ConQUESTA: Topology-Aware Post-Training Structural Pruning of Quantum Neural Networks

What it claims

ConQUESTA prunes two-qubit entanglers from an already trained quantum neural network (QNN) by searching whole pruning masks, transpiled together, instead of ranking gates one at a time. The authors argue that removing a set of entanglers changes placement, SWAP insertion and depth in ways that summing single-gate savings does not predict. At 55% pruning, the search beats the best simple baseline by 3.46 pp (Breast Cancer), 1.74 pp (Digits01) and 8.89 pp (Iris) on a combined resource score. The authors claim better compiled efficiency at comparable accuracy, not higher accuracy.

How it works

1. Screening. Each trainable entangler gets a first-order task saliency (expected absolute gradient of the loss). It also gets a leave-one-gate-out compiler cost: 0.60 × change in native 2Q count, 0.25 × change in routed SWAPs, and 0.15 × change in depth. Negative deltas are clipped to zero. The cost takes the median over transpiler seeds, then the 75th percentile over topologies. The priority is cost × (1 − saliency), after normalisation. A diversified candidate pool is drawn from several views and capped at K = 20.
2. Mask search. NSGA-II searches fixed-cardinality masks over validation accuracy and 2Q, depth and SWAP reduction. Each mask is transpiled on line, ring and heavy-hex maps at optimization level 3 with transpiler seeds 101/202/303.
3. Selection. A mask is feasible if validation accuracy drops at most 0.01 below the dense model. The selector maximises the RT1 score (the same 0.60/0.25/0.15 weights, averaged over topologies) among feasible masks. If no mask is feasible, it first minimises the accuracy violation.
4. Recovery. With the mask frozen, surviving parameters get 8 fine-tuning epochs at learning rate 0.008. The test split is used only after this step.

An eight-seed comparison with Beam search shows no significant difference (p = .56 on accuracy, p = .22 on resource score). The authors treat NSGA-II as an implementation choice, not a contribution.

Evidence

The benchmarks are Iris (4 qubits, 12 entanglers), Breast Cancer (6 qubits, 30 entanglers) and Digits 0-vs-1 (8 qubits, skip-CRY). Each runs with 8 model seeds at 35% and 55% pruning. All baselines (random, magnitude, task-only, compiler-only, one-shot) share the same checkpoint, cardinality and recovery budget.

• Joint search versus one-shot ranking on the same priority score gives paired median RT1 gains of 4.88 pp (BC), 2.74 pp (Digits01) and 16.78 pp (Iris). The p-values are .008 to .023.
• Accuracy. Search does not win on accuracy everywhere. On Digits01, magnitude and task-only pruning match it. On Iris, compiler-only is 0.42 pp higher. Test-accuracy deltas against the dense model at 55% are +0.88 (BC), −0.52 (Digits01) and −1.25 pp (Iris).
• Recovery feasibility is 39/48 to 45/48 runs across both ratios. The two remaining Digits01 failures miss the threshold by 0.39 pp.
• Weight sensitivity. Uniform, depth-oriented and routing-oriented weights shift depth or SWAP savings by up to 3.82 pp. Mean validation accuracy moves by at most 0.44 pp.
• Scaling. In a fixed k = 4 synthetic probe from 4 to 12 qubits, the mask space grows from 495 to 12.1M while evaluated masks grow from 189 to 305. Search time rises from 265 s to 3254 s, and peak RSS from 0.79 to 1.55 GB.

How to use it

The page gives no code, repository or licence. Search is a one-time offline cost: about 4.2, 46.1 and 22.2 min (Iris, BC, Digits01), against dense training of 0.12, 13.2 and 8.7 min.

Limits

• Per-mask compile time grows from 1.40 s to 10.66 s with qubit count, which makes compilation the next bottleneck. The compile cache hit rate is zero.
• The scaling result holds only for fixed k. At a fixed 55% ratio, k would exceed the K = 20 pool at 56 and 132 entanglers, and this case is not tested.
• The grid-topology transfer and noise tests use only three seeds. Iris loses 2.2 to 3.3 pp under depolarizing noise.
• The authors present the method as resource reduction, not error mitigation.
• With n = 8, the authors call the statistics supporting, not confirmatory.