Uncertainty-aware quantum machine learning based on Dempster-Shafer theory

Marie Kempkes, Jakob Spiegelberg, Evert van Nieuwenburg, Vedran DunjkoarXiv 2026 · arXiv

In the race to prove "quantum advantage," the field of Quantum Machine Learning (QML) has largely obsessed over expressivity and raw accuracy. But in real-world, high-stakes applications—like autonomous driving or medical diagnostics—a model that is 99% accurate but doesn't know when it's guessing is a liability.

A recent paper by Marie Kempkes and colleagues introduces a framework to fix this. By integrating Dempster-Shafer Theory (DST) into Parameterized Quantum Circuits (PQCs), they’ve developed a way for quantum models to quantify their own uncertainty in a principled, efficient manner.

TL;DR

This work introduces a quantum classification framework that uses Dempster-Shafer Theory to assign "masses" (weights) to sets of labels, allowing models to express uncertainty when data is ambiguous. By training PQCs to output these masses directly, the authors enable efficient out-of-distribution detection and multi-model evidence fusion with polynomial resource requirements.

What Problem Does This Paper Address?

Standard machine learning models typically output a single probability distribution over classes. However, this fails to distinguish between two types of uncertainty:

  1. Aleatoric uncertainty: Inherent randomness in the data (e.g., a blurry image).
  2. Epistemic uncertainty: A lack of knowledge because the model hasn't seen similar data before.

Dempster-Shafer Theory (DST) addresses this by assigning "mass" not just to individual labels, but to sets of labels. For example, a model might say, "I'm 80% sure this is either a '3' or an '8', but I don't know which." Classically, DST is computationally expensive because the number of subsets grows exponentially ($2^n$) with the number of classes. This paper leverages the structure of quantum states to handle this complexity.

Methods

The authors propose a classification model where an $n$-qubit PQC represents $n$ classes.

  • Quantum Encoding: The probability of measuring the circuit in a specific basis state $|i\rangle$ is mapped to a DST "mass" $m(S_i)$. Because there are $2^n$ basis states, the quantum state naturally encodes the entire power set of label assignments.
  • Efficient Training: Instead of trying to optimize all $2^n$ masses (which would be impossible), they use a loss function that only requires singleton plausibilities—the total mass assigned to any set containing a specific label. Crucially, these can be estimated using simple local measurements on individual qubits, making the training process efficient.
  • The "Caution" Parameter ($\lambda$): The loss function includes a trade-off parameter $\lambda$. A high $\lambda$ forces the model to be decisive (assigning mass to single labels), while a lower $\lambda$ allows the model to be "cautious" by shifting mass to larger subsets when it is uncertain.
Figure 1. Conceptual overview of the proposed method. A parame- terized quantum circuit is trained using the loss in Eq. (7). The result- ing basis state probabilities encode masses from Dempster–Shafer theory (see Section III A), which are subsequently used in different post-processing steps (Section III B).

Datasets

The framework was evaluated using standard computer vision benchmarks:

  • MNIST: Handwritten digits (in-distribution).
  • FashionMNIST: Clothing items (used as out-of-distribution data).
  • CIFAR-10: More complex natural images.

Key Insights and Ideas

  • Quantum-DST Synergy: DST aligns remarkably well with quantum mechanics. Just as a quantum state can be a superposition of basis states, a DST mass function is a distribution over the power set of outcomes.
  • Sparsity is the Savior: The authors prove that in an ideal "zero-loss" scenario, the learned mass function is 1-sparse (concentrated on a single set). Numerical experiments show that even under realistic training conditions, the mass function remains relatively sparse, which is vital for practical implementation.
  • Evidence Fusion: One of the most powerful features of DST is the ability to fuse evidence from multiple independent sources. The authors demonstrate that combining several "weak" quantum classifiers (each seeing only a patch of an image) using DST fusion rules leads to significantly better performance than any single model.
Figure 2. Illustration of the influence of the loss function parameter λ. (a) Exemplary mass vector for a representative MNIST image of the digit 3 for different values of λ. Smaller values of λ increase the mass assigned to larger subsets, resulting in more imprecise hypotheses. (b) Average mass assigned to subsets of different cardinalities |S| as a function of λ, quantitatively confirming the trend observed in panel (a).

Key Results

  • Out-of-Distribution (OOD) Detection: Using the "summed singleton plausibility" metric, the model achieved an AUROC of 0.83 for distinguishing between MNIST (seen) and FashionMNIST (unseen) data. This indicates the model successfully "knows what it doesn't know."
  • Mass Fusion Gains: In a multi-view setup, a fused classifier achieved higher accuracy than the mean of individual classifiers, proving that DST rules effectively aggregate complementary evidence.
  • Sparsity Scaling: While the number of subsets grows, the number of subsets required to capture the majority of the mass ($k_\tau$) grows slowly if the model is trained to high precision, as shown in Figure 3.
Figure 3. Empirical sparsity of the learned mass vector beyond the ideal setting of Proposition 2. The quantity kτ denotes the minimum number of subsets required to account for a fraction τ of the total mass. Panels (a) and (c) show kτ as a function of the error threshold ε for different numbers of qubits, while panels (b) and (d) show the dependence on the number of qubits for fixed values of ε. As the training error increases, the mass functions become less sparse. Moreover, for fixed error thresholds, kτ grows rapidly with the number of qubits.

Limitations and Future Work

  • Simulated vs. Real Hardware: The experiments were conducted on simulated quantum systems. Real-world quantum noise might affect the precision of mass assignments.
  • Identifiability: Since the loss function only constrains singleton plausibilities, the full mass function is not uniquely determined. The "inductive bias" of the specific PQC architecture plays a significant role in how uncertainty is distributed.
  • Classical Comparison: While the quantum approach is elegant, it remains to be seen if it offers a definitive "advantage" over optimized classical evidential deep learning models in terms of raw scaling for very large label sets.

By moving beyond simple accuracy and embracing the nuances of uncertainty, this research provides a roadmap for building more reliable and "honest" quantum AI.

Summary updated

Abstract

A large body of work in quantum machine learning has centered on characterizing the expressivity and performance of quantum models. In contrast, little attention has been paid to the reliability of predictions beyond average performance measures such as accuracy. In this work, we introduce an uncertainty-aware quantum classification framework based on Dempster-Shafer theory. This theory allows to express uncertainty through weights assigned not only to individual classification labels but to sets of labels. We introduce a training strategy that enables parameterized quantum circuits to learn such assignments directly from data. This provides an efficient way to prepare the corresponding quantum states and enables the application of different post-processing strategies. In particular, it offers a practical setting for quantum Dempster-Shafer fusion procedures with polynomial resource requirements, even when the corresponding mass functions are defined over exponentially many subsets. Overall, our work represents a step toward enabling quantum models to express uncertainty about their predictions, thereby extending the evaluation of quantum models beyond predictive performance alone.
BibTeX
@misc{kempkes2026uncertainty,
  author = {Marie Kempkes and Jakob Spiegelberg and Evert van Nieuwenburg and Vedran Dunjko},
  title = {{Uncertainty-aware quantum machine learning based on Dempster-Shafer theory}},
  year = {2026},
  eprint = {2611.16903},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.16903},
}
Figure