Which data samples matter when training a quantum model?

Problem Statement

While the expressivity, loss landscapes, and generalization of Quantum Machine Learning (QML) models have been extensively studied from a model-centric perspective, there is a lack of understanding regarding the compatibility between specific quantum models and particular datasets. Existing research often characterizes model capacity and trainability but does not directly address how individual training samples contribute to the learning process or how to efficiently select data subsets that maximize model performance. The authors posit that a data-centric approach is necessary to determine which data points are most critical for a model's learning dynamics and to improve the data efficiency of QML training pipelines.

Methodology

The authors extend the V-usable information framework, originally introduced by Xu et al. for classical models, to the domain of parametrized quantum models. This framework quantifies the information a specific model family can extract from a dataset relative to its computational and architectural constraints.

1. Theoretical Framework:

• V-Usable Information ($I_V$): Defined as the reduction in uncertainty about labels $Y$ when data $X$ is provided, achievable only by a model family $V$. It is calculated as $I_V(X \to Y) = H_V(Y) - H_V(Y|X)$, where $H_V(Y)$ is the predictive entropy with a null input (no data) and $H_V(Y|X)$ is the conditional entropy given data.
• Pointwise V-Information (PVI): An extension that assigns an information value to individual data points $(x, y)$. PVI can be negative, indicating that a specific sample might hinder learning (e.g., due to noise or mislabeling) or that the model predicts it better without it.
• V-Generalization Proxy ($\Delta_V$): A metric defined as the absolute difference between V-usable information on test data versus training data, serving as an information-theoretic measure of generalization performance.

2. Quantum Adaptation:

• Null Input Construction: A critical challenge in the quantum setting is defining a "null input" ($\emptyset$) that provides no task-specific information while preserving the model's inductive bias. The authors evaluated three candidates: a Haar-random state approximation (N1), a Hadamard product state (N3), and a zero state (N2). They determined that the Hadamard product state ($H^{\otimes S}|0\rangle^{\otimes S}$, labeled N3) offers a practical, high-entropy baseline comparable to Haar-random states without requiring ancilla qubits or density matrix simulations. The zero state (N2), which applies no gates ($U_{\emptyset} = I$), performed substantially worse.
• Estimation: The V-usable information is approximated empirically by training an ensemble of $K$ independent quantum model instantiations with random parameter seeds.

3. Experimental Setup:

• Datasets:
  • Phase-Classification: Ground states of generalized cluster Hamiltonians (16 qubits) distinguishing symmetry-protected topological, ferromagnetic, antiferromagnetic, and trivial phases.
  • Adhoc: A synthetic 2D classification task with controlled ambiguity, redundancy, and noise.
  • NTangled: A quantum-specific task on 8 qubits distinguishing states by Concentrated Entanglement values.
• Models: Parametrized quantum circuits (PQCs) with feature-map encodings and trainable ansatz layers (e.g., 10 layers of rotations and CNOTs).
• Scale: Experiments were conducted using state-vector simulations for smaller systems and tensor-network methods to evaluate a 50-qubit model, demonstrating tractability beyond standard state-vector simulation limits.

Key Contributions

1. Formulation of V-Usable Information for QML: The authors provide an architecture-independent formulation of V-usable information applicable to feature-map-ansatz circuits, data re-uploading models, and integrated parametrized encodings.
2. Quantum Null Input Strategy: They address the non-trivial construction of a null input for quantum models, validating the Hadamard product state (N3) as an effective, low-overhead baseline.
3. Sample-Level Diagnostics: By utilizing PVI, the work offers a fine-grained view of learning dynamics, identifying which samples are memorized, which support generalization, and how sample utility evolves during training.
4. Data-Subset Selection Strategy: The authors introduce a preselection strategy that identifies informative data subsets without requiring a trained model. Instead of discarding low-PVI samples (which requires prior training), they propose selecting samples with low PVI at initialization. The hypothesis is that these samples are not well-represented by the model's initial inductive bias and thus provide a stronger, more diverse learning signal.
5. Scalability Demonstration: The pipeline is demonstrated to be tractable for 50-qubit models using tensor-network methods, bridging the gap between theoretical metrics and large-scale quantum simulation.

Results

• Learning Dynamics: In the phase-classification task, PVI maps revealed that the model struggled to learn information near phase boundaries and in Phase IV, reflected by low PVI values. The V-usable information metric tracked training progress more continuously than accuracy, which plateaued and fluctuated.
• Generalization: The V-generalization proxy ($\Delta_V$) showed strong qualitative agreement with the difference in training and test loss, validating it as a meaningful measure of generalization.
• Data Efficiency:
  • On the phase-classification, adhoc, and NTangled datasets, the PVI-guided subset selection strategy outperformed uniform random sampling, entropy sampling, k-center sampling, and TDA-inspired sampling, particularly when selecting subsets of 10% or larger.
  • The PVI-selected subsets yielded larger initial gradient norms and higher effective dimensions, suggesting improved expressivity.
  • 50-Qubit Proof-of-Concept: Training a 50-qubit model on only 20% of the data (selected via PVI) achieved a test accuracy of 0.716, compared to 0.742 for the full dataset. The difference (2.6%) was within the standard deviation of the subset runs, demonstrating that the selected subset retained most of the full dataset's performance while reducing training samples by 80%.

Significance and Claims

The paper claims to establish V-usable information as a powerful diagnostic tool for learning dynamics and a practical route to data-efficient quantum machine learning.

• Data-Centric Insight: It shifts the focus from purely model-centric properties to the compatibility between models and specific datasets, revealing structures that standard metrics (like accuracy) or generalization bounds may miss.
• Resource Efficiency: The proposed PVI-based selection strategy allows for the identification of critical training samples without the computational cost of training on the full dataset first, directly addressing the resource intensity of QML training.
• Theoretical Extension: The work complements recent progress on model-dependent generalization theory by providing pointwise, data-centric diagnostics.
• Limitations: The authors modestly note that the metric is not normalized (requiring care when comparing across tasks with different label distributions) and that the choice of null input remains a modeling assumption. They also highlight that the framework's robustness to quantum hardware noise requires further evaluation.

In conclusion, the authors argue that empirical, data-dependent studies using V-usable information are essential for assessing QML models, as theoretical guarantees do not always predict the practical behavior of models trained on finite datasets.