WiMi Launches H-QNN Achieving Superior MNIST Accuracy Over Classical MLP
WiMi Hologram Cloud has introduced a Hybrid Quantum-Classical Neural Network (H-QNN) for MNIST binary classification, integrating parameterized quantum circuits with classical MLP layers to map 28×28 images into high-dimensional quantum feature space. In tests distinguishing digits “0” and “1,” H-QNN outperformed equivalent-scale classical MLPs under the same epochs and sample sizes, demonstrating enhanced sensitivity and generalization.
1. Technology Overview
WiMi’s H-QNN combines a trainable quantum feature encoding module with a classical neural classifier. The pipeline begins with binarization and normalization of 28×28 MNIST images, followed by compression into blocks for quantum encoding via rotation and entanglement gates (Ry, Rz, CNOT, CZ). Measured quantum states yield feature vectors fed into a lightweight MLP, enabling hybrid gradient-based optimization across quantum and classical parameters.
2. Experimental Results
In a binary classification task distinguishing handwritten digits “0” and “1,” H-QNN achieved higher accuracy than a classical MLP of equivalent parameter scale under identical training epochs and dataset sizes. The quantum feature space mapping boosted discrimination of high-dimensional patterns, maintaining strong performance even with reduced sample sets.
3. Competitive Advantages
H-QNN addresses classical CNN and MLP limitations such as overfitting and gradient vanishing by leveraging quantum superposition and entanglement for exponential feature mapping. This hybrid architecture sidesteps current hardware qubit constraints while preserving quantum acceleration potential, positioning WiMi at the forefront of practical quantum machine learning applications.