BEIJING, Oct. 20, 2025 (GLOBE NEWSWIRE) -- BEIJING, Oct. 20, 2025––WiMi Hologram Cloud Inc. (
NASDAQ: WiMi) ("WiMi" or the "Company"), a leading global Hologram Augmented Reality ("AR") Technology provider, today announced the development of single-qubit quantum neural network technology for multi-task design. This technology has extremely disruptive significance; this technology, by demonstrating the feasibility of high-dimensional quantum systems in efficient learning, provides a realistic path for the deep integration of future quantum computing and artificial intelligence.Nowadays, training large neural networks often requires billions of parameters and massive data center resources, and the sharp rise in power consumption and hardware costs has become a real bottleneck in the development of artificial intelligence. At the same time, although traditional neural networks have achieved high accuracy in multi-class classification problems, as the number of categories increases, the model structure also expands accordingly, leading to a decline in inference latency and computational efficiency.The rise of quantum computing provides new possibilities for this dilemma. Quantum bits (qubits) and quantum multi-level systems (qudits) can utilize superposition and entanglement to achieve natural representation of high-dimensional data spaces, thereby breaking the resource limitations of classical computing. In this field, Quantum Neural Networks (QNN) have become a frontier direction of research. Compared to traditional deep learning, QNN can achieve complex mappings through shallow quantum circuits, greatly improving model compactness and computational efficiency.In the wave of quantum machine learning, the single-qudit quantum neural network technology proposed by WiMi not only meets the actual needs of high-dimensional data classification but also breaks through the implementation bottlenecks under the constraints of quantum hardware, becoming an important step in promoting industrial progress.The core idea of the single-qudit quantum neural network technology proposed by WiMi is to use the state space of a single high-dimensional qudit to directly handle multi-class classification tasks. Unlike classical neural networks that rely on thousands of neurons and complex hierarchical structures, SQ-QNN leverages the high-dimensional characteristics of quantum systems to efficiently encode and distinguish category information within a compact circuit scale.In this design, each category corresponds to one dimension of the quantum system, and the overall classification process is completed through the action of a high-dimensional unitary operator. WiMi uses the Cayley transform of skew-symmetric matrices to construct the unitary operator; this method not only possesses good mathematical stability but also ensures efficiency in quantum circuit implementation. In this way, the evolution of the quantum state directly establishes a mapping relationship with the category labels, greatly reducing the circuit depth and training overhead.Additionally, this technology introduces a hybrid training method when optimizing network parameters. It combines extended activation functions with the optimization framework of Support Vector Machines (SVM). The extended activation function originates from the truncated multivariate Taylor series expansion and can effectively introduce nonlinear representational capabilities in the quantum state space, while SVM optimization further ensures the stability of parameter optimization and the acquisition of global optimal solutions.The entire technical logic of WiMi's SQ-QNN can be divided into three levels: quantum state encoding, unitary evolution design, and hybrid training optimization.First is the quantum state encoding. In multi-class classification problems, assuming the number of categories is $d$, a $d$-dimensional qudit system is constructed to carry the data. After appropriate data preprocessing, the input samples are mapped to the amplitude or phase information of the quantum state. In this process, traditional feature extraction steps are greatly simplified, allowing data to directly enter the neural network in quantum form.Second is the unitary evolution design. WiMi proposes using the Cayley transform of skew-symmetric matrices to generate $d$-dimensional unitary operators. The properties of skew-symmetric matrices make their Cayley transform results naturally satisfy unitarity, thereby ensuring the physical rationality and implementability of quantum state evolution. Through this unitary operator, the input quantum state completes the mapping and differentiation of category information in the high-dimensional Hilbert space. Unlike the multi-layer propagation in classical neural networks, this scheme can achieve complex decision boundaries through a single-step evolution, significantly reducing the circuit depth.Finally, it is the hybrid training optimization. In the parameter training phase, this scheme does not solely rely on quantum computing but adopts a hybrid quantum-classical training method. The introduction of extended activation functions enables the quantum neural network to possess nonlinear classification capabilities while maintaining a shallow structure. At the same time, the support vector machine optimization mechanism provides an efficient path for parameter search, allowing the network to quickly converge to the global optimal solution. Under this training framework, the burden on quantum hardware is effectively shared, and training efficiency is significantly improved.
