acceptodds
Withdrawn from ICLR 2027

SpooQ: Anatomy-Guided Quantum Skeleton Action Recognition

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Quantum machine learning (QML) promises compact models that capture correlations costly to represent classically, yet has not been demonstrated on complex spatio-temporal problems such as skeleton-based action recognition. We introduce SpooQ, to our knowledge the first quantum machine learning architecture for skeleton-based action recognition, designed for near-term quantum hardware by encoding human anatomy directly into its circuit structure. On the NW-UCLA and a single-person subset of NTU RGB+D 60 cross-view benchmarks, SpooQ outperforms a parameter-matched state-of-the-art graph convolutional network by 7.0 and 3.5 accuracy points, reaching 84.3% and 68.5%, respectively. Under 1024-shot simulations using noise models derived from the IBM Brisbane and Prague backends, SpooQ retains approximately 80% accuracy and outperforms a flat quantum baseline by 27–38 accuracy points. These results extend QML beyond static benchmarks to a demanding spatio-temporal task and demonstrate robustness under finite-shot, backend-derived noise.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Quantum, Skeleton, QML

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_4HN5OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SpooQ, which the authors claim is the first quantum machine learning architecture for skeleton-based action recognition. The core motivation is to extend QML beyond static benchmarks to complex spatio-temporal tasks while addressing the constraints of noisy intermediate-scale quantum (NISQ) devices. Instead of encoding entire poses into a monolithic quantum circuit, SpooQ decomposes the skeleton into six anatomical parts (arms, legs, head, torso) and assigns compact quantum circuits at three scales—local part circuits, relational region circuits (arms, legs, upper body), and a global body circuit. The method is evaluated on NW-UCLA and a single-person-filtered subset of NTU RGB+D 60 (cross-view), outperforming parameter-matched classical baselines, including CTR-GCN.

Critical strengths
  1. Applying QML to skeleton-based action recognition with anatomy-guided circuit design is novel. Treating anatomical structure as a circuit design template is conceptually sound.
  2. The paper's experiments show that the anatomy hierarchy is helpful (Classical SpooQ > CTR-GCN), and quantum circuits add further gains beyond the hierarchy.
Critical weaknesses
  1. The current results are on filtered datasets with very limited scale: the NTU60 experiments use only single-person clips (30,683 training clips) rather than the full dataset. The "parameter-matched" comparison to CTR-GCN is misleading since the parameter counts are remarkably small (13K vs. 1.4M for full CTR-GCN).
  2. The paper compares against only a reduced CTR-GCN, a flat quantum baseline, and a classical version of SpooQ. However, it omits recent strong skeleton-based methods, such as PoseC3D, ST-GCN++, 2s-AGCN, and especially modern transformer-based approaches (e.g., STTFormer, MotionBERT).
  3. The claim that quantum circuits "capture correlations costly to represent classically" is not tested, no explicit comparison is made to nonlinear classical feature maps (e.g., random Fourier features, polynomial kernel approximations) at matched parameter counts.
  4. The experimental setup uses analytic simulation (default.qubit), which does not reflect real NISQ execution realities.
Questions for authors
  1. What would be the resource requirements on actual quantum hardware?
  2. How many classes retain sufficient samples after filtering? Does the performance on mutual-action classes differ from single-actor classes?
  3. Under Brisbane noise, SpooQ retains ~95% of its accuracy (80.1/84.3) while the Quantum Baseline retains only ~60% (42.6/70.9). Can you provide more insight into why?
  4. Can you provide comparisons with other QML approaches such as quantum recurrent networks and QLSTM?
  5. SpooQ processes frames independently through quantum circuits and then uses a GRU for temporal aggregation. What is the maximum sequence length you can handle given quantum circuit depth constraints?
Reviewer_WasdOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SpooQ, a quantum machine learning architecture for skeleton-based action recognition. It combines anatomically organized variational quantum circuits, pairwise readout, and classical temporal modeling. Experiments on NW-UCLA and a single-person-filtered NTU60 split show improvements over compact classical and flat quantum baselines, while backend-derived noise simulations demonstrate greater robustness than the flat quantum baseline.

Critical strengths
  1. Well-motivated architecture. Anatomical decomposition provides a clear structural inductive bias for organizing compact quantum circuits across multiple body scales.
  2. Informative empirical validation. The width-matched readout control supports the value of joint measurement features, while finite-shot noise simulations show substantially better robustness than the flat quantum baseline.
Critical weaknesses
  1. Limited comparisons with classical neural models. The classical baselines are limited to an author-designed MLP counterpart and a substantially downsized CTR-GCN. Without broader comparisons against recent, competitive neural architectures under comparable resource budgets, the results are insufficient to establish the relative benefits of introducing quantum computation for skeleton-based action recognition.
  2. Lack of validation on physical quantum hardware. All training and noise evaluations are performed using quantum-circuit simulators. Consequently, the reported results do not establish effectiveness and robustness under actual hardware constraints. Practical advantages over classical alternatives in execution time, resource consumption, and predictive performance remain unverified.
Questions for authors
  1. Could you compare SpooQ with recent, competitive neural architectures and stronger classical replacements for the quantum circuits under comparable parameter budgets, while also reporting computational costs, to clarify the benefits of the quantum components?
  2. Could you validate SpooQ on physical quantum hardware and compare its accuracy and end-to-end inference cost with classical baselines, accounting for state preparation, measurement shots, and hardware connectivity constraints?
Reviewer_T8AzOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SpooQ is a hybrid quantum–classical architecture for skeleton action recognition that combines anatomical part, region, and whole-body circuits with temporal mixing and a GRU classifier. It evaluates single-qubit and pairwise readouts on NW-UCLA and a filtered NTU60 subset, with classical comparisons, component ablations, and finite-shot noise simulations. Although the perspective of this manuscript is interesting, it does not show the impact for the task of SBAR, especially compared with the latest methods such as Skateformer, BlockGCN, ProtoGCN...

Critical strengths
  1. The anatomical hierarchy provides a clear way to allocate compact quantum circuits to local body parts and coordinated regions.

  2. At similar parameter counts, SpooQ exceeds Classical SpooQ by 5.3 accuracy points on NW-UCLA and 1.9 points on filtered NTU60.

  3. The width-matched product-readout control and three-seed noise experiments provide useful evidence about relational measurements and robustness.

Critical weaknesses
  • Classical SpooQ uses only two hidden units per replacement MLP, limiting the classical comparator family used to establish the quantum contribution.

  • The flat quantum baseline has fewer parameters and a different classifier, so its comparison with SpooQ does not isolate the benefit of anatomical decomposition in Section 4.

  • The noise simulations omit backend connectivity constraints and execution-cost accounting, leaving practical feasibility on near-term hardware insufficiently established.

  • As shown in Tabel 7, the filtered NTU60 split leaves several interaction classes with only three test clips, so aggregate accuracy provides limited evidence of performance across all 60 classes.

  • CTRGCN is accepted in ICCV2021. Lack comparison with the latest SBAR methods such as BlockGCN, ProtoGCN, InfoGCN....

  • SBAR has many benchmarks can validate the effectiveness of the methods. However, the current version shows the insufficient experiments.

Questions for authors
  1. Does the advantage remain against stronger classical feature maps, such as structured quadratic blocks, under the same total parameter budget?

  2. How does SpooQ compare with a flat quantum model using matched parameter counts and the same classifier?

  3. What are the routed circuit depths, gate counts, and noise-simulation results when backend connectivity is respected?

  4. Can you report macro-averaged accuracy and separate results for single-person and interaction classes on filtered NTU60?

  5. Can you clarify the impact or contribution of this work compared with the mainstream methods (BlockGCN, ProtoGCN and so on) for the Action Recognition task?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.