Quantum models now express their own uncertainty about predictions, going beyond simply stating what they believe is most likely. A framework based on Dempster-Shafer theory allows circuits to learn assignments reflecting confidence in multiple possible labels simultaneously; it offers efficient preparation of quantum states with polynomial resource requirements even when dealing with complex uncertainties represented by exponentially many possibilities.
This framework uses Dempster-Shafer theory, a method for representing uncertain data, to efficiently prepare quantum states even with complex possibilities. The approach allows combining multiple sources of information and evaluating systems beyond simple accuracy measures using manageable computational resources. The Leiden University and Volkswagen Group Innovation team has tackled a key limitation in quantum machine learning: reliably assessing prediction confidence; current models excel at stating *what* they believe but struggle to express *how sure* they are.
Their new framework integrates Dempster-Shafer theory, a system for representing degrees of belief about different outcomes, much like assigning probabilities while acknowledging incomplete knowledge, into parameterised quantum circuits. These flexible “computer programs” for quantum computers have settings that can be adjusted during training, similar to tuning knobs on an old radio. The approach efficiently prepares quantum states even when dealing with complex uncertainties represented by exponentially many possibilities using ‘mass functions’, which rank certainty levels across multiple guesses.
Polynomial scaling unlocks strong uncertainty handling with quantum Dempster-Shafer evidence theory
Quantum Dempster-Shafer fusion procedures are now achievable using polynomial resources, a sharp improvement over prior methods that struggled beyond relatively simple scenarios. Previously hampered by computational limits when dealing with complex data, this breakthrough allows for the efficient combination of information from multiple classifiers. The framework enables parameterised quantum circuits to learn and express subtle uncertainties about predictions via assignments reflecting confidence in sets of labels rather than single ones, extending evaluation metrics beyond mere predictive performance.
Scaling as a polynomial in problem size, quantum Dempster-Shafer fusion procedures combine information from several sources using weights assigned to sets of possible outcomes instead of individual predictions. This represents an advance on previous methods struggling with more complicated situations and facilitates handling mass functions defined across exponentially many subsets. They successfully tested their approach on benchmark world datasets, showing potential for sophisticated evaluation metrics that extend past simple prediction accuracy; however, these results currently focus on simulation, meaning practical implementation on real-world quantum hardware remains a significant hurdle before widespread application is viable.
Quantifying prediction certainty may affect quantum model performance
This work demonstrates a pathway toward quantum models articulating confidence levels, a vital step for real-world deployment, but validation remains limited to simulations, hindering immediate use on actual hardware. Does incorporating uncertainty awareness introduce any trade-offs regarding overall predictive accuracy? Acknowledging the current reliance on simulations rather than physical quantum computers raises legitimate concerns about whether adding uncertainty estimates impacts how well these models predict outcomes and if knowing *when* a system isn’t confident diminishes its overall accuracy.
The team has demonstrated a quantum system capable of expressing its own uncertainty, extending evaluation beyond mere predictive performance. This advance could enable more reliable quantum machine learning models by integrating Dempster-Shafer theory, a method for representing degrees of belief, into adaptable quantum circuits. They have enabled systems to assign confidence levels not just to individual answers but also to sets of possibilities simultaneously; efficient processing using resources scaling predictably as the problem grows overcomes previous computational limitations when dealing with complex data, even with exponentially increasing options.
This research demonstrates that quantum computing can be used to express prediction uncertainties through a new framework based on Dempster-Shafer theory. By assigning weights to both single labels and combinations of labels, this approach allows quantum models to move beyond simply providing an answer and instead indicate their confidence in it. This offers a way to process information defined across many possible outcomes efficiently, overcoming challenges faced by earlier methods. The work represents progress toward evaluating quantum machine learning models for reliability, not just accuracy.
👉 More information
🗞 Uncertainty-aware quantum machine learning based on Dempster-Shafer theory
✍️ Marie Kempkes, Evert van Nieuwenburg and Vedran Dunjko (Affiliation: Leiden University); Jakob Spiegelberg (Affiliation: Volkswagen Group Innovation)
🧠 ArXiv: https://arxiv.org/abs/2611.16903


See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
