Announcement • May 31
WiMi Hologram Cloud Inc. Achieves Breakthrough In Deep Convolutional Neural Network Technology Based On Quantum Parameterized Circuits
WiMi Hologram Cloud Inc. has proposed a quantum deep convolutional neural network (QDCNN) model for image recognition tasks. The model uses quantum parameterized circuits as its core computing structure, performing feature extraction through quantum convolutional layers and completing classification via a quantum classification layer. The architecture follows the hierarchical design philosophy of classical CNNs while leveraging the parallel computing capabilities of quantum circuits, enabling improved efficiency when processing high-dimensional data. The QDCNN architecture consists of four main modules: data encoding, quantum convolutional layers, quantum feature fusion, and quantum classification. The process begins by mapping classical image data into quantum state space using encoding methods such as amplitude encoding, angle encoding, or hybrid approaches. These techniques convert pixel values into the probability amplitudes of qubits, allowing quantum systems to process image data. Once encoded, the quantum convolutional layers perform feature extraction. Similar to classical convolution kernels, these layers operate on local qubits using parameterized quantum gates that act as filters. Due to quantum superposition, these operations can process multiple states simultaneously, enabling highly parallel feature extraction for complex image structures. The core of the model lies in its parameterized quantum circuits, which include rotation gates, control gates, and entanglement gates. Rotation gates adjust qubit state angles, control gates establish correlations between qubits, and entanglement gates create complex relationships across multiple qubits. Together, these gates enable feature extraction with higher expressive power compared to classical methods. As layers deepen, the network learns hierarchical features. Shallow layers capture basic features like edges and textures, while deeper layers identify more complex structures. Because quantum states remain in superposition, feature extraction occurs across an exponentially large state space, improving computational efficiency. Following feature extraction, a quantum feature fusion module integrates information from different qubits using additional quantum gate operations. Through entanglement, features from different image regions are combined into higher-dimensional representations with stronger discriminative ability. Unlike classical fusion relying on matrix multiplication, this approach integrates information through quantum state evolution, reducing computational overhead. The final stage is quantum classification, where key qubits are measured to produce probability distributions. These measurement outcomes determine the image category, similar to fully connected layers in classical networks, but performed in quantum state space to exploit parallelism. For training, WiMi employs a quantum-classical hybrid approach due to current hardware limitations. In this setup, quantum circuits handle forward computations, while classical computers perform parameter updates. During training, data is encoded, processed through the quantum circuit, and measurement results are analyzed to compute error. Classical optimization algorithms then calculate gradients and update circuit parameters, which are reloaded for subsequent iterations.
This hybrid method is based on variational quantum algorithms (VQAs), which combine parameterized quantum circuits with classical optimization to enable learning under limited quantum resources. Applying VQA principles ensures the feasibility of training the QDCNN model. From a computational standpoint, the model offers potential advantages over classical CNNs. While classical models scale polynomially with data size, the QDCNN leverages quantum superposition and parallelism to process large state spaces simultaneously, theoretically enabling exponential speedups for certain tasks. Experimentally, WiMi validated the model on quantum simulation platforms. Results show that the QDCNN can effectively learn image features and achieve stable classification performance. Although current experiments are limited by qubit availability, they demonstrate the feasibility of the approach.
Additionally, WiMi has developed a supporting software and algorithm framework that includes quantum circuit construction, data encoding, training optimization, and model evaluation. This framework allows the model to run on quantum simulators and early-stage quantum hardware.
