Stop Counting Qubits, Count Energy Levels
Quantum machine learning has an awkward reality: theoretical advantages look great on paper but shrink on noisy experimental hardware. The problem lies in the building blocks. Current schemes almost universally use two-level qubits as neurons—each particle uses only two energy levels, so scaling the network means adding particles. More particles mean more control complexity and more accumulating errors. According to a November 1 announcement from USTC, the joint team of Zhaokai Li and Yiheng Lin, publishing in *National Science Review* (NSR), has for the first time experimentally realized a quantum neural network based on high-dimensional quantum neurons. They used all five energy levels of a single calcium ion, achieving 95.7% classification accuracy on 395 test samples from a fashion image dataset.
Scaling Inward Instead of Outward
Two-level schemes have only one path to a larger computational space: add particles. 10 qubits are not enough? Use 20. 20 are not enough? Use 50. Each added particle brings more control overhead and another channel for error accumulation—precious resources on NISQ devices.
The USTC team turned inward. Quantum carriers typically have a series of usable energy levels; conventional two-level encoding only touches a small subset, leaving the rest idle. By incorporating more levels into a single quantum neuron—without adding any physical particle—the usable state space doubles and doubles again. One five-level calcium ion carries as much information dimensionality as multiple qubits in the old scheme.
In one sentence: the old approach keeps adding rooms; the new one partitions more rooms on the same floor.
Energy Levels as a Training Resource, Not Just Storage
The second contribution upgrades high dimensionality from "storage capacity" to a "computational resource." The team built a quantum-classical co-training framework for multi-level quantum neurons: the quantum side handles concise system control and low-complexity readout, while the classical side tunes parameters. During training, the extra energy levels are controllable, trainable, and directly involved in the machine learning task.
This clarifies something important: the value of high-dimensional encoding is not just "storing more" but "actually using it." A multi-level system has, for the first time, completed the functional leap from high-dimensional information representation to trainable intelligent computation.
Why This Matters at the Hardware Layer
A bottleneck quantum AI cannot avoid when scaling is hardware resources. If error rates cannot be pushed down, networks cannot spread out, and quantum advantage stays at the demonstration stage. The team's answer: fully exploit the internal resources of the particles at hand before adding more.
This is good news for the trapped-ion route. On the same calcium-ion platform, last week another USTC team (Yiheng Lin and Changling Zou, October 23, *Nature Physics*) used four levels of a single ion for autonomous quantum error correction; now a five-level ion runs a quantum neural network. Multi-level encoding is becoming a general-purpose template for trapped ions—used for error correction and machine learning alike.
The sobering caveats: a 395-sample test set is small, 95.7% is an experimental figure, and practical quantum machine learning still faces three hurdles—noise, scale, and algorithm maturity. The real significance is validating a route: on a finite physical carrier, information-processing capacity can indeed be dug deeper, not only spread wider.
For those tracking the quantum computing industry, the signal is that the spec sheet for trapped-ion hardware may need revising: beyond qubit-count metrics, "usable energy levels per particle" is becoming a parameter worth listing separately—it determines how many dimensions a single chip can yield.
---
Sources and limitations: USTC news release, November 1 (via Anhui Net, reporter Wei Xinxin); the study was published online in *National Science Review* (NSR) by the joint team of Zhaokai Li and Yiheng Lin. The 95.7% accuracy on 395 fashion-image test samples, the five-level calcium-ion processor, and the quantum-classical co-training framework all come from official statements. Results have not been independently replicated. Note: the quantum error correction work (Lin and Zou, *Nature Physics*, October 23) and this quantum neural network work come from different teams on the same platform—do not conflate them.