Quantum machine learning holds immeasurable promises, but current models typically require numerous measurements to produce reliable predictions, creating significant cost and time barriers for widespread use. Along with Brookhaven National Laboratory and Samuel Yen Chen of Gabriel Matos, Chen Yu Liu, Leonardo Prasidi and Kuan Chen Chen, who work at Imperial College London, present a new approach called “Yomo” that dramatically reduces this measurement. Yomo achieves accurate inferences up to a single measurement, and even far fewer measurements, by replacing traditional output methods with stochastic aggregation techniques and using loss functions that facilitate clear predictions. This innovative design overcomes the limitations inherent in existing quantum machine learning models, consistently outperforms them in image recognition tasks, paving the way for more affordable and accessible quantum computation.
Traditionally, quantum machine learning algorithms rely on repeated measurements or shots of observability to obtain reliable predictions. This large reliance on shot budgets leads to high inference costs and time overhead, especially since access to quantum hardware is usually priced proportionally with the number of shots. Yomo replaces Pauli's predicted and probabilities aggregation mechanism and introduces a loss function that facilitates sharp predictions. The experiments show that mugwort always outweighs existing QML models in various shot budgets and simulated noise conditions. The team rigorously tested models for the MNIST and CIFAR-10 datasets, achieving high classification accuracy even in single shot regimes. Specifically, this study confirms Yomo's ability to surpass traditional expectations QML models in shot efficiency. This is formally proven through the theoretical boundaries of the measured shots needed to achieve the target error probability.
This achievement directly addresses key barriers to practical quantum machine learning adoption by significantly reducing financial and computational costs. The team validated the performance of the model under realistic conditions and simulated noise derived from current single kits and 2 quit error rates for existing quantum hardware. Yomo uses a loss function designed to replace traditional expected output with a probability aggregation mechanism and facilitate accurate prediction. Experiments on image datasets including MNIST and CIFAR-10 show that Yomo is consistently superior to existing methods across a variety of measurement budgets under conditions simulated using hardware noise. In team analysis, Yomo avoids the limitations of models that rely on multiple measurements, allowing accurate inferences even for just a single measurement.
The authors acknowledge that the yomo is best suited for workflows in which model training is performed using classical simulations of quantum states, and that deployment takes place on Quantum devices. They also note that future research is needed to identify specific digit Qubits that quantum inference with Yomo surpasses classical simulations of runtimes and may provide valuable guidance for practical applications. The team suggests that this crossover point is a key factor in determining when quantum reasoning will have an advantage.
