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

AI summary

Overview

Research area: Quantum machine learning and quantum compiler-aware optimization — specifically, post-training structural pruning of parameterized quantum circuits / quantum neural networks (QNNs) for deployment on hardware with sparse qubit connectivity.

Technical level: Intermediate. Readers should be comfortable with variational quantum circuits, two-qubit entanglers, transpilation (logical-to-physical mapping, SWAP insertion, circuit depth), and basic pruning/saliency concepts. No deep optimization theory is required.

Scope (one sentence): The paper introduces ConQUESTA, an accuracy-constrained, compiler-in-the-loop mask search that decides which trained entanglers in an already-trained QNN to delete so that transpiled two-qubit gate count, SWAP count, and depth all drop, evaluated on three small benchmarks plus a 4–12 qubit scaling probe.

What This Paper Is About

A trained quantum neural network cannot simply be dropped onto real quantum hardware: physical devices have sparse coupling graphs, so the compiler must map logical qubits onto physical ones and insert routing (SWAP) operations, which adds two-qubit gates and depth. Deciding which entanglers to remove before deployment is hard because the savings from removing several gates are not the sum of their individual savings — removing a set of edges can change qubit placement, SWAP insertion, and depth in non-additive ways.

The paper's goal is a post-training pruning method that preserves task accuracy while reducing compiler-induced cost. ConQUESTA treats pruning as a constrained search over complete pruning masks, where each candidate mask is transpiled as a whole across multiple coupling topologies, rather than scoring gates one at a time.

Key Contributions

1. Compiler-aware post-training mask search. A screening score combines first-order task saliency with a leave-one-gate-out physical cost (weighted two-qubit, SWAP, and depth deltas), followed by evaluation of complete fixed-cardinality masks after multi-topology transpilation instead of summing local gate scores.

2. Matched eight-seed evidence at 55% pruning. Search improves mean resource score R_T1 over the strongest matched simple baseline by 3.5 / 1.7 / 8.9 percentage points on Breast Cancer / Digits01 / Iris, and direct fixed-mask recovery raises feasibility from 16/24 to 22/24 runs.

3. Deployment-policy and scaling evidence. Re-weighting the same stored masks changes depth and routing priorities with at most a 0.44 pp mean validation shift, while a controlled 4–12 qubit fixed-k=4 probe expands the raw mask space by 24,400× but evaluated masks by only 1.62×.

4. Clear positioning between local pruning and full quantum architecture search. The trained QNN is frozen as the starting point, only masks over existing trainable entanglers are searched, and the structure is frozen before recovery and held-out test evaluation.

Main Findings

• Resource savings at matched sparsity (55% pruning). Mean R_T1 was 61.55 ± 2.10% on Breast Cancer (6-qubit dense-CRY), 71.97 ± 1.99% on Digits01 (8-qubit skip-CRY), and 82.21 ± 5.87% on Iris (4-qubit dense-CRY), versus best simple baselines of 58.10 ± 2.00% (task-only), 70.23 ± 2.20% (task-only), and 73.31 ± 3.22% (compiler-only).

• Accuracy was not uniformly superior. On Digits01, magnitude and task-only pruning matched search validation accuracy; on Iris, compiler-only was 0.42 pp higher. The supported claim is better compiled physical efficiency at comparable task accuracy, not universal accuracy dominance.

• Mask-level search beats one-shot ranking with the same screening heuristic. Paired median resource-score gains over one-shot ConQUESTA were 4.88 pp on Breast Cancer (95% bootstrap CI 0.01–7.07 pp, p = .023, |r_rb| = .89), 2.74 pp on Digits01 (CI 1.85–5.45 pp, p = .008, |r_rb| = 1.00), and 16.78 pp on Iris (CI 15.52–18.89 pp, p = .016, |r_rb| = .94).

• Recovery restores the constraint, not test accuracy. Direct fixed-mask recovery raised feasibility from 16/24 to 22/24 runs at 55% pruning (66.7% to 91.7%), and from 39/48 to 45/48 across both 35% and 55%. Mean test-accuracy deltas versus the dense model at 55% were +0.88 pp (Breast Cancer), −0.52 pp (Digits01), and −1.25 pp (Iris). The two remaining Digits01 failures (seeds 55 and 77) reached 98.61% validation accuracy against a 99.0% threshold, missing by 0.39 pp.

• Resource weights behave as a deployment policy, not a tuned optimum. Re-scoring the same stored masks under uniform (1/3, 1/3, 1/3), depth-oriented (.40, .20, .40), and routing-oriented (.40, .40, .20) policies exactly reproduced the default selection in all 48 benchmark–seed–ratio conditions before any weight change; the largest mean validation shift was 0.44 pp. At 55%, minimum mean Jaccard similarity to the default mask was .93 (Breast Cancer), .99 (Digits01), and .94 (Iris), and 87.5–100% of selected masks lay on the Pareto frontier.

• Screening suppresses combinatorial growth. In the dense-CRY fixed-k=4 probe, raw prunable entanglers grew from 12 to 132, so the full mask space grew from 495 to 12,082,785 (24,400×), while unique evaluated masks grew only from 189 to 305 (1.62×). The explored fraction fell from 38.18% to 0.0025%, and search time grew from 265 to 3254 s.

• Per-mask compilation becomes the next bottleneck. Time per unique mask rose from 1.40 to 10.66 s and peak RSS from 0.79 to 1.55 GB, with a compile-cache hit rate of zero, so the trend is attributable to screening and bounded search rather than cache reuse.

• NSGA-II is a chosen implementation, not a claimed algorithmic advantage. Against an eight-seed Beam comparison it gave 92.87 ± 3.40 vs. 92.54 ± 3.35 validation accuracy and 63.54 ± 2.88 vs. 61.65 ± 3.43 resource score, with paired p = .56 and p = .22.

• Secondary stress tests (three seeds, not repeated in the eight-seed campaign). Masks selected over line/ring/heavy-hex retained savings on an unseen grid: 2Q/depth/SWAP reductions of 54.4/46.8/72.7% (Breast Cancer), 65.6/71.1/86.7% (Digits01), and 79.5/73.8/100.0% (Iris). Noise-induced accuracy shifts ranged from −0.8 to +0.8 pp (Breast Cancer), +0.9 to +2.8 pp (Digits01), and −3.3 to −2.2 pp (Iris).

• One-time offline cost. Estimated NSGA-II search times averaged approximately 4.2, 46.1, and 22.2 min for Iris, Breast Cancer, and Digits01, versus dense-training times of about 0.12, 13.2, and 8.7 min.

Methodology in Plain English

1. Start from a trained circuit. Each benchmark QNN is already trained; the question is only which of its trainable two-qubit entanglers to delete. The benchmarks are Iris (4 qubits, 12 prunable entanglers), Breast Cancer (6 qubits, 30 entanglers), and Digits 0-vs-1 (8 qubits).

2. Score each entangler two ways. Task importance is a first-order saliency — the expected absolute value of the parameter times the gradient of the loss with respect to that parameter. Physical cost comes from removing that single gate and transpiling, measuring the fractional change in native two-qubit count, routed SWAPs, and depth, combined with weights .60 / .25 / .15. The robust cost takes the median across transpiler seeds, then the 75th percentile across scoring topologies, with negative deltas clipped to zero.

3. Combine into one screening priority. After min–max normalization, the priority is (normalized cost) × (1 − normalized importance): physically expensive but task-insensitive gates rank high. Candidate pools are drawn from several views (joint priority, high physical cost, low saliency, gradient-only), capped at K = 20.

4. Search whole masks, not single gates. NSGA-II searches fixed-cardinality masks over four objectives: validation accuracy, two-qubit reduction, depth reduction, and SWAP reduction. Each mask is transpiled in full on line, ring, and heavy-hex maps at optimization level 3 with transpiler seeds 101/202/303. The deployment score R_T1 averages the weighted (.60 two-qubit, .25 SWAP, .15 depth) median fractional reductions across topologies.

5. Apply an explicit accuracy constraint. A mask is feasible if validation accuracy is at least the unpruned model's minus ε = 0.01. Final selection maximizes R_T1 among feasible masks; if none is feasible, it minimizes accuracy violation first and then maximizes R_T1.

6. Recover, then freeze, then test. Direct fixed-mask recovery uses eight fine-tuning epochs at learning rate 0.008 with early stopping disabled, compared to dense training for up to 35 epochs with patience 8. Because the mask is fixed, recovery only updates surviving parameters and cannot undo the compiled reductions. The held-out test split never participates in mask, optimizer, recovery, or resource-policy selection.

7. Control the scaling question separately. A synthetic dense-CRY family at 4, 6, 8, 10, and 12 qubits with fixed k = 4 isolates compiler/search cost from application scaling, reporting raw candidates, pool size, mask-space size, unique evaluated masks, transpilation counts, wall-clock, and peak memory.

Why This Matters

Impact on research. The paper reframes QNN pruning as a constrained, compiler-in-the-loop combinatorial search rather than a per-gate ranking, and shows empirically that the difference is measurable (up to 16.78 pp paired median resource gain on Iris versus one-shot ranking with the same heuristic). It also contributes a negative-ish but useful methodological point: alternative resource weights do not materially disturb accuracy (≤ 0.44 pp), so the scalarization can be treated as a deployment knob. Finally, it positions the method deliberately between local post-training pruning and full quantum architecture search, as a complement rather than a replacement.

Potential real-world applications (these are plausible implications, not deployments demonstrated in the paper):

• Compiling small QNN classifiers onto superconducting devices with sparse coupling maps, where routing overhead dominates gate budgets.
• Deploying variational circuits on near-term hardware with tight coherence budgets, where depth and two-qubit count are the binding constraints.
• Pre-deployment circuit compression in hybrid quantum–classical pipelines for small-feature classification tasks such as the medical and tabular benchmarks studied here.
• Providing a reusable cost model for hardware-aware circuit toolchains that must target several device topologies from one trained model.

Industry relevance. Hardware vendors and compiler teams care about SWAP and two-qubit overhead because those dominate error rates and runtime on sparse devices. A post-training step that requires only a one-time offline search (roughly 4.2–46.1 minutes in this study) and eight recovery epochs, while leaving the trained parameters' structure frozen, is operationally attractive for deployment pipelines. The scaling result — raw space up 24,400× but evaluated masks up only 1.62× — speaks to whether such a tool remains usable as circuits grow.

Future Directions

• Fixed-ratio scaling beyond the current pool. The scaling probe fixes k = 4; at a 55% ratio, k would grow with N and exceed the K = 20 pool at N = 56 and N = 132. The authors explicitly leave fixed-ratio scaling outside the claim, so extending the screening mechanism to growing k is the obvious next step.

• Attacking the per-mask compilation bottleneck. With time per unique mask rising from 1.40 to 10.66 s and a zero compile-cache hit rate, caching, surrogate cost models, or cheaper transpilation proxies could dominate future gains more than better search operators.

• Noise- and fidelity-aware objectives. The current objective is resource efficiency, and noise effects were benchmark-dependent (−0.8 to +0.8 pp on Breast Cancer, +0.9 to +2.8 pp on Digits01, −3.3 to −2.2 pp on Iris). Folding hardware error rates directly into mask selection, rather than treating the method as resource-only, is an open question.

• Broader structural and hardware generality. Evidence covers only dense-CRY and skip-CRY ansätze on three small benchmarks, with grid and noise results from a secondary three-seed stress test. Whether the approach transfers to other ansatz families, larger qubit counts, and arbitrary device layouts remains to be shown.

Target Audience

Quantum machine learning researchers working on circuit compression, pruning, and architecture search; quantum compiler and toolchain engineers concerned with routing and two-qubit overhead; and practitioners deploying small variational circuits on near-term sparse hardware. It also suits graduate students who want a worked example of rigorous, matched-baseline experimental design in quantum ML — including explicit feasibility constraints, held-out isolation, paired statistics with the caveat that n = 8 is a small-sample regime, and careful separation of what the evidence does and does not support.

Authors’ abstract

Deploying quantum neural networks (QNNs) on sparse hardware requires preserving task accuracy while reducing compiler-induced two-qubit gates, SWAPs, and depth. We present ConQUESTA (\underline{C}onstrained \underline{Qu}antum \underline{E}ntangler \underline{S}election with \underline{T}opology \underline{A}wareness), a post training pruning framework that combines task saliency with compiler cost, screens candidate entanglers, and searches complete pruning masks under an explicit accuracy constraint. Unlike one-shot ranking, each mask is transpiled as a whole across multiple coupling topologies. Across 4-, 6-, and 8-qubit QNNs and eight model seeds at 55% pruning, ConQUESTA achieves mean resource reductions of 61.6% on Breast Cancer, 72.0% on Digits01, and 82.2% on Iris, exceeding the strongest matched simple baseline by 3.5, 1.7, and 8.9 percentage points, respectively. Fixed-mask recovery raises feasibility from 16/24 to 22/24 runs. In a fixed-k=4 scaling study on a scaling probe of 4 to 12-qubits, the raw mask space grew from 495 to 12.1M while the evaluated masks increase only from 189 to 305. Resource-weight changes alter depth/routing priorities with at most a 0.44-point mean validation shift.