduction strategies, and (iii) a nuanced understanding of how architectural choices affect both predictive performance and the physical meaning of feature importance. The results suggest that, for near‑term quantum experiments, carefully designed neural‑network classifiers combined with Shapley‑driven measurement selection can substantially lower experimental overhead while preserving reliable entanglement detection. Future directions include scaling the methodology to larger qubit numbers, integrating shadow‑tomography techniques, and testing on real experimental datasets.