Towards Sample Efficient Entanglement Classification for 3 and 4 Qubit Systems: A Tailored CNN-BiLSTM Approach

Technical Summary: Sample-Efficient Entanglement Classification via Tailored CNN-BiLSTM

Problem Statement
Accurate classification of multipartite entanglement in high-dimensional quantum systems is essential for quantum communication and information processing. However, conventional characterization methods (e.g., entanglement witnesses, positive partial transpose criteria) and existing machine learning (ML) approaches face a severe scaling challenge. Traditional methods require resources that grow exponentially with system size, while standard ML classifiers typically demand massive training datasets. This creates a significant experimental bottleneck, as acquiring sufficient quantum data is resource-intensive and time-consuming. The core challenge addressed is achieving reliable entanglement classification with minimal training data (sample efficiency) to alleviate the burden of data acquisition in quantum experiments.

Methodology
The authors propose a hybrid neural network architecture integrating Convolutional Neural Networks (CNNs) and Bidirectional Long Short-Term Memory networks (BiLSTMs). The model is designed to extract local, spatially invariant features via CNNs and capture complex sequential dependencies via BiLSTMs. The study investigates two specific fusion paradigms for connecting these modules:

1. Architecture 1 (Archi1 - Flattening-based): Measurement outcomes are organized into a vector, processed by 1D convolutional layers, and then flattened into a single vector before being fed into the BiLSTM. This approach is computationally efficient but discards the sequential correlations between different measurement bases.
2. Architecture 2 (Archi2 - Dimensionality-Transforming): This is the novel contribution. Instead of flattening, the feature maps output by the CNNs undergo a dimensionality transformation. The spatial dimension of the feature maps is reshaped into a temporal sequence dimension, while the feature channel dimension becomes the feature value dimension. This preserves the correlations between measurement outcomes, allowing the BiLSTM to process the data as a sequence in both forward and backward directions.

The input consists of simulated projective measurement outcomes from mutually unbiased bases. The models are trained using the Adam optimizer with cross-entropy loss. The study focuses on 3-qubit and 4-qubit systems, classifying states into Stochastic Local Operations and Classical Communication (SLOCC) equivalence families (e.g., separable, GHZ, W, and various bipartite entangled states).

Key Results
The proposed models were evaluated on datasets ranging from 100 samples (extreme scarcity) to 400,000 samples (full data).

• Performance under Data Scarcity: When trained on only 100 samples, Architecture 2 maintained classification accuracies exceeding 90% for both 3-qubit and 4-qubit systems. In contrast, standalone CNNs, BiLSTMs, MLPs, and Architecture 1 showed significantly lower performance in this low-data regime.
• Performance with Full Data: Under full-data conditions (400,000 samples), both architectures achieved near-perfect accuracy, with Archi2 reaching 99.976% (3-qubit) and 99.999% (4-qubit).
• Convergence: Archi2 demonstrated rapid loss convergence, stabilizing within approximately 100 epochs even with 100 training samples.
• Noise Robustness: In the presence of dephasing noise and random measurement noise, Archi2 consistently outperformed Archi1. With only 100 training samples, Archi2 maintained accuracy above 88% under noise, whereas Archi1 dropped below 80%. This suggests Archi2 possesses inherent denoising capabilities by exploiting temporal correlations.
• Computational Trade-off: Archi2 requires significantly longer training times than Archi1 and baseline models (e.g., ~25 hours vs. ~2.5 hours for 4-qubit full-data training) due to the Backpropagation Through Time (BPTT) algorithm and increased sequence length. However, the authors argue this computational cost is justified given the massive reduction in experimental data acquisition requirements.

Key Contributions

1. Sample Efficiency: The work demonstrates that tailored CNN-BiLSTM fusion can reduce the required training data by four orders of magnitude (from 400,000 to 100 samples) while maintaining >90% accuracy.
2. Dimensionality Transformation: The introduction of a specific reshaping strategy (Archi2) that converts CNN feature maps into sequential data for BiLSTMs, preserving physically meaningful correlations between measurement bases that are lost in standard flattening approaches.
3. Physics-Aware Design: The architecture is grounded in the physics of quantum measurement, acknowledging that outcomes from complementary bases are interdependent. By structuring the data to reflect these dependencies, the model achieves higher information density and better generalization.
4. Robustness: The model exhibits superior resilience to experimental noise without requiring noise-aware training or additional regularization.

Significance and Claims
The paper claims to offer a practical pathway toward scalable entanglement verification in complex quantum systems. By shifting the complexity from the experimental domain (data acquisition) to the computational domain (training time), the approach addresses a critical bottleneck in quantum information science. The authors assert that their method enables reliable classification even under the stringent data constraints of current Noisy Intermediate-Scale Quantum (NISQ) devices. While the study is demonstrated on pure states of 3- and 4-qubit systems, the authors suggest the architecture is inherently scalable and adaptable to larger systems, mixed states, and other quantum learning tasks through future modifications like attention mechanisms or graph neural networks.