Hybrid Quantum-Classical Neural Network
The unveiling of WiMi Hologram Cloud Inc.’s Hybrid Quantum-Classical Neural Network (H-QNN) technology marked a major advancement in the real-world implementation of quantum machine learning. Specifically created for effective MNIST binary image classification, this new system marks a calculated shift from theoretical investigation to practical industry application. The development occurs at a crucial moment for the business, which is trying to use its intellectual property to counteract a difficult stock market year.
You can also read Infleqtion, Indiana Quantum Corridor’s GPS-Free Timing trial
Bridging Quantum-Classical
Multi-layer perceptrons (MLPs) and convolutional neural networks (CNNs), two classic deep learning architectures, have long been the industry standard for image recognition. These models often encounter “bottlenecks” while processing high-dimensional data, such as computational cost, gradient vanishing, and overfitting.
The H-QNN incorporates a trainable quantum feature encoding module at the classical network’s front end. Through the use of quantum superposition and entanglement, the system is able to map raw image data into an exponentially huge Hilbert space, expressing features with a level of complexity that is significantly higher than what is possible with classical methods. Get beyond the high noise and qubit restrictions that current quantum hardware frequently encounters, WiMi has developed a “synergistic enhancement” that combines quantum feature mapping with the sophisticated parameter optimization of classical deep learning.
Three Fundamental Components of Architectural Innovation
The three functional levels of the H-QNN architecture are carefully separated:
- Data Preprocessing: The MNIST dataset’s 28×28 pixel images are normalized and binarized. WiMi uses a statistical feature distribution-based screening technique to make sure the data is “quantumizable,” which lowers the production of invalid quantum states.
- Feature extraction and quantum encoding: a parameterized quantum circuit (PQC) is used in this step. Numerical information is embedded into quantum amplitudes or phases by the PQC using rotation gates (Ry, Rz) and entanglement gates (CNOT, CZ). As a result, every image sample has a distinct global representation in the quantum state space.
- Classical Neural Classifier: A lightweight MLP receives the quantum stage measurement results. The model simultaneously changes the quantum circuit parameters and classical weights via classical backpropagation.
WiMi developed a hybrid optimization technique based on the Parameter Shift Rule to guarantee training stability. This technique effectively converges the entire network by enabling accurate gradient estimation inside the quantum circuits.
You can also read LANL’s Center for Quantum Computing to Advance Moore’s Law
Success and Scalability of Experiments
Through extensive testing aimed at differentiating between the handwritten “0” and “1,” the H-QNN showed definite computational benefits. The H-QNN outperformed classical MLP models of a comparable scale in terms of classification accuracy, according to experimental results. Additionally, the model demonstrated improved generalization performance and increased resistance to noise, even when trained on fewer datasets, indicating that quantum feature mapping successfully reduces overfitting.
One of the most notable results was the decrease in computing time. In simulated settings, the H-QNN outperformed conventional deep networks in terms of processing time, cutting it by about 30%. Additionally, WiMi confirmed the scalability of the quantum feature space by reporting a nonlinear development in feature expression capabilities as the system grew from 4 to 8 qubits.
A Path Forward for Quantum Intelligence
Although the MNIST dataset is used to illustrate the recent success, WiMi sees the H-QNN as a general-purpose framework. The company intends to expand this technique to more intricate computer vision applications, such as handwriting recognition, video frame feature extraction, and medical picture analysis.
In contrast to simulation environments, future research will concentrate on confirming the H-QNN’s performance on real quantum gear. Building a complete quantum intelligence ecosystem is another goal of WiMi, which also seeks to investigate integrations with other quantum algorithms including quantum support vector machines (QSVM) and quantum convolutional networks.
You can also read Pakistan’s Inaugural National Quantum Computing Hackathon
Context of the Market and Investor Prospects
Despite advancements in technology, WiMi has experienced difficulties with its financial success. The stock of the company has dropped more than 80% in the last year and is currently trading close to its 52-week low. But according to market analysts, WiMi is selling at an exceptionally low Price/Book ratio of 0.11, which raises the possibility that the market is undervaluing the company’s technological assets and broad range of quantum-AI research.
This release is the most recent in a string of quantum-AI announcements from WiMi, which also launched its QB-Net in late 2025 and its MC-QCNN for multi-channel learning in January 2026. A consistent dedication to becoming a leading supplier of holographic cloud and quantum computing systems is demonstrated by these periodic technical improvements. The “important pillar” of future artificial intelligence, according to WiMi, will be hybrid architectures like the H-QNN as quantum hardware continues to advance.
