acceptodds
Under review as a conference paper at ICLR 2027

Learning What to Encode: Representation Selection for Competitive Quantum Tabular Classification

Abstract

Establishing the predictive competitiveness of quantum machine learning against strong classical baselines remains a central challenge in tabular classification. We propose learned representation selection (LRS), a hybrid quantum–classical classifier that learns a discrete representation choice for each encoding slot under a fixed quantum input budget. Instead of prescribing a representation mixture, LRS assigns each slot a feature from complementary representations, allowing the encoded input to adapt to the prediction task. A select–freeze–refit procedure separates learning these assignments from training the final predictor: the selection stage learns slot-wise choices, the freeze stage fixes the resulting discrete configuration, and the refit stage trains the predictor with that configuration held constant. The selected features are angle-encoded into a variational quantum circuit, whose expectation values are combined with a classical feature branch. Across twenty tabular datasets, LRS achieves accuracy and macro-F1 comparable to tuned random forests, XGBoost, and LightGBM, with advantages on several datasets. These results support slot-wise discrete representation selection as a practical approach to competitive hybrid quantum classification under a constrained encoding budget.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Quantum Machine Learning, Hybrid Quantum–Classical Learning, Representation Allocation, Variational Quantum Circuits, Tabular Classification, Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_utRaOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes learned representation selection for hybrid quantum–classical tabular classification. Under a fixed encoding budget, each input slot selects between a PCA coordinate and an NCA coordinate through a learned binary gate. Selection is trained jointly with the predictor, after which the mask is frozen and the predictor is reinitialized and retrained. Experiments on twenty binary classification datasets compare the method with tuned tree ensembles and variants of the representation, circuit topology, and encoding width. The results show competitive average predictive performance, although the reported pairwise comparisons do not establish statistically significant improvements after multiple-comparison correction.

Critical strengths

S1: The evaluation includes twenty datasets, tuned classical baselines, shared outer test folds, and preprocessing fitted within the corresponding training partitions. These choices make the reported predictive comparisons more informative than comparisons against weak or untuned baselines. The paper also provides useful diagnostics and candidly reports limitations, including substantial retention of initial selection masks, slightly higher average scores from NCA-only inputs, and the absence of evidence for quantum advantage.

Critical weaknesses

W1: The methodological contribution is insufficiently distinguished from existing differentiable feature selection and learned quantum encoding. Concrete Autoencoders already learn discrete feature selections and include experiments that retrain predictors on the selected features; Stochastic Gates jointly learns feature selection and prediction. Both are cited but receive no experimental comparison. Relevant quantum embedding approaches, including Neural Quantum Embedding and Q-FLAIR, are also absent from the comparisons. The specific restriction to one PCA/NCA choice per slot may be distinctive, but the experiments do not establish its benefit over adapted existing selectors under a matched encoding budget.

W2: The experiments do not convincingly isolate the value of learning the selection mask. Appendix E reports that only 13.3% of eligible source choices differ from initialization and that 52 of 100 final masks retain every initial choice on non-padded slots, yet no frozen random-mask baseline is evaluated. NCA-only inputs achieve slightly higher mean accuracy and macro-F1, and none of the twelve corrected comparisons against the ablation variants reaches the 0.05 significance threshold. Furthermore, Count-12 changes the fitted NCA representation, circuit grouping, and training procedure, so it does not isolate the selection mechanism. A matched random-mask comparison, an ablation of fresh refitting, and a classical-only predictor using the same selector are needed to establish which proposed components account for the observed performance. These gaps leave the central claimed contribution insufficiently supported.

Questions for authors

None.

Reviewer_YPe5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes learned representation selection (LRS) for hybrid quantum–classical binary classification under a fixed encoding budget. Two preprocessing pipelines based on PCA and supervised NCA provide candidate coordinates. A global binary mask selects one coordinate from each index-matched pair, using hard forward decisions and surrogate gradients. The mask is learned jointly with a hybrid predictor, then frozen; the predictor is reinitialized and trained on the selected representation. Both the quantum and classical branches receive this representation. The main configuration uses twelve qubits and two fixed six-wire CNOT rings. Experiments on twenty binary tabular datasets use five outer cross-validation folds and compare against tuned tree ensembles and representation, topology, and width variants. The reported aggregate accuracy and macro-F1 are close to those of the tree references. However, the experiments do not yet isolate the incremental benefit of learned selection and fresh refitting sufficiently to establish the central methodological contribution.

Critical strengths
  1. Competitive predictive performance on the evaluated tabular tasks. The complete hybrid pipeline achieves mean accuracy and macro-F1 of 87.15% and 83.76% across twenty binary classification datasets, close to the results of tuned RF, XGBoost, and LightGBM (Table 1). The paper additionally compares against trees trained on the PCA/NCA candidate bank (Table 4). This provides useful empirical evidence that the evaluated hybrid architecture can achieve aggregate predictive performance comparable to these strong classical references under a fixed quantum encoding budget.

  2. Detailed and transparent experimental reporting. The paper reports common outer test folds, preprocessing and tuning protocols, per-dataset results, and multiple-comparison-corrected paired tests. It also discloses findings that qualify its conclusions, including the NCA-only variant's slightly higher mean scores and substantial retention of initial mask choices. This reporting helps readers assess both the observed competitiveness and the limitations of the evidence.

Critical weaknesses
  1. The additional methodological complexity is not sufficiently justified by the demonstrated benefits. The pipeline combines PCA/NCA preprocessing, discrete source selection with surrogate gradients and a margin penalty, a separate refitting stage, and quantum-classical fusion. Although representation, topology, and width ablations are provided, the incremental benefits of learned selection, fresh refitting, and the quantum branch are not sufficiently isolated. Count-12 simultaneously changes the NCA fitting dimension, circuit grouping, and training procedure, so its comparison cannot isolate the selection strategy (Section 4.2 and Appendix B.1). The paper also lacks a direct comparison of fresh refitting with retaining or continuing to train the selection-stage predictor, and a matched classical-only or classical-replacement control. Notably, the simpler NCA-only variant achieves slightly higher mean accuracy and macro-F1 than LRS (87.26/83.91 versus 87.15/83.76), although LRS has better average ranks among the seven variants (Table 2). These results leave the added value of the proposed machinery over simpler alternatives insufficiently established.

  2. The rationale and robustness of the paired selector are insufficiently supported. Section 3.1 acknowledges that PCA and NCA coordinates are not intrinsically aligned, yet the selector must retain exactly one coordinate from each index-matched pair. This guarantees the input count but does not explain why this particular restriction is appropriate: it excludes potentially useful subsets containing both coordinates from one pair and neither from another. No pairing-sensitivity analysis or comparison with a less restrictive, same-budget selector is reported. Moreover, while Table 6 specifies logit initialization, temperature annealing, and the margin setting, the experiments do not establish robustness to these choices or include a no-margin control. This matters given Appendix E's finding that 52 of 100 final masks retain every initial choice on non-padded slots. Without frozen-initial-mask and repeated random-mask controls, it remains unclear how much the selector improves on a fixed random mixture. These observations motivate further validation; they do not by themselves show that the pairing is invalid or that mask learning is ineffective.

Questions for authors
  1. Component contributions: Under comparable training budgets, how much does fresh refitting improve over continuing training with a frozen mask? How much does the quantum branch improve over a matched classical-only predictor?

  2. Pairing rationale: Why should each slot select exactly one of the same-index PCA and NCA coordinates, given that they are not intrinsically aligned? Does changing the pairing materially affect performance?

  3. Selector robustness: Does learning the mask outperform freezing its random initialization? How sensitive is this result to initialization, temperature annealing, and the margin penalty?

  4. Significance: The limited encoding budget motivates input selection, but the method appears to optimize within this constraint rather than address the constraint itself. Could the authors explain why this constitutes a meaningful advance, and whether the evidence supports a broader contribution than this interpretation suggests?

Reviewer_2XBAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed a learned representation selection (LRS), a hybrid quantum–classical classifier that learns a discrete representation choice for each encoding slot under a fixed quantum input budget.

Critical strengths

The experiments seem to be solid.

Critical weaknesses

The intuition is not sufficient. More proofs should be introduced.

Questions for authors

Could the authos give more proofs to wxplain the intuition?

Reviewer_kMyhOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes learned representation selection for hybrid quantum–classical tabular classification under a fixed encoding budget. The method learns a global binary mask that selects between PCA and NCA coordinates at each input slot, using hard forward decisions and surrogate gradients. After selection, the mask is frozen and the predictor is reinitialized and trained on the selected representation. The predictor combines a variational quantum circuit with a classical neural network branch. The authors evaluate the method on 20 binary classification datasets, reporting performance comparable to tuned tree ensembles and examining different representations, circuit topologies, and encoding widths.

Critical strengths

The algorithmic pipeline is clearly presented, and the select–freeze–refit procedure is easy to follow. The experiments cover 20 tabular datasets and include comparisons with several tuned classical baselines, providing a useful assessment of the overall predictive performance.

Critical weaknesses
  1. Insufficient justification and validation of the selection updates.

    The method uses hard binary decisions with sigmoid surrogate gradients, but the benefit of this particular update mechanism is not sufficiently established. Appendix E reports that only 13.3% of non-padded slot choices across the outer folds differ from their initialization, and 52 out of 100 final masks retain all initial choices on these slots. It is unclear how much the learned selection contributes to the reported performance.

  2. Missing ablations of the quantum and classical branches.

    Although the paper includes representation, topology, and width ablations, it does not evaluate quantum-only and classical-only versions of the predictor. Both branches receive the same selected representation, so the current results do not establish their individual contributions or whether combining them improves performance over either branch alone.

  3. Limited methodological novelty.

    The method primarily combines PCA/NCA representations, a straight-through binary selector, and a hybrid quantum–classical classifier within a select–freeze–refit procedure. I do not see a substantial original algorithmic contribution beyond this combination. The paper provides limited insight into why this particular combination should be advantageous, also why the selected representation should be processed by both quantum and classical branches. This makes it difficult to evaluate the contribution of the paper.

Questions for authors
  1. Given that many mask entries remain unchanged after initialization, what is the significance of the selection step? How much does learned selection improve over simply retaining the initial mask or using a random mask?

  2. How do quantum-only and classical-only predictors perform with the same input representation and comparable training and tuning budgets? Does the hybrid model improve accuracy or macro-F1 over both?

  3. What is the substantive algorithmic contribution beyond combining the existing representation, selection, and classification components? What explains the advantage of the proposed hybrid architecture?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.