QML-SAKURA — Quantum-Classical Hybrid Machine Learning

Challenge

VQCs are expressive but suffer exponential gradient vanishing under random initialization (the barren-plateau problem), and a clean advantage over classical NNs has not been demonstrated on real-world data. NISQ noise complicates real-hardware evaluation, so research claims must be cleanly separated from operational viability.

Solution

Pennylane + PyTorch makes the VQC a differentiable module; block-identity initialization, local cost functions, and shallow ansätze are combined to suppress barren plateaus. Gradient variance is monitored mid-training, and the runtime auto-falls-back to a classical baseline when the variance dips below threshold.

Results

• On a 4-class, 1k-sample MNIST subset, hybrid VQC matches classical MLP accuracy (in classical simulation)
• Block-identity initialization improves early gradient variance by 3×+ (simulation)
• Barren-plateau detector cuts wasted training steps by 18% on average
• We make no quantum-hardware advantage claim — output is shared internally as a NISQ-era applicability guide