WiMi Hologram Cloud has released a new feature mapping technology for quantum neural networks called Repeated Amplitude Encoding (RAE). The method aims to increase the representational capacity of quantum models by encoding the same classical data multiple times across separate qubit blocks, rather than relying on a single encoding pass.
The problem with current encoding
Existing quantum neural networks typically encode input data using parameterized quantum gates. These gates perform linear or unitary transformations, and the resulting feature maps are constrained by circuit depth, qubit count, and the number of trainable parameters. Although quantum states theoretically occupy an exponentially high-dimensional space, practical encoding methods often fail to exploit this advantage, leading to poor mapping capability and weak category scalability in complex classification tasks.
How Repeated Amplitude Encoding works
Traditional amplitude encoding maps a normalized classical feature vector into the probability amplitudes of a single quantum state. This approach is efficient in terms of qubit usage, but the feature distribution after a single encoding can be diluted by linear operations during circuit evolution, limiting the model's ability to capture complex nonlinear structures.
RAE addresses this by repeatedly encoding the same set of classical data across multiple qubit blocks. This repetition preserves more discriminative information throughout the quantum circuit, allowing the model to maintain higher expressive power while keeping resource usage controllable.
Experimental results
WiMi tested RAE on the MNIST image classification benchmark. Researchers embedded the method into several typical quantum neural network architectures and compared it against standard amplitude encoding and angle encoding. Under a fixed number of classes, models using RAE outperformed the control methods in classification accuracy, convergence stability, and robustness to parameter initialization.
Tradeoffs
The primary tradeoff is increased qubit usage: RAE requires multiple qubit blocks to encode the same data repeatedly, which raises the total qubit count compared to single-pass encoding. However, WiMi claims the method maintains controllable resource usage overall. The company has not disclosed specific qubit overhead numbers or circuit depth comparisons.
When to use it
RAE is relevant for quantum machine learning tasks where classification accuracy and model expressiveness are critical, particularly with high-dimensional data. It is not a general-purpose quantum computing improvement but a targeted technique for the feature mapping stage of quantum neural networks.
Bottom line
WiMi's Repeated Amplitude Encoding offers a practical engineering approach to improving quantum neural network performance without requiring fundamentally new hardware. The MNIST results suggest it can provide more discriminative feature representations under the same task complexity. Whether this translates to larger, real-world datasets remains to be seen.