BEIJING, June 24, 2026 /PRNewswire/ -- WiMi Hologram Cloud Inc. (NASDAQ: WIMI) ("WIMI" or the "Company"), a leading global Hologram Augmented Reality ("AR") Technology provider, has completed systematic benchmark testing on fully parameterized quantum convolutional neural networks. Its research team has proposed a quantum neural network model inspired by classical convolutional neural networks. Throughout the entire computational process, this model only adopts two-qubit interactions. It not only retains the simplicity of the network architecture but also greatly lowers the implementation difficulty of quantum circuits, laying a crucial foundation for its practical deployment on noisy intermediate-scale quantum computers in the future.

Unlike traditional deep learning networks, classical convolutional neural networks generally rely on massive parameters and sophisticated hierarchical structures to perform feature extraction. For instance, in image classification tasks, convolutional layers continuously scan input images to extract local features, after which pooling layers compress feature dimensions, and fully connected layers ultimately generate classification decisions. Although this architecture has achieved remarkable success, the training and inference costs surge exponentially as model scales keep expanding.

The research team at WIMI holds that quantum computing is inherently capable of processing high-dimensional feature spaces. A quantum system consisting of n qubits can represent a 2ⁿ-dimensional state space simultaneously, which enables it to tackle complex pattern recognition tasks with far fewer parameters than classical algorithms. This potential advantage has driven the research team to rethink the implementation of convolutional neural networks and attempt to build convolutional quantum neural network architectures via quantum gate operations.

To this end, WIMI has designed a fully parameterized Quantum Convolutional Neural Network (QCNN) architecture. The network mainly comprises a quantum data encoding layer, a quantum convolutional layer, a quantum pooling layer, a feature compression layer and a quantum classification layer. Distinct from many complex quantum networks built on multi-qubit gate operations, this model takes two-qubit interactions as its basic computational unit, which effectively controls circuit depth and mitigates noise accumulation.

In the data input phase, raw classical images first go through preprocessing modules for dimensionality reduction and normalization. Given the limited number of qubits supported by current quantum hardware, researchers need to map high-dimensional image data onto a finite qubit space. To address this challenge, WIMI has investigated multiple classical data preprocessing strategies including principal component analysis, feature selection and image compression techniques, ensuring the input data adapts to quantum computing resource constraints while preserving core information.

Once preprocessing finishes, data proceeds to the quantum encoding stage. As a cornerstone of quantum machine learning, quantum encoding converts classical data into quantum state representations. WIMI has systematically compared a range of quantum encoding schemes such as angle encoding, amplitude encoding and hybrid encoding methods. Experimental results reveal that different encoding strategies directly determine the network's representation capacity and training efficiency. Angle encoding leverages rotation gates to map data into the parameter space of quantum states, featuring straightforward implementation and strong noise robustness. By contrast, amplitude encoding can express higher-dimensional data with fewer qubits yet comes with relatively higher implementation complexity.

After data is encoded into quantum states, the network executes quantum convolution operations. Classical convolutional neural networks extract features by sliding convolutional kernels over local regions, while QCNN implements convolution via parameterized two-qubit gates. WIMI has constructed a series of trainable quantum gate arrays that establish correlations between different features through quantum entanglement. Since quantum states can exist as superpositions of multiple states concurrently, the network can process numerous potential feature combinations in parallel within a single computation, delivering more efficient feature extraction than classical convolution methods.

Notably, the model follows a fully parameterized design principle. While certain quantum gate parameters are often fixed in conventional quantum neural networks, the architecture proposed by WIMI allows all key quantum gate parameters to be updated during training. This design substantially boosts the model's expressive power, enabling it to learn intricate data distribution patterns.

Following quantum convolution, the network enters the quantum pooling stage. Classical pooling layers reduce feature dimensions through max pooling or average pooling, whereas quantum pooling screens valid information via measurement, entanglement reconstruction and quantum state compression. By gradually cutting down the number of qubits involved in computation, the network reduces subsequent computational complexity while retaining critical feature information.

From an information processing perspective, this mechanism mirrors the feature abstraction process in classical deep learning. As network layers deepen, low-level features including edges, textures and shapes in input images are progressively converted into high-level semantic abstract representations, supporting the final classification task.

WIMI has tested multiple QCNN configurations covering diverse parameterized quantum circuit architectures, data encoding schemes, loss functions and optimization algorithm combinations. Experimental results demonstrate that QCNN achieves outstanding classification performance across most test scenarios. Most importantly, even with far fewer parameters than classical convolutional neural networks, QCNN matches or even exceeds the classification accuracy of traditional CNNs. This outcome proves that quantum neural networks deliver higher parameter utilization efficiency; in other words, quantum models can learn richer data features with fewer trainable parameters and thus achieve superior overall performance.

Further analysis by WIMI attributes this competitive edge to the high-dimensional feature representation capability enabled by quantum entanglement. In classical neural networks, information exchange between neurons is constrained by connection topology, yet quantum entanglement builds non-classical correlations among multiple qubits, empowering the network with stronger representation capacity under limited computing resources. To enhance training stability, the research team has thoroughly explored optimization workflows. Training quantum neural networks frequently suffers from vanishing gradients and flat parameter regions, an issue known as the barren plateau phenomenon. To resolve this problem, the team evaluated various optimizers including stochastic gradient descent, Adam optimizer and quantum-specific optimization algorithms, and analyzed how different cost functions affect model training.

Experimental results verify that well-designed parameter initialization strategies and optimization pipelines can effectively ease training difficulties, allowing the network to converge stably within fewer training epochs. This research delivers valuable practical experience for training large-scale quantum neural networks in the future.

It is foreseeable that with the continuous advancement of quantum computing technologies, quantum convolutional neural networks will evolve into a core component of next-generation intelligent computing. From current classical data classification experiments to large-scale artificial intelligence applications in complex future scenarios, QCNN exhibits enormous potential to reshape the evolution of machine learning. WIMI's research not only advances the theoretical development of quantum machine learning but also pioneers innovative technical routes for developing efficient, low-parameter and high-performance new artificial intelligence systems, laying a solid foundation for the advent of the quantum intelligence era.

About WiMi Hologram Cloud

WiMi Hologram Cloud Inc. (NASDAQ: WiMi) focuses on holographic cloud services, primarily concentrating on professional fields such as in-vehicle AR holographic HUD, 3D holographic pulse LiDAR, head-mounted light field holographic devices, holographic semiconductors, holographic cloud software, holographic car navigation, metaverse holographic AR/VR devices, and metaverse holographic cloud software. It covers multiple aspects of holographic AR technologies, including in-vehicle holographic AR technology, 3D holographic pulse LiDAR technology, holographic vision semiconductor technology, holographic software development, holographic AR virtual advertising technology, holographic AR virtual entertainment technology, holographic ARSDK payment, interactive holographic virtual communication, metaverse holographic AR technology, and metaverse virtual cloud services. WiMi is a comprehensive holographic cloud technology solution provider. For more information, please visit http://ir.wimiar.com.

Translation Disclaimer

The original version of this announcement is the officially authorized and only legally binding version. If there are any inconsistencies or differences in meaning between the Chinese translation and the original version, the original version shall prevail. WiMi Hologram Cloud Inc. and related institutions and individuals make no guarantees regarding the translated version and assume no responsibility for any direct or indirect losses caused by translation inaccuracies.

Source: WiMi Hologram Cloud Inc. Related Stocks: NASDAQ:WIMI