A quantum processor classified handwritten digits with 96.08% accuracy

On an IBM quantum processor, the system also reached its result about 70% faster than the compared classical model.

Editorial illustration — not from the study.

The study introduces a way to split image data into smaller pieces before encoding it into a quantum system, reducing the demands placed on current quantum hardware. Tests found that accuracy increased as the number of qubits rose from 1 to 4 to 16, leading the researchers to develop a 49-qubit system that could directly process full 28-by-28-pixel MNIST images.

The researchers also created an automated method for selecting the quantum circuit components used by the image classifier. On IBM’s Heron r2 processor, the approach reached 96.08% accuracy and converged about 70% faster than a classical convolutional neural network under identical training conditions, with a similar number of adjustable parameters.

What the quantum system achieved

The researchers introduced fragment encoding to reduce the input dimensionality of images for quantum processing. In their tests, accuracy improved as the system grew from 1 to 4 to 16 qubits. They then developed a 49-qubit quantum convolutional neural network capable of directly processing 28-by-28-pixel MNIST images, avoiding classical dimensionality-reduction preprocessing. An automated framework used measures of expressibility, entanglement and circuit complexity to select the quantum circuits used as building blocks. On IBM’s Heron r2 quantum processor, the system achieved 96.08% accuracy and converged approximately 70% faster than the compared classical convolutional neural network under identical training conditions.

Why the result matters

Current quantum processors are limited by noise and by the number of data inputs they can handle. The study presents an approach that reduces those demands while allowing a quantum system to process full-size MNIST images directly. Its results suggest that simple image-classification tasks can be carried out on real quantum hardware today, although the findings do not establish how the approach would perform on more complex images or tasks.

Evidence and limitations

The evidence comes from experiments run on IBM’s Heron r2 quantum processor, including comparisons with other quantum and classical image-classification models. The abstract reports results for MNIST, a relatively simple handwritten-digit dataset, and does not provide details such as the number of test examples, uncertainty estimates or performance on other datasets. The reported speed comparison applies to the stated identical training conditions and does not by itself show an advantage for broader machine-learning workloads.

Generative AI could make rule-based decision tools easier to use

A paper in Operations Research proposes a framework for combining generative AI with mathematical optimization. The approach is intended to make complex decision tools more understandable, interactive and adaptable while keeping rules and constraints in place.

22 hours ago·6 views

AI & Computing

Adding economic data made construction-material price forecasts up to 59% better

A forecasting framework predicted construction-material prices at the detailed six-digit level used in the industry’s classification system. Adding raw-material prices, commodity indexes and economic indicators improved every tested model, while a long short-term memory model performed best and improved on ARIMA by up to 59%.

1 day ago·9 views

AI & Computing

A female virtual body made men’s and women’s responses to abuse more alike

Men and women responded similarly when embodied in a female virtual body during a simulated domestic violence scenario. The study found comparable physical arousal, reaction times, presence and sense of inhabiting the avatar, although some interpretations of the experience differed.

2 days ago·15 views