Insider Brief
- WiMi announced a hybrid quantum convolutional neural network (QCNN) architecture featuring a Quantum Kernel Convolution (QKC) scheme designed to run on current NISQ quantum hardware.
- The approach combines classical neural networks with quantum convolution and pooling layers to perform image classification while reducing parameter counts and computational complexity.
- Experimental tests using the MNIST dataset showed classification performance comparable to traditional CNNs despite using fewer parameters.
- Photo from Pexels by Yusuf P.
PRESS RELEASE — WiMi Hologram Cloud Inc. (NASDAQ: WiMi) (“WiMi” or the “Company”), a leading global Hologram Augmented Reality (“AR”) Technology provider, announces the release of a core technology for hybrid Quantum Convolutional Neural Network (QCNN), proposing and implementing a Quantum Kernel Convolution (QKC) scheme capable of running on current noisy intermediate-scale quantum (NISQ) devices, thereby providing a practically feasible engineering path for quantum-enhanced image classification models.
The core objective of this technology is not simply to embed quantum circuits into classical neural networks, but rather, starting from the computationally intensive core operation of convolution, to rethink the computational approach to feature extraction and dimensionality reduction. WiMi points out that classical convolutional layers essentially rely on sliding windows and linear weighted summation to accomplish local feature extraction, whereas quantum computing inherently possesses the capability of high-dimensional Hilbert space representation and quantum parallelism. If local image patches can be mapped into quantum states and feature mixing can be achieved through controlled entanglement evolution, it becomes possible to realize an equivalent or even more expressive feature extraction mechanism under a lower parameter scale.
WiMi points out that this pooling approach is essentially an information reallocation and selection mechanism, which can achieve dimensionality compression without explicitly discarding information, thereby significantly reducing the computational burden on subsequent quantum circuits and classical networks.
At the overall system architecture level, this hybrid QCNN adopts a layered design of classical-quantum synergy. The classical neural network is responsible for completing preliminary normalization of input data, dimensionality adjustment, and final classification decisions, while the quantum convolutional layer is embedded at the critical position of feature extraction, functioning as a quantum acceleration module. This design enables the model to fully leverage mature classical deep learning toolchains while introducing quantum advantages at key computational nodes, thereby avoiding, from an engineering perspective, the scalability issues that fully quantum models face under current hardware conditions.
In terms of technical implementation, WiMi, based on the Qiskit quantum computing development framework, completed a complete engineering realization from quantum circuit construction, parameterized training, to integration with classical deep learning frameworks. The quantum convolutional layer is encapsulated as a reusable module interface that can be directly embedded into existing deep learning training workflows. During the training process, the model adopts a hybrid optimization strategy: classical backpropagation algorithms are used to update the parameters of the classical network, while the parameter-shift rule is utilized to estimate gradients for the quantum circuit parameters, achieving end-to-end joint training. This implementation path effectively addresses the challenge of gradient propagation between quantum and classical components, providing engineering assurance for the trainability of hybrid models.
In the experimental validation phase, WiMi selected the MNIST handwritten digit dataset as the benchmark task and conducted a systematic evaluation of the proposed hybrid QCNN model. The experimental results show that, with a significantly lower number of parameters compared to traditional CNN models, this hybrid model is still able to achieve classification accuracy comparable to that of classical models. Particularly noteworthy is that after replacing some classical convolutional layers with quantum convolutional layers, the overall parameter scale and computational complexity of the model are effectively controlled while maintaining stable convergence performance. These results demonstrate that quantum kernel convolution possesses practical feasibility in real tasks, rather than remaining merely at the theoretical level.
Furthermore, through analysis of intermediate quantum states and measurement outcomes, WiMi verified the effectiveness of the entanglement-based quantum pooling mechanism in the dimensionality reduction process. Experiments show that quantum pooling not only compresses feature dimensions but also preserves discriminative information critical to the classification task. This finding provides a new entry point for interpretability research in quantum neural networks and lays the foundation for subsequent extensions to more complex datasets and tasks.
This hybrid quantum convolutional neural network technology is not an isolated algorithmic innovation, but rather an important step taken around WiMi’s long-term strategic goal of deployable quantum-enhanced artificial intelligence. By emphasizing low depth, modularity, and high compatibility with existing AI ecosystems, this technology provides a realistic path for quantum computing to move from the laboratory to practical applications. In the future, further exploration will be conducted on the application potential of this architecture in higher-resolution images, multi-channel data, and other perception tasks, while continuously optimizing circuit designs in conjunction with the development of quantum hardware.
The release of WiMi’s hybrid neural network quantum kernel convolution technology marks an important step forward for quantum machine learning, moving from proof-of-concept toward engineering implementation. It not only demonstrates the practical value of quantum computing in real-world image recognition tasks but also provides clear design ideas for the future construction of quantum-classical collaborative computing systems. With the continuous improvement of quantum hardware performance and the ongoing maturation of development toolchains, the hybrid QCNN framework built by WiMi is expected to play a role in a broader range of artificial intelligence applications, becoming an important component of next-generation intelligent computing technology.
