WiMi Hologram Cloud Inc. (NASDAQ: WiMi) has launched a next-generation hybrid quantum neural network structure (H-QNN) designed for image multi-classification. This technology integrates classical convolutional neural networks (CNNs) – leveraging their spatial feature extraction capabilities – with the high-dimensional nonlinear mapping features of quantum neural networks (QNNs). The resulting hybrid structure demonstrates stronger generalization ability and computational efficiency, systematically optimizing quantum-classical hybrid learning. WiMi’s H-QNN achieves classification accuracy and stability superior to similar algorithms, establishing a technical foundation for future quantum intelligent vision systems.
Hybrid Quantum Neural Network Structure Overview
WiMi has released a hybrid quantum neural network structure (H-QNN) designed for image multi-classification. This technology integrates classical convolutional neural networks (CNNs) for spatial feature extraction with quantum neural networks (QNNs) for high-dimensional nonlinear mapping. The H-QNN utilizes a three-stage distributed structure: convolutional feature extraction, quantum mapping, and hybrid decision-making. This design aims to improve generalization ability and computational efficiency compared to traditional quantum hybrid models that simply use quantum parts as classification heads.
The H-QNN addresses key challenges in quantum machine learning through several innovations. WiMi employs an angle encoding method combined with principal component analysis (PCA) to overcome quantum encoding dimension limitations and maximize quantum information utilization. A reconfigurable parameter sharing strategy and mixed state perturbations are used within the quantum state transformation module to alleviate gradient vanishing and maintain stable convergence, avoiding the “barren plateau” phenomenon.
To further enhance performance, WiMi incorporates transfer learning and an early stopping strategy based on quantum Fidelity. The system utilizes a heterogeneous computing architecture, running classical components on CPU/GPU platforms while executing quantum modules on FPGA-accelerated quantum simulators. This design enables nanosecond-level response times for quantum state updates, significantly improving overall training speed and demonstrating performance advantages over pure CPU or GPU simulations.
H-QNN Architectural Design and Modules
WiMi has developed a hybrid quantum neural network (H-QNN) structure designed for image multi-classification. This technology integrates classical convolutional neural networks (CNNs) for spatial feature extraction with quantum neural networks (QNNs) for high-dimensional nonlinear mapping. The H-QNN utilizes a three-stage distributed structure – convolutional feature extraction, quantum mapping, and hybrid decision-making – allowing the quantum portion to reconstruct information at the feature space level, not just classify. This architecture aims to address limitations of traditional quantum hybrid models.
The H-QNN incorporates several key innovations to improve performance. An angle encoding method, combined with principal component analysis (PCA), addresses quantum encoding dimension limitations by maximizing the utilization of quantum information. To prevent gradient vanishing during training, WiMi implemented a reconfigurable parameter sharing strategy and introduced mixed state perturbations. Additionally, a transfer learning mechanism migrates pre-trained quantum layer parameters to new tasks, reducing training time and enhancing model stability.
This H-QNN system employs a heterogeneous computing architecture, leveraging both CPU/GPU platforms for classical computing and FPGA-accelerated quantum simulators for quantum computations. The FPGA module enables nanosecond-level response times for quantum state updates, significantly improving overall training speed. An early stopping strategy, based on quantum Fidelity, further prevents overfitting by monitoring quantum state evolution during training, demonstrating performance advantages over pure CPU or GPU simulations.
By embedding trainable quantum layers into the foundation of classical neural networks, this technology achieves efficient utilization of quantum computing resources, enabling quantum advantages to be embodied in real visual tasks.
Quantum Encoding and Training Strategies
WiMi has developed a hybrid quantum neural network structure (H-QNN) designed for image multi-classification. This technology combines classical convolutional neural networks (CNNs) for feature extraction with quantum neural networks (QNNs) for high-dimensional mapping and discrimination. The H-QNN utilizes a three-stage process – convolutional feature extraction, quantum mapping, and hybrid decision-making – allowing quantum components to contribute to information reconstruction at the feature space level, not just classification.
The system addresses quantum encoding limitations with an angle encoding method combined with principal component analysis (PCA). Optimizing PCA’s cumulative variance contribution rate ensures high fidelity when mapping input features to quantum state amplitudes, maximizing quantum information use. Furthermore, WiMi employs a reconfigurable parameter sharing strategy and mixed state perturbations within the quantum layers to mitigate gradient vanishing—a common problem in training—and maintain stable convergence.
To enhance training and performance, WiMi incorporates transfer learning, migrating pre-trained quantum layer parameters to new tasks, reducing training epochs and improving stability. An early stopping strategy, based on quantum Fidelity metrics, prevents overfitting by monitoring quantum state evolution. This H-QNN utilizes a heterogeneous computing architecture, executing classical parts on CPUs/GPUs and quantum parts on FPGA-implemented quantum simulation modules, achieving nanosecond-level response times.
Heterogeneous Computing and System Implementation
WiMi has developed a hybrid quantum neural network structure (H-QNN) integrating classical convolutional neural networks (CNN) with quantum neural networks (QNN). This design features a three-stage distributed structure – convolutional feature extraction, quantum mapping, and hybrid decision-making – enabling quantum components to handle both nonlinear discrimination and information reconstruction within the feature space. The H-QNN aims to overcome limitations of traditional quantum hybrid models that often simply use quantum parts for classification alone.
The technology addresses quantum encoding challenges with a joint dimensionality reduction scheme utilizing angle encoding and principal component analysis (PCA). Optimizing PCA’s cumulative variance contribution rate ensures high information fidelity during the mapping of input features to quantum state amplitudes, maximizing quantum information utilization. WiMi also employs a transfer learning mechanism and parameter sharing structure to mitigate risks of gradient vanishing and overfitting common in quantum neural network training.
System implementation leverages a heterogeneous computing architecture, running classical computations on CPU/GPU platforms and quantum parts on quantum simulation modules implemented on FPGA. The FPGA module provides nanosecond-level response times for quantum state updates, significantly boosting overall training speed. This hybrid approach demonstrates performance exceeding pure CPU or GPU simulations, paving the way for practical applications of quantum intelligence in areas like computer vision and edge computing.
