acceptodds
Under review as a conference paper at ICLR 2027

Prediction as Feedback: Closed-Loop Deep Unfolding for Visual Classification

Abstract

Deep unfolding has emerged as a promising paradigm for visual classification by integrating model-based optimization with deep representation learning. However, existing unfolding-based classifiers often retain an open-loop design without an explicit feedback path from intermediate predictions to subsequent solver updates. To bridge this gap, we propose Pre2Inf, a closed-loop deep unfolding framework that uses prediction as feedback to guide subsequent sparse inference, enabling bidirectional interaction between sparse inference and class prediction. Specifically, we formulate a class-aware sparse inference model and unfold its half-quadratic splitting solver into sequential stages. Each stage incorporates tailored modules for the resulting subproblems, with intermediate predictions mapped to sample-adaptive relevance weights that modulate atom-wise regularization in subsequent coefficient updates. Moreover, we analyze the stability of prediction-guided coefficient updates by characterizing their sensitivity to relevance perturbations. Experiments on eight pixel-level and two image-level classification benchmarks demonstrate consistent improvements over existing unfolding methods. Ablation studies further show that prediction feedback improves class separability and reduces reconstruction ambiguity during inference.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Deep unfolding, prediction feedback, closed-loop inference, sparse representation, visual classification

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_zFAQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes Pre2Inf, a deep unfolding classifier that feeds intermediate class predictions back into later sparse coefficient updates. It uses confidence-weighted predictions to adjust atom-wise regularization, improving results on eight hyperspectral datasets and providing smaller gains on CIFAR-10/100.

Critical strengths

1 Predictions directly change the regularization in the next coefficient solve, giving the method a concrete interpretation beyond simply adding another prediction head. Combining semantic feedback with reconstruction similarity is also a sensible way to retain information from the input.

2 This paper compares open-loop, naive, delayed, and confidence-aware feedback, alongside component removal studies. The full model improves over the open-loop version across the four analyzed hyperspectral datasets, and the CIFAR experiments offer some evidence that the idea extends beyond hyperspectral classification.

Critical weaknesses

1 The experiments use a fixed seed without reporting variation across training runs or sampled splits. Random pixel splits also leave the possibility of overlapping train/test neighborhoods unaddressed. This matters especially for the small CIFAR gains of 0.47 and 0.61 percentage points.

2 Entropy-based gating reduces uncertain feedback, but a confidently wrong prediction can still steer inference toward the wrong class. The naive variant’s sharp accuracy drop on one dataset highlights this risk. The theoretical analysis bounds individual coefficient updates, while convergence of the full loop assumes a contraction condition that is not established for the trained network.

Questions for authors

1 Could you please report results across multiple seeds and spatially separated hyperspectral splits, with validation data used explicitly for hyperparameter selection?

2 How does the full model behave when intermediate predictions are confidently wrong? A controlled feedback-corruption test would help show whether later stages correct or reinforce these errors.

3 Table 3 shows substantial closed-loop gains already at Stage 1, although feedback is initially disabled. Could you separate benefits learned during training from the effect of feedback during inference—for example, by disabling feedback in a trained model?

4 What are the parameter count, inference latency, and memory cost relative to the open-loop model and representative unfolding baselines? How does the coefficient solve scale as the number of classes and dictionary atoms increases?

Reviewer_uHKeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Pre2Inf, a closed-loop deep unfolding framework that feeds intermediate class predictions back into sparse inference. The method combines prediction confidence with reconstruction-based evidence to adapt atom-wise regularization and improve discriminative representations. The authors conduct experiments on eight hyperspectral classification datasets and CIFAR-10/100, and evaluate the proposed feedback mechanism through theoretical analysis and ablation studies.

Critical strengths
  1. Intermediate predictions modulate atom-wise regularization in subsequent sparse inference, providing a clear and interpretable closed-loop design.

  2. The experiments cover multiple datasets and classification settings, with comparisons of feedback strategies and component-wise ablations providing reasonably broad empirical validation.

Critical weaknesses
  1. The SA results are inconsistent across tables. For example, Tables 2 and 12 report an OA of 97.88% for the full model, whereas Tables 3, 4, and 6 report 97.55%. The authors do not explain whether these differences arise from different experimental settings, making it difficult to establish the exact gain from prediction feedback on SA.

  2. Some OA and per-class accuracy values in Table 14 appear incompatible with the stated use of identical test splits across methods. For example, A2S2K exceeds the average OA of DBCT and SMFAT by 2.44 percentage points, whereas its largest per-class advantage over the average of these two methods is only 1.815 percentage points. Under standard evaluation definitions and a common test set, the overall advantage cannot exceed the largest per-class advantage; thus, rounding cannot explain this discrepancy. This issue directly affects the reliability of the main comparison and requires further explanation of the underlying predictions and metric calculations.

  3. The current experiments do not sufficiently distinguish the independent benefit of prediction feedback from training-induced changes in representations. In Table 3, the closed-loop variant already substantially outperforms the open-loop variant at the first stage, when semantic feedback is disabled, while accuracy does not consistently improve in subsequent stages. I do not consider the current controls sufficient to attribute the performance gains specifically to prediction feedback during inference.

Questions for authors
  1. Can OA and per-class accuracies be recomputed from the same set of predictions? Could you also explain the discrepancies in the SA results across tables?

  2. Could you specify the source of the dictionary atoms, classifier architecture, dimensionality reduction settings, stage-wise (\beta_t), main regularization parameters and loss weights, and the precise location of stop-gradient in Equation (3)? Do dictionary construction and preprocessing use only the training information permitted by the evaluation protocol?

  3. What validation data are used to select hyperparameters and checkpoints? How is spatial overlap handled when training and test pixels are randomly sampled from the same image? Such overlap may affect the interpretation of the generalization performance of spectral–spatial classification methods.

Reviewer_WfN8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Pre2Inf unfolds HQS updates for class-aware sparse inference into a network. It converts intermediate class probabilities and reconstruction–atom similarity into atom-wise regularization to guide subsequent coefficient updates. The feedback uses stop-gradient, delayed activation, and entropy-based confidence weighting. Training combines classification, HQS consistency, and class-aware reconstruction losses. Table 2 reports OA improvements on eight HSIC datasets, while Table 5 compares open-loop and closed-loop variants on CIFAR-10/100. The appendix analyzes single-step update stability and provides component ablations.

Critical strengths
  1. Feedback has an explicit role in optimization. Equations (5), (7), and (8) map predictions to diagonal regularization in the coefficient subproblem, making their effect on subsequent inference concrete. Appendices A.1–A.2 establish uniqueness and perturbation bounds for fixed relevance weights, making the mechanism straightforward to inspect.
  2. The results show practical gains and expose meaningful failures. In Table 2, average OA is 89.11 versus 86.04 for SMFAT. In Table 4, Full outperforms Open on all four datasets, whereas Naive achieves only 30.62 on Dataset 9. These results show that feedback design matters and warrants investigation.
Critical weaknesses
  1. Feedback reliability remains unverified. Equation (17) treats low entropy as reliability, but the coarse confidence bins in Figure 5 do not establish whether confidently wrong predictions are repeatedly amplified. Appendix A.2 only bounds single-step perturbations, while A.3 assumes contraction. Neither establishes the claimed closed-loop correction mechanism.
  2. Attribution and fair comparison are insufficient. Although first-stage feedback is zero, Closed already leads by 1.06–2.61 percentage points in Table 3; these gaps cannot be directly attributed to online correction. Table 2 omits the directly relevant NSR-Net, and §5.1 uses a fixed random split without multi-seed variability or spatially disjoint evaluation.
Questions for authors
  1. Why does Closed already outperform Open at the first stage in Table 3, when feedback is zero? Does Open retain the same similarity MLP, stage-wise supervision, loss weights, and training budget? Please identify Datasets 8/9/10/14.
  2. How many incorrect predictions fall into each confidence bin in Figure 5? For confidently wrong predictions or controlled erroneous feedback, what is the error amplification rate of Full relative to Open?
  3. Which samples are used to construct the dictionary and fit dimensionality reduction? Is tuning performed on a separate validation set? Do the gains persist against NSR-Net under matched supervision, across multiple seeds, and with spatially disjoint splits?
  4. Does the implemented network satisfy the contraction assumption in Appendix A.3? Which optimization properties remain guaranteed after replacing the exact proximal operator with the learned residual module?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.