New feature map boosts quantum neural network accuracy by up to 44%

New Feature Map Improves Quantum Neural Network Accuracy

Quantum neural networks (QNNs) have long been hampered by a subtle but consequential limitation of amplitude encoding: the encoding process normalizes input vectors, making any non‑zero real scaling of the data indistinguishable. In practice, this means that inputs such as `2.0` and `4.0` are mapped to the same quantum state, preventing the network from learning functions that depend on absolute magnitude.

Vincent Gurgul (Humboldt‑Universität zu Berlin) and David Prinz (Rheinische Friedrich‑Wilhelms‑Universität Bonn) propose a deterministic preprocessing step—a fixed nonlinear feature map, denoted \(\Phi\)—that resolves this ambiguity before the data is encoded. The map expands the input into a higher‑dimensional space where scaled versions become orthogonal, yet it adds no trainable parameters and retains the logarithmic qubit scaling that makes amplitude encoding attractive for near‑term quantum hardware.

---

Why Scale Invariance Matters

Amplitude encoding represents a `d`‑dimensional feature vector using only \(\log_2 d\) qubits by normalizing the vector to unit length. While efficient, this normalization discards the radial information (the overall magnitude) of the data. Consequently, any downstream variational circuit—no matter how deep or expressive—cannot recover that lost information. The result is a representational bottleneck that limits the class of functions a QNN can approximate, especially those that rely on absolute scale.

---

The Deterministic Feature Map \(\Phi\)

The proposed map applies a simple, deterministic nonlinear transformation to each input component, effectively embedding the original vector into a space where the norm is preserved. Because the transformation is fixed, it does not increase the number of trainable parameters, and the overall qubit requirement grows only by a constant factor (the dimensionality triples, but the qubit count still scales as \(O(\log_2 d)\)).

Key properties of \(\Phi\):

• Scale discrimination: Scaled inputs produce distinct quantum states, eliminating the global‑phase equivalence that plagued conventional amplitude encoding.
• Parameter‑free: No additional optimization is required, simplifying the training pipeline.
• Resource‑efficient: Maintains the logarithmic qubit advantage over angle encoding, which typically needs one qubit per feature.

---

Empirical Gains on Benchmark Datasets

The authors evaluated the new encoding on three standard classification benchmarks (including the Breast Cancer and Wine datasets) and a synthetic nonlinear dataset. Across all tasks, the QNN equipped with \(\Phi\) outperformed both conventional amplitude encoding and angle encoding:

• Balanced accuracy improvements ranged from 21 % to 44 %.
• On the synthetic dataset, the new method achieved a balanced accuracy of 0.7717, compared to 0.5996 (amplitude) and 0.6458 (angle).
• For the Wine dataset, the QNN reached 0.9320 accuracy, surpassing the 0.6816 (angle) and 0.7600 (amplitude) baselines.

These results demonstrate that the bottleneck is not a lack of quantum circuit depth but the loss of scale information during encoding. By restoring that information, the QNN can learn richer decision boundaries, narrowing the performance gap with classical neural networks while still operating with far fewer qubits.

---

Implications for Quantum Machine Learning

The work highlights two broader lessons for the field:

1. Encoding choices are as critical as circuit design. Even the most sophisticated variational ansätze cannot compensate for information lost at the data‑loading stage.
2. Clever, deterministic preprocessing can unlock expressive power without sacrificing hardware efficiency. The feature map shows that modest classical computation before quantum processing can yield substantial quantum performance gains.

Future research may explore extensions of \(\Phi\) to other data modalities, integration with hybrid classical‑quantum pipelines, and theoretical analyses of how different preprocessing strategies affect the expressive capacity of QNNs.

---

Practical Takeaway

For practitioners building quantum machine learning models on near‑term devices, adopting a deterministic, parameter‑free feature map like \(\Phi\) offers a low‑overhead route to improve accuracy. It preserves the qubit‑efficiency of amplitude encoding while eliminating a fundamental representational limitation, making quantum neural networks more viable for real‑world classification tasks.