HQNN vs Classical Neural NetworkThree Production Benchmarks, Compared Side by Side
Compare a hybrid quantum neural network against classical MLP and CNN baselines on trainable parameters, training time, and accuracy — using real numbers from three in-house benchmarks.
What Is a Hybrid Quantum Neural Network?
Classical Layers, a Quantum Circuit, and a Classical Read-Out
A classical neural network — an MLP, a CNN, or an LSTM — learns purely from stacked linear layers, nonlinearities, and (for CNN/LSTM) convolution or recurrence, trained by backpropagation. A hybrid quantum neural network (HQNN) keeps that architecture intact and replaces one internal layer with a parameterized quantum circuit (PQC): classical layers feed a compact feature vector into the circuit, its measurement outputs feed a classical read-out layer, and the whole stack — classical weights and quantum gate angles alike — trains end to end with one optimizer. It is a classical network with one layer swapped out, not a quantum replacement for it.
Every case on this page follows that pattern: an LSTM or linear front end carries most of the representation learning, and the quantum layer contributes one narrow, well-defined transformation.
What the Quantum Layer Actually Contributes
The quantum layer does two things a classical layer doesn't. Angle encoding maps the feature vector into a quantum state by using each feature to set a qubit's rotation angle — embedding the data into an exponentially large Hilbert space without adding classical parameters. Trainable rotations interleaved with entangling gates (CNOT or CZ) then couple qubits together, spreading local feature information across multi-qubit correlations that a classical layer of the same width can't represent directly. The circuit is measured — typically Pauli-Y or Pauli-Z — and those expectation values become the classical read-out layer's input.
This is why HQNNs are described as expressive for their parameter count: a shallow circuit on a handful of qubits can, in principle, represent a feature map that would need far more classical weights to approximate. Whether that turns into a measurable gain on a given dataset is exactly what the tables below test.
Training Realities: Gradients, Shots, and Barren Plateaus
Training a quantum layer isn't free. Gradients are estimated with methods like the parameter-shift rule, the adjoint method, or SPSA rather than exact autodiff, and on real hardware each estimate needs repeated circuit runs — typically 1,000 to 50,000 shots — to average out measurement noise. More shots cut variance at the cost of runtime, one reason every case here trains on a simulator rather than a physical device.
The deeper issue is the barren plateau: as depth and qubit count grow, the achievable gradient shrinks, sometimes exponentially, leaving a loss surface too flat to train. Hardware noise makes it worse — a "noise-induced barren plateau" — and better per-layer fidelity only pushes the onset later, without removing it. That's why every HQNN here stays shallow: one to four layers across 5 to 12 qubits.
All of it runs on a noiseless simulator, capped at 16 qubits. Real hardware today offers roughly 127–433 qubits, with a practical industrial advantage generally expected to need over 1,000; 96–99% gate fidelities still accumulate error; coherence times of 100–200 microseconds cap usable depth; and cloud access runs an estimated $0.30–$1.60 per second, plus queue delays. None of that changes the numbers below, but it's why a hybrid split — mostly classical, one quantum layer — is the practical shape for anything meant to run today.
Why compare them
- A classical MLP, CNN, or LSTM is a familiar baseline with mature tooling and no dependency on a simulator or quantum device.
- An HQNN keeps that same backbone and swaps in a small parameterized quantum circuit for one internal layer — the rest of the training loop is unchanged.
- Running both on the same split and metric is the only way to tell whether the quantum layer contributes anything, rather than just adding engineering surface area.
Limits to keep explicit
- Quality gains from the quantum layer are not guaranteed, and below they are real but modest — single-digit percentage points, not order-of-magnitude wins.
- Every result here comes from a noiseless simulator capped at 16 qubits; real hardware adds gate noise, limited coherence time, and a per-second cost.
- Barren plateaus get worse as qubit count and depth grow, which is why every case here uses a shallow, compact circuit rather than a deep one.
HQNN vs Classical Neural Network, Side by Side
Seven dimensions that decide whether a quantum layer is worth adding, drawn from the benchmarks below and the platform's HQNN algorithm page. For the same comparison built around a quantum-kernel classifier instead, see QSVM vs classical SVM.
| Dimension | HQNN | Classical Neural Network |
|---|---|---|
| Trainable parameters | Often fewer: 320,882 (wind) and 11,300 (predictive maintenance). | Larger backbone typically needed to match — 481,013 in the wind MLP baseline. |
| Training time | 2.5–4 min (wind), 30 min (predictive maintenance), 58 sec (hydraulic). | Comparable in the source reporting; standard backprop, no circuit-simulation overhead. |
| Data requirements | Same preprocessed inputs as the backbone; the quantum layer adds an encoding step, not new data. | Whatever the architecture normally needs, trained on the same split. |
| Inference cost | Slower in 2 of 3 cases (10s vs 0.05s; 0.03s vs 0.02s); faster in the third (5s vs 70s CNN). | Fast on ordinary hardware in most cases, though CNN/LSTM add their own cost. |
| Hardware today | Noiseless simulator only, capped at 16 qubits; real devices offer ~127–433 qubits at $0.30–$1.60/sec. | Any GPU or CPU — no specialized access or per-second cost. |
| Noise behaviour | Noise-induced barren plateaus shrink gradients as qubits/depth grow; better fidelity pushes the onset later, not away. | No quantum-noise failure mode, though deep/recurrent nets have their own vanishing-gradient issues. |
| Where each wins | Parameter efficiency in 2 of 3 cases; faster inference than a heavier CNN in the third. | Simplicity, faster training/inference in most cases, no quantum access dependency. |
Three Worked Comparisons: HQNN vs Classical Baselines
Run in full on the platform: wind-power forecasting and predictive maintenance (which also covers hydraulic condition monitoring as a second case).
| Case | Dataset | HQNN result | Classical baseline | Parameters | Training time |
|---|---|---|---|---|---|
| Wind-power forecasting | 26,304 hourly readings, 7 features | NMAE 7.11%, MSE 0.00789 (9 qubits) | NMAE 7.30%, MSE 0.00809 (MLP) | 320,882 vs 481,013 | 2.5–4 min; inference 0.03s vs 0.02s |
| Predictive maintenance | 10,000 telemetry rows, 5 features | 94% accuracy, ROC-AUC 0.983 (5 qubits) | 93% accuracy, ROC-AUC 0.981 (MLP) | 11,300 (HQNN); MLP count not reported | 30 min; inference 10s vs 0.05s |
| Hydraulic condition monitoring | 2,205 cycles, 11 dims × 60 steps | 100% accuracy (12 qubits) | 99.6% accuracy (CNN) | 852,000 (HQNN); CNN count not reported | 58 sec; inference 5s vs 70s |
The pattern across all three rows: HQNN matches or slightly beats its classical counterpart, by roughly one percentage point of accuracy or a similar fraction of forecasting error — not an order of magnitude. Wind forecasting is the cleanest efficiency story, reaching lower error with about a third fewer parameters. Predictive maintenance shows the same shape at smaller scale, though its classical baseline isn't parameter-counted in the source. Hydraulic monitoring is the outlier on inference cost: HQNN beats a heavier CNN there (5s vs 70s), the reverse of the other two cases, because that CNN does more work per prediction over a 60-step sequence.
None of the source reports claims a wholesale win — each frames the gain as real but conditional, worth checking case by case, especially where a compact parameter count matters or a classical model has already plateaued.
When to Choose HQNN vs. When to Choose a Classical Network
None of the three benchmarks below show a one-sided win, so the decision comes down to what the project actually needs.
Choose a classical network when
- Fast iteration and mature tooling matter more than parameter count — classical baselines here trained or ran inference in comparable or less time in two of three cases.
- No team or budget yet exists for simulator or quantum-hardware access, or the timeline doesn't allow for it.
- The problem doesn't obviously need a high-dimensional feature map — most tabular and time-series problems don't, and a well-tuned MLP, CNN, or LSTM is the simpler answer.
Choose an HQNN when
- A compact parameter count matters — edge or memory-constrained deployment, as in wind forecasting's 320,882 versus 481,013 parameters.
- A strong classical baseline has plateaued and a differently-shaped feature map is worth testing before further tuning.
- The goal includes benchmarking quantum-hardware readiness now, ahead of workloads that may later move from simulator to device.
In the three cases benchmarked here, the difference was real but modest in both directions — worth deciding project by project, not by default toward either architecture.
Business Impact
The value estimates below come from the three source reports, not from this comparison itself — they describe what each production use case's report projects for a deployed model.
$200K/unit
Predictive-maintenance HQNN annual savings
$70K–85K/unit
Hydraulic condition-monitoring annual savings
$50K–300K/yr
Wind-forecasting O&M optimization value
3
In-house HQNN benchmarks behind these numbers
ML and quantum benchmark teams
For teams that need the same data split, preprocessing, and evaluation metric across a classical backbone and its HQNN variant.
Applied quantum researchers
For researchers testing whether a parameterized quantum layer changes accuracy, error, or parameter efficiency on their own data.
Product and platform leads
For teams deciding whether an HQNN result is strong and stable enough to justify moving from a simulator toward real hardware.
How it works
One dataset, two architectures, one reproducible comparison.
Prepare Data
Use the same split, scaling, and features for both the classical backbone and the HQNN's front end
Run the Classical Baseline
Train the MLP, CNN, or LSTM baseline and record parameters, training time, and accuracy
Build the HQNN
Wrap the same backbone around a compact parameterized circuit — angle encoding plus entangling layers
Compare
Review trainable parameters, training time, accuracy or error, and inference cost side by side
Export
Download the method, architecture details, executable code, and reproducible benchmark report
Related pages
Algorithms, applied use cases and benchmarks connected to this page
Algorithms
Quantum Neural Networks: QNN & HQNN
Understand quantum neural networks (QNN) and hybrid QNN (HQNN): how each works, when to use which, real benchmarks, and NISQ hardware requirements.
Manufacturing
Predictive Maintenance and Condition Monitoring with HQNN
Predictive maintenance and hydraulic condition monitoring with HQNN, real accuracy vs. classical baselines, and executable code.
Energy
Wind Energy Forecasting with HQNN
Forecast short-term wind power with a Hybrid Quantum Neural Network, CAISO/NREL weather data, and a benchmarked classical MLP comparison.
Comparisons
QSVM vs Classical SVM for Classification
Compare QSVM with classical SVM on a 569-sample breast-cancer benchmark with 30 FNA features, 4 qubits, 2 layers, 96% accuracy, and 0.99 ROC-AUC.
Frequently Asked Questions
Is a quantum neural network better than a classical neural network?
Not decisively, based on the three benchmarks compared here. HQNN matched or slightly beat its classical MLP or CNN counterpart — about one percentage point of accuracy, never an order of magnitude. The consistent advantage was parameter efficiency rather than raw quality: the wind-forecasting HQNN reached lower error with roughly a third fewer trainable parameters than its MLP baseline.
HQNN vs LSTM for time series: which is better for forecasting?
In wind-power forecasting, an LSTM wrapped around a compact quantum layer reached 7.11% NMAE against 7.30% for a classical MLP, with about a third fewer parameters. In hydraulic condition monitoring, the same recipe hit 100% accuracy against a CNN's 99.6%, with faster inference. Both are real but small wins — a case for testing the quantum layer alongside an LSTM, not replacing it.
How many qubits does an HQNN need?
Between 5 and 12 in these benchmarks, sized to how much the classical backbone has already compressed the input rather than to the raw feature count: 5 qubits for predictive maintenance, 9 for wind forecasting, 12 for the higher-dimensional hydraulic time series. A pure QNN over the same raw features typically needs more qubits than an HQNN whose backbone compresses them first.
Can I train an HQNN on a laptop?
Yes, at these scales. Every benchmark here ran on a classical simulator on ordinary CPU cores, in under an hour: 58 seconds (hydraulic), 2.5–4 minutes (wind), about 30 minutes (predictive maintenance). Real quantum-hardware access, by contrast, is priced at an estimated $0.30–$1.60 per second of execution.
What is a barren plateau, and does it affect every HQNN?
A barren plateau is where the training loss's gradient shrinks toward zero, making the circuit effectively untrainable. It worsens as qubit count and depth grow, and worsens further with hardware noise. It's a risk for any variational circuit, not a guaranteed failure: every case here avoids it by staying shallow — one to four layers across 5 to 12 qubits.
Does an HQNN need real quantum hardware to run?
No. Every result here ran on a noiseless classical simulator capped at 16 qubits. Real devices are an option — today's offer roughly 127–433 qubits at 96–99% gate fidelity — but the source material treats simulator results as the practical baseline and hardware as a later validation step.
Benchmark HQNN Against a Classical Neural Network
Run the same comparison on your own data: a classical MLP, CNN, or LSTM baseline against a hybrid quantum neural network, with parameters, training time, and accuracy reported side by side.
Try your first case for free