WiMi Studies Hybrid Quantum-Classical Learning Architecture for Multi-Class Image Classification
Rhea-AI Summary
WiMi (NASDAQ: WIMI) announced a hybrid quantum-classical learning architecture for multi-class image classification on December 4, 2025. The design reuses measurement results from both retained and discarded qubits in quantum convolutional neural networks (QCNN), feeding them into separate classical fully connected branches that are fused and jointly trained with quantum parameters.
The approach aims to reduce information loss from quantum pooling on NISQ devices, improve expressive power of hybrid models, and enable co-evolution of quantum gate angles and classical weights via cross-entropy backpropagation. WiMi positions this as a practical path for QML under current hardware constraints, with applications in intelligent vision, medical diagnosis, and autonomous driving.
Positive
- None.
Negative
- None.
Key Figures
Market Reality Check
Peers on Argus 1 Up
Sector peers showed mixed action: ABLV appeared in momentum scans, up ~30.6% without news, while other advertising/communications names moved modestly. With WIMI up 4.42% and only one peer in strong upside momentum, the setup looks more stock-specific than a broad sector rotation.
Historical Context
| Date | Event | Sentiment | Move | Catalyst |
|---|---|---|---|---|
| Nov 20 | Quantum QGAN R&D | Positive | -5.6% | Announced dual‑discriminator hybrid quantum‑classical QGAN architecture to enhance training stability. |
| Nov 14 | Quantum GAN research | Positive | -1.7% | Detailed QGAN and hybrid quantum‑classical CNN models for faster synthetic image generation. |
| Oct 24 | Post‑quantum blockchain | Positive | -0.5% | Outlined blockchain privacy system using post‑quantum cryptography and threshold key‑sharing. |
| Oct 24 | Post‑quantum blockchain | Positive | -0.5% | Repeated disclosure of post‑quantum blockchain privacy protection research details. |
| Oct 23 | Hybrid QCNN model | Positive | +5.9% | Presented shallow hybrid quantum‑classical CNN to improve image classification efficiency. |
Recent technology R&D announcements, especially around quantum and blockchain, often coincided with flat-to-negative next-day moves, suggesting limited short-term price follow-through on such news.
Over the last few months, WiMi repeatedly highlighted quantum and post‑quantum R&D. On Oct 23, a shallow hybrid quantum‑classical CNN for image classification saw a 5.9% rise. Later blockchain privacy and post‑quantum security research on Oct 24 had a mild -0.51% reaction. Two November QGAN and hybrid quantum‑classical releases (Nov 14 and Nov 20) drew -1.68% and -5.61% moves, indicating that similar technical advances have not consistently driven upside.
Regulatory & Risk Context
Short interest at 4.13% of float with 2.71 days to cover indicates relatively low crowding on the short side, limiting the potential impact of short covering on volatility around news.
Market Pulse Summary
This announcement details a hybrid quantum‑classical learning architecture that reuses discarded qubit information, aiming to improve multi‑class image classification on NISQ hardware. It extends a series of WiMi quantum and AI research updates seen since October 2025. Investors comparing this with prior events might watch for concrete commercialization steps, changes in filing activity, and how the stock behaves relative to its 200‑day moving average and low short interest of 4.13%.
Key Terms
quantum convolutional neural networks (QCNN) technical
quantum entanglement technical
cross-entropy loss technical
AI-generated analysis. Not financial advice.
On the basis of in-depth research on quantum convolutional neural networks (QCNN), it proposed a new type of hybrid quantum-classical learning technology. This technology, through innovatively recycling discarded qubit state information and joint training with classical fully connected layers, achieves significant performance improvements in multi-class image classification tasks.
This achievement not only optimizes the efficiency of quantum networks under the conditions of noisy intermediate-scale quantum (NISQ) devices but also demonstrates the possibility of quantum information reuse, opening up a brand-new development path for hybrid quantum-classical models.
Image classification is one of the core applications of artificial intelligence. From face recognition to medical image analysis, deep convolutional neural networks (CNN) have become the mainstream. However, as the model depth increases, its training time and computational energy consumption grow exponentially, with the dependence on hardware computing power becoming increasingly strong. Even under the support of GPU clusters or TPU arrays, model optimization is still constrained by bottlenecks. On the other hand, issues such as data security, privacy protection, and computational energy efficiency are forcing academia and industry to rethink the underlying architecture of intelligent computing.
Quantum computing provides a completely new approach. It utilizes quantum superposition and entanglement effects to process information simultaneously in an exponential space, bringing theoretical acceleration advantages for complex pattern recognition tasks. Based on this characteristic, quantum machine learning (QML) is considered the next stage of artificial intelligence development. However, current quantum computers are still in the NISQ stage, with limited qubit numbers and susceptibility to noise interference, making how to achieve stable and scalable quantum learning algorithms under this hardware constraint the core problem that urgently needs to be solved.
Traditional QCNN, as a representative structure of QML, inherits the hierarchical feature extraction idea of CNN and achieves quantum feature mapping and quantum pooling through quantum gate operations. However, unlike classical CNN, the pooling operation in QCNN usually means that the "discarded" qubits—measured or dimension-reduced qubits—will no longer participate in subsequent computations. These discarded qubits often have entanglement relationships with the retained qubits, and their internals still contain potential correlation information. Previous research has mostly ignored this "discarded" quantum information, while WiMi precisely starts from this point, proposing a new idea: can the discarded qubits be allowed to re-participate in decision-making, thereby enhancing the model's overall expressive ability.
To solve this key problem, WiMi designed a hybrid quantum-classical learning architecture. The core innovation of this architecture lies in simultaneously utilizing the information of retained qubits and discarded qubits, thereby achieving maximum utilization of quantum information at the feature level.
In traditional QCNN, after several quantum convolutional layers and quantum pooling layers, some qubits are measured or removed to reduce the system dimension and achieve downsampling operations. The pooling operation in classical CNN selectively retains high-activation features, while the pooling in QCNN is usually achieved by measuring or discarding some qubits. Due to the existence of quantum entanglement, there is often non-local quantum correlation between discarded qubits and retained qubits; directly discarding this part of qubits is equivalent to information loss.
In the architecture proposed by WiMi, all discarded qubits, after measurement, have their measurement results retained and input into an independent classical fully connected branch. At the same time, the measurement results of the retained qubits are input into another fully connected branch. These two branches perform nonlinear transformations and feature compression respectively, and then undergo vector-level concatenation and weight integration in the fusion layer. Finally, the fused comprehensive features complete the final prediction through a joint classification layer.
This structure can be regarded as a quantum-classical dual-channel feature fusion network. It not only compensates for the quantum information loss in the pooling stage of QCNN but also enables the co-evolution of quantum parameters (determined by quantum gate angles) and classical parameters (determined by weight matrices) through joint optimization strategies, thereby achieving adaptive improvement in global performance.
In this architecture, the training process adopts a joint optimization mechanism based on classical cross-entropy loss. WiMi treats the measurement probability distribution output by the quantum circuit as a feature vector, which, together with the output of the classical layer, is input into a fusion network for backpropagation.
The significance of this technology lies in that it redefines the information utilization method in hybrid quantum-classical learning models. Traditional quantum neural networks pursue quantum purity in structure, that is, maintaining the fully quantumized processing process as much as possible. However, WiMi's research shows that, under current quantum hardware conditions, the synergistic integration of quantum and classical is instead the key to achieving practical performance breakthroughs. By fully utilizing the information of discarded qubits, it breaks the inherent assumption that quantum pooling means information loss, enabling quantum computing to achieve a balance between information utilization rate and energy efficiency.
WiMi's hybrid quantum-classical learning technology for multi-class image classification represents a new direction in quantum intelligence. It does not rely on idealized quantum hardware but explores feasible optimal paths under the real NISQ constraints. This achievement demonstrates the powerful potential of quantum machine learning in image understanding, pattern recognition, and cross-domain feature fusion, and also provides a practical engineering sample for the deep integration of quantum information science and artificial intelligence.
In the future where quantum computing is gradually moving toward practicalization, hybrid quantum-classical models will become the key bridge connecting theory and industry. Through continuous optimization of quantum circuit design, information recycling strategies, and cross-domain training methods, this technology will bring disruptive innovative power to fields such as intelligent vision, medical diagnosis, and autonomous driving.
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.
View original content:https://www.prnewswire.com/news-releases/wimi-studies-hybrid-quantum-classical-learning-architecture-for-multi-class-image-classification-302633366.html
SOURCE WiMi Hologram Cloud Inc.
FAQ
What did WiMi (WIMI) announce on December 4, 2025 about quantum-classical learning?
How does WiMi's hybrid architecture address information loss in QCNN on NISQ devices?
What training method does WiMi use for the quantum-classical model (WIMI)?
Which image tasks does WiMi say the hybrid model targets for WIMI technology?
Why does WiMi believe hybrid quantum-classical models are practical now (WIMI)?
What is the core innovation WiMi claims in its WIMI hybrid QCNN architecture?
