Beijing, Feb. 06, 2026 (GLOBE NEWSWIRE) -- WiMi Releases Hybrid Quantum-Classical Neural Network (H-QNN) Technology for Efficient MNIST Binary Image Classification
BEIJING, Feb.06, 2026––WiMi Hologram Cloud Inc. (NASDAQ: WiMi) ("WiMi" or the "Company"), a leading global Hologram Augmented Reality ("AR") Technology provider, today announced the release of a Hybrid Quantum-Classical Neural Network (Hybrid Quantum-Classical Neural Network, H-QNN) technology for efficient MNIST binary image classification. This breakthrough achievement marks a new progress in quantum machine learning moving from theoretical exploration toward practicalization, and also embodies the enterprise's core competitiveness in the field of quantum intelligent algorithm research. This technology takes an efficient hybrid structure, scalable quantum feature mapping mechanism, and quantum state optimization strategy as its core, successfully achieving excellent classification performance on the MNIST handwritten digit dataset, proving the practical feasibility and computational advantages of quantum neural networks in high-dimensional image recognition tasks.
Under the traditional framework of deep learning, convolutional neural networks (CNN) have long played the core role in image feature extraction, while multi-layer perceptrons (MLP) undertake the final classification tasks. However, these models are still constrained by the bottlenecks of classical computing architectures, especially in the complex feature mapping and nonlinear discrimination of high-dimensional data, where models are prone to issues such as overfitting, gradient vanishing, and high computational complexity. The emergence of quantum computing provides new solutions to this problem. Quantum neural networks (QNN) can represent complex feature distributions in an exponentially large Hilbert space by leveraging quantum superposition and entanglement characteristics, thereby theoretically achieving feature expression capabilities far surpassing those of classical neural networks.
The hybrid quantum-classical neural network (H-QNN) technology proposed by WiMi was born precisely under this technological trend. H-QNN introduces a trainable quantum feature encoding module at the front end of the classical network, mapping raw image data into a high-dimensional quantum feature space, then performing nonlinear feature transformations using quantum circuits, and finally conducting subsequent classification decisions through the classical network. This structure fully combines the exponential expressive power of quantum computing in feature mapping with the mature mechanisms of classical deep learning in large-scale parameter optimization, achieving synergistic enhancement between quantum and classical computing. Unlike pure quantum networks or classical deep models, H-QNN not only avoids the limitations of high noise and limited qubit numbers in quantum hardware but also retains the potential acceleration advantages of quantum algorithms in feature extraction.
From the perspective of architectural design, the H-QNN technology consists of three main parts: the data preprocessing module, the quantum encoding and feature extraction module, and the classical neural classifier. First, the data preprocessing module performs binarization and normalization operations on the 28×28 pixel images of MNIST, and reduces the image dimensionality to a quantumizable data format through compression and block strategies. At this stage, WiMi uses a screening method based on statistical feature distribution to ensure that the data input into the quantum system has high feature representativeness, thereby reducing the generation of invalid quantum states.
After entering the quantum encoding stage, H-QNN adopts a Parameterized Quantum Circuit (PQC) as the core computing unit. The PQC is composed of several layers of quantum gates, which structurally include rotation gates (R_y, R_z) and entanglement gates (CNOT, CZ) and other operations, used to construct nonlinear quantum feature space mappings. Image pixels or locally extracted feature vectors are mapped to quantum states, embedding numerical information into quantum amplitudes or phases through quantum rotation encoding. This process achieves the quantum expression of high-dimensional nonlinear features, making each sample possess a unique global representation in the quantum state space.
In the core stage of feature extraction, H-QNN simulates complex high-dimensional decision boundaries through quantum state evolution. The superposition and entanglement properties of quantum states enable the model to simultaneously capture multiple feature correlation relationships in a single evolution. Through the iterative action of multi-layer quantum circuits, the quantum states of the input data are mapped to new feature distributions, and the output results, after measurement, yield a set of interpretable feature vectors. Unlike traditional convolutional or fully connected layers, this quantum transformation layer can theoretically achieve exponential feature space expansion without significantly increasing the number of parameters.
The measurement results serve as intermediate feature vectors input into the classical neural classifier part. This part adopts a lightweight multi-layer perceptron structure, composed of several fully connected layers and nonlinear activation functions. Through classical backpropagation algorithms, the model can simultaneously update the quantum circuit parameters and classical weights, thereby achieving hybrid optimization. To maintain the training stability of the quantum and classical parts, WiMi introduces a hybrid optimization strategy based on gradient estimation. This strategy precisely calculates the gradients of trainable parameters in the quantum circuit through the Parameter Shift Rule, thereby ensuring the differentiability and convergence of the overall network during the training process.
