WiMi Releases Next-Generation Quantum Neural Network Feature Mapping Technology: Repeated Amplitude Encoding Significantly Enhances Expressive Power of Quantum Models
From a technical background perspective, existing mainstream quantum neural networks generally rely on parameterized quantum gates to encode input data during the feature mapping stage. These quantum gates are mathematically linear or unitary transformations in essence, and the feature mappings formed by their combinations are often limited by circuit depth, the number of qubits, and the scale of trainable parameters. Although the quantum state itself resides in an exponentially high-dimensional space, in practical models, the limited encoding methods make it difficult to fully unleash this high-dimensional advantage, resulting in issues such as insufficient mapping capability and weak category scalability in complex classification tasks.
To address the above bottlenecks, WiMi, starting from the fundamental mechanism of quantum state representation, re-examined the way classical data enters the quantum system. The traditional amplitude encoding method typically maps a set of normalized classical feature vectors into the probability amplitudes of a single quantum state. Its advantage lies in high qubit usage efficiency, but the disadvantage is that the feature distribution after a single encoding is easily diluted by linear operations during the evolution of the quantum circuit, thereby limiting the ability of subsequent quantum neural networks to model complex nonlinear structures.
To verify the effectiveness of this technology in real tasks, WiMi used the classic image classification benchmark dataset MNIST as the experimental platform and conducted a systematic evaluation of the repeated amplitude encoding method. In the experiments, researchers embedded this method into various typical quantum neural network architectures and compared it with mainstream data loading methods such as traditional amplitude encoding and angle encoding.
The experimental results show that, under the condition of a fixed number of classes, quantum neural networks adopting repeated amplitude encoding outperform the control methods in classification accuracy, convergence stability, and robustness to parameter initialization. This indicates that, even under the same task complexity, repeated amplitude encoding can provide the model with more discriminative feature representations.
About
Translation Disclaimer
The original version of this announcement is the officially authorized and only legally binding version. If there are any inconsistencies or differences in meaning between the Chinese translation and the original version, the original version shall prevail.
View original content:SOURCE
