Beijing-based, Nasdaq-listed WiMi Hologram Cloud Inc., which focuses on holographic cloud services, has launched a hybrid quantum neural network structure (H-QNN) for image multi-classification. The company says this marks a key step in moving quantum artificial intelligence research from theory to practical applications.
WiMi’s H-QNN integrates the spatial feature extraction capabilities of classical convolutional neural networks (CNN) with the high-dimensional nonlinear mapping features of quantum neural networks (QNN). According to the company, this forms a new type of hybrid structure that possesses stronger generalization ability and computational efficiency in multi-class classification scenarios, which will lay a solid technical foundation for quantum intelligent vision systems.
The overall system consists of three main modules: a Feature Dimensionality Reduction and Encoding Module, a Quantum State Transformation Module, and a hybrid Decision and Transfer Learning Module.
First, the Feature Dimensionality Reduction and Encoding Module, which is based on the classical convolutional neural network (CNN) structure, extracts low-dimensional feature representations of images through several convolutional layers and pooling layers. The feature vectors after PCA dimensionality reduction are standardized and then input into the quantum encoding circuit. At this stage, Angle Embedding is used to map real-valued features to quantum state amplitudes, achieving efficient encoding through multi-layer quantum rotation gates (Ry, Rz) and thereby reducing quantum gate depth and lowering encoding noise.
Next, the Quantum State Transformation Module undertakes the core tasks of high-dimensional feature mapping and nonlinear discrimination. This module includes several layers of quantum circuits, with each layer composed of parameterized rotation gates and controlled entanglement gates (CNOT or CZ), forming nonlinear coupling and entanglement of quantum states. To alleviate gradient vanishing, WiMi has adopted a reconfigurable parameter sharing strategy, allowing different quantum layers to share some trainable parameters, while introducing mixed state perturbations to maintain gradient balance during the training process. This structural design effectively avoids the barren plateau phenomenon, enabling the model to maintain stable convergence in multi-class tasks.
Finally, the Hybrid Decision and Transfer Learning Module integrates the results of quantum computing with the classical decision layer. The measurement probability distribution output by the quantum circuit is converted into feature vectors and fused with the output of the classical fully connected layer. This fused vector is input into the Softmax layer for final classification judgment. To further enhance the generalization performance in multi-class tasks, WiMi has introduced a transfer learning mechanism, migrating the parameters of quantum layers pre-trained in small-sample tasks to new tasks, thereby reducing the number of training epochs and enhancing model stability.
In actual implementation, this structure supports running on simulation environments and hardware quantum processing units (QPUs). The simulation environment uses high-performance GPU clusters to complete training of classical modules, while quantum modules are executed in quantum simulators or FPGA-accelerated quantum kernel estimation environments, achieving heterogeneous collaboration of classical and quantum computing resources.
According to WiMi, this technology is innovative in the following ways:
1. At the architectural design level, it achieves deep integration of convolutional neural networks (CNN) and quantum neural networks (QNN). Whereas traditional quantum hybrid models usually simply embed the quantum part as a classification head, the H-QNN proposed in this research adopts a three-stage distributed structure of “convolutional feature extraction—quantum mapping—hybrid decision-making”, enabling the quantum part not only to undertake nonlinear discrimination but also to achieve information reconstruction at the feature space level.
2. At the encoding strategy level, the joint dimensionality reduction scheme of angle encoding and principal component analysis (PCA) effectively solves the quantum encoding dimension limitation problem. By optimizing the cumulative variance contribution rate of PCA, it ensures that the mapping between input features and quantum state amplitudes maintains high information fidelity, thereby maximizing the utilization rate of quantum information.
3. At the training strategy level, a transfer learning mechanism and parameter sharing structure is introduced, establishing balanced gradient flow between different quantum layers. In addition, WiMi says it has designed an early stopping strategy based on the quantum Fidelity metric, which determines whether the training has reached the optimal point by monitoring the stability of quantum state evolution, thereby preventing overfitting.
4. Finally, at the system implementation level, it adopts a heterogeneous computing architecture, running the classical computing part on CPU/GPU platforms, while the quantum part is executed in quantum simulation modules implemented on FPGA. The FPGA module realizes reconfigurable execution logic for parameterized quantum circuits, capable of completing quantum state updates within nanosecond-level response times, thereby significantly improving the overall training speed of the system.
