WiMi Hologram Cloud Inc. (NASDAQ: WiMi) has developed a quantum kernel convolution (QKC) scheme designed to run on noisy intermediate-scale quantum (NISQ) devices. This is a surprising move from a company that specializes in holographic augmented reality technology. Rather than waiting for the arrival of fault-tolerant quantum computers, WiMi’s approach rethinks computationally intensive convolution processes, particularly feature extraction and dimensionality reduction, by mapping image patches to quantum states and leveraging controlled evolution of entanglement. WiMi points out that classical convolutional layers essentially rely on sliding windows and linear weighted addition to achieve local feature extraction, whereas quantum computing inherently features high-dimensional Hilbert space representations and quantum parallel processing capabilities. This hybrid quantum convolutional neural network (QCNN) integrates a quantum acceleration module for feature extraction within a classical deep learning framework, providing a practical path towards quantum-enhanced image classification.
Quantum Kernel Convolution (QKC) for NISQ Device Implementation
WiMi Hologram Cloud Inc. This is not a far-flung vision of future quantum capabilities. The QKC scheme is specifically designed to work with current noisy intermediate-scale quantum (NISQ) devices, avoiding the need for currently unavailable stability in fault-tolerant quantum computers. This focus on immediate implementation distinguishes WiMi’s approach from many other quantum machine learning initiatives. A key element of this approach is a new pooling mechanism. WiMi describes it as “an information reallocation and selection mechanism that can achieve dimensionality reduction without explicitly discarding information,” which reduces the computational load of subsequent quantum and classical circuits. At the architectural level, WiMi’s hybrid QCNN adopts a layered design and strategically integrates quantum processing and established classical deep learning techniques. While classical neural networks handle preliminary data normalization, dimension adjustment, and final classification, quantum convolutional layers act as dedicated quantum acceleration modules for feature extraction.
This synergy allows the model to benefit from a mature classical toolchain while introducing quantum advantages at key computational nodes, avoiding the scalability limitations of fully quantum models due to current hardware constraints. Built on the Qiskit framework, this implementation encapsulates quantum convolutional layers as reusable modules and seamlessly integrates them into existing deep learning workflows. WiMi details that the model adopts a hybrid optimization strategy during the training process. The classical backpropagation algorithm is used to update the parameters of the classical network, and the parameter shift rules are utilized to estimate the gradients of the quantum circuit parameters, achieving end-to-end joint training. Initial tests on the MNIST dataset showed comparable classification accuracy to traditional CNNs, but with a significantly lower number of parameters, suggesting a viable path toward practical quantum-enhanced image classification.
This hybrid model can achieve classification accuracy comparable to traditional models despite having significantly fewer parameters compared to traditional CNN models.
