WiMi Hologram Cloud Releases H-QNN Tech, Demonstrating Progress in Practical Quantum Computing

WiMi Hologram Cloud Inc. (NASDAQ: WiMi) has announced a breakthrough in practical quantum computing with the release of its Hybrid Quantum-Classical Neural Network (H-QNN) technology on February 06, 2026. This new approach tackles the limitations of traditional deep learning in high-dimensional image recognition, specifically demonstrating efficient binary image classification using the MNIST handwritten digit dataset. H-QNN combines the power of quantum feature mapping with established classical deep learning mechanisms, potentially overcoming issues like overfitting and computational complexity. This achievement, according to WiMi, “marks a new progress in quantum machine learning moving from theoretical exploration toward practicalization,” and embodies their core competitiveness in quantum intelligent algorithm research.

H-QNN Architecture Integrates Quantum and Classical Components

A novel approach to machine learning, dubbed the Hybrid Quantum-Classical Neural Network (H-QNN), is demonstrating promising results in image recognition, potentially bridging the gap between theoretical quantum computing and practical application. Developed by WiMi Hologram Cloud Inc., the H-QNN tackles limitations inherent in both traditional deep learning and nascent quantum networks by strategically combining their strengths. The technology’s initial success centers on efficient classification of the MNIST dataset of handwritten digits, a benchmark in computer vision. The core innovation lies in a trainable quantum feature encoding module positioned at the front end of a classical neural network.

This module maps image data into a “high-dimensional quantum feature space” before processing with quantum circuits, a process designed to overcome bottlenecks in classical computing architectures. Traditional convolutional neural networks (CNNs) and multi-layer perceptrons (MLPs) often struggle with complex feature mapping and nonlinear discrimination in high-dimensional data, leading to issues like overfitting and high computational complexity; the H-QNN aims to circumvent these challenges. “QNN can represent complex feature distributions in an exponentially large Hilbert space by leveraging quantum superposition and entanglement characteristics,” highlighting the theoretical advantage of quantum feature expression.

The H-QNN architecture comprises three key modules: data preprocessing, quantum encoding/feature extraction, and a classical neural classifier. Data preprocessing involves binarization, normalization, and dimensionality reduction of the 28×28 pixel MNIST images. WiMi employs a “screening method based on statistical feature distribution” to ensure high feature representativeness, minimizing the creation of invalid quantum states. The quantum encoding stage utilizes a Parameterized Quantum Circuit (PQC) – a series of rotation and entanglement gates – to map pixel data into quantum states, embedding numerical information into quantum amplitudes and phases. Crucially, the system doesn’t rely solely on quantum processing.

After quantum feature extraction, the results are fed into a lightweight multi-layer perceptron for final classification, leveraging the strengths of classical deep learning in parameter optimization. WiMi has implemented a “hybrid optimization strategy based on gradient estimation” to ensure stable training across both quantum and classical components, calculating gradients within the quantum circuit using the Parameter Shift Rule. Experimental results on the MNIST dataset, distinguishing between handwritten “0” and “1”, show that H-QNN achieves significantly higher classification accuracy than equivalent-scale classical MLP models, even with smaller datasets, suggesting improved generalization and robustness.

Parameterized Quantum Circuits Encode MNIST Image Features

A novel approach to image classification leveraging the principles of quantum computing is demonstrating promising results with the iconic MNIST handwritten digit dataset. WiMi Hologram Cloud Inc. has unveiled a Hybrid Quantum-Classical Neural Network (H-QNN) designed to overcome limitations inherent in traditional deep learning architectures when processing complex, high-dimensional data. Central to this design is a “trainable quantum feature encoding module” utilizing Parameterized Quantum Circuits (PQCs). Unlike purely quantum or classical models, the H-QNN avoids the pitfalls of noisy quantum hardware while retaining the potential for quantum speedup.

Following this, the PQC constructs nonlinear quantum feature mappings, and the resulting data is fed into a classical neural classifier – a lightweight multi-layer perceptron. Experimental results reveal a significant performance boost. WiMi observed “a nonlinear growth in the model’s feature expression capability when the number of qubits expanded from 4 to 8,” confirming the scalability of the quantum feature space. Furthermore, simulations indicate a roughly 30% reduction in computation time compared to traditional deep networks, hinting at substantial acceleration potential with mature quantum hardware. This technology, WiMi believes, is not limited to MNIST, but represents “a general quantum-enhanced neural network framework” applicable to diverse computer vision tasks.

Hybrid Optimization Strategy Enables Gradient-Based Training

A novel approach to training hybrid quantum-classical neural networks (H-QNNs) is yielding promising results, allowing for efficient learning even with the constraints of current quantum hardware. WiMi Hologram Cloud Inc. has developed a system that moves beyond theoretical quantum machine learning, demonstrating practical feasibility in image recognition tasks. Central to this advancement is a strategy for optimizing both the quantum and classical components of the network using gradient-based methods – a cornerstone of modern deep learning. The challenge lies in bridging the gap between the fundamentally different ways quantum and classical systems are optimized.

Traditional deep learning relies on backpropagation to adjust parameters, but directly applying this to the quantum realm is problematic. This isn’t simply about adding quantum elements to an existing classical network; it’s about synergistic enhancement. The system leverages the exponential expressive power of quantum computing for feature mapping while retaining the robust parameter optimization capabilities of classical deep learning. WiMi’s experiments, focused on classifying handwritten digits from the MNIST dataset, reveal significant performance gains. Beyond accuracy, computational efficiency is also improved. The company envisions extending this framework to applications like medical image analysis and video processing, adapting the quantum encoding and circuit depth to suit diverse datasets.

H-QNN Achieves Enhanced Accuracy and Efficiency on MNIST Dataset

Initial trials focused on the MNIST handwritten digit dataset, a standard benchmark for computer vision algorithms, revealing significant gains in both accuracy and computational efficiency. Data preprocessing within the H-QNN begins with binarization and normalization of the 28×28 pixel MNIST images, followed by dimensionality reduction to a quantum-compatible format. The core of the quantum processing utilizes Parameterized Quantum Circuits (PQCs), comprising rotation and entanglement gates, to construct nonlinear feature mappings. Crucially, experiments reveal a performance edge over comparable classical models.

In binary classification of digits “0” and “1” from the MNIST dataset, H-QNN demonstrated significantly higher accuracy “under the same number of training epochs and sample scale” than classical MLPs. Furthermore, the model exhibited robust generalization even with limited datasets, suggesting the quantum feature mapping effectively mitigates overfitting.