acceptodds
Withdrawn from ICLR 2027

DASH-LISTA: Disentangling History Feedback in Hybrid-ISTA Networks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

History can improve an unfolded solver's learned function or become integral to a particular trained computation. We introduce DASH-LISTA to separate these roles through a state-preserving Hybrid-ISTA interface that controls correction information and injection location. Matched retraining, fixed-backbone replacement and equal-norm interventions measure architecture benefit, model dependence and directional effects. Under a shared feedback-gain cap, four-tap FIR improves on two-time and nonlinear current controls by and dB in three independently trained confirmation instances, each tested on 100,000 previously unused signals; all paired gains are positive. Original-gain effects remain tail-sensitive, with one short-history ordering reversal. Equal-norm reflection identifies useful older directions. In images, current replacement retains a dual-branch cost after 6,400 adaptation steps and relaxation to six independent bounded stage gains: about dB in each of two checkpoint cohorts. The same relaxation slightly improves refinement-only reconstruction. On test50, dual-branch reflection at stages four and five has a larger terminal effect through the direct path in all six checkpoints. Independent retraining yields nearly coincident image endpoint means. These results separate architecture benefit from checkpoint dependence and locate pathway-specific effects of older directions.

Primary area:
optimization
Keywords:
deep unfolding, learned optimization, sparse recovery, compressive sensing, history feedback, mechanistic analysis

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

This paper explores the historical information in Hybrid-ISTA networks for sparse signal and image reconstruction, and the authors introduce DASH-LISTA, which incorporates corrections based on previous iteration updates into the original solver.

Critical strengths
  1. This paper explores how historical information affects reconstruction performance and compares differnet feedback design.

  2. The proposed FIR feedback achieves clear improvements over several baselines on synthetic sparse recovery tasks.

Critical weaknesses
  1. DASH-LISTA is to introduce historical iteration updates into the existing Hybrid-ISTA framework. Similar ideas of using historical information have already been explored in iterative optimization and memory-based unfolding networks.

  2. The convergence analysis relies on conditions that are not verified for the actual trained models. It is therefore unclear what new theoretical insight the paper provides beyond analyzing the behavior of the proposed corrections.

  3. Although DASH-LISTA shows significant improvements on synthetic sparse recovery tasks, the results on real image reconstruction are much less convincing. In Table 4, the simpler current-only method achieves slightly better PSNR than FIR feedback on both validation and COCO, while also being faster.

  4. The authors show that removing historical information from an already trained FIR model can lead to performance degradation. However, this does not necessarily demonstrate that historical feedback is better than training a simpler model from scratch. In fact, the independently trained current-only model performs similarly or better.

  5. Given the additional complexity and inference cost, I am not convinced that the proposed method offers a clear practical advantage. The large difference between synthetic and real-image results also raises questions about how broadly the conclusions apply.

Questions for authors
  1. What is the main methodological novelty of DASH-LISTA compared with existing memory-based unfolding methods? Beyond adding historical corrections and studying different injection locations, what is the key new contribution?

  2. In Table 4, the current-only method achieves better reconstruction quality with lower inference latency. Could the authors provide a practical setting where historical feedback clearly outperforms the simpler alternative?

  3. Why does FIR feedback provide such large improvements on synthetic sparse recovery but not on real image reconstruction?

  4. Can the theoretical analysis provide any guarantees or predictions for the actual trained DASH-LISTA models?

  5. The authors clearly state:" Generative AI tools, including ChatGPT and Codex, assisted with literature retrieval, research and experiment design, mathematical derivations and proof drafting, implementation, result analysis, manuscript drafting and editing, and figure preparation." What did the authors really contribute during the whole research?

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

This paper asks whether history in a hybrid-ISTA sparse coding solver is important. They introduce a memory module to the original Hybrid-ISTA architecture which they call DASH. They ablate their proposed history setup, and generally compare to multiple baselines, including: current gain, two-time, nonlinear current, and equal-norm intervention. They evaluate the role of history in a synthetic setting, and an image setting. They find that history is relevant in the synthetic setting but that it's not clearly necessary in the image setting.

Critical strengths

The paper studies the role of history for ISTA-type solvers for signal reconstruction, which is important problem. The paper provides a nuanced result, which shows that history can improve performance in certain settings (e.g., their synthetic setting) but it's unclear the extent to which history improves performance in other settings (e.g., their image setting).

The paper seems methodologically correct (with a few issues around standard variant, see below), although given the difficulty understanding the paper, it's possible there are issues based on parts that I missed / didn't understand properly.

Critical weaknesses

Clarity: This paper is poorly written, to the point where the paper would have to be substantially re-written to be acceptable for publication. The paper is extremely dense and challenging to understand, commonly uses terminology and notation without definition. While I'm not an expert in this domain, it required substantially more than the typical amount of effort for a new research domain to be able to understand the the basic questions that the paper is asking, and what it contributes.

Soundness: In Table 3, in the synthetic setting, the standard deviation of the paired gain is larger than the difference between Two-time / Nonlinear and the full method in many of the instances of the table, which calls into question whether the result (that the added history is better than these two baselines) is true. In Table 4, in the image setting, the difference between methods is small, and also falls within the reported standard deviation, which suggests that it's unclear whether the reported results are in fact real.

Impact: It's not clear to me that this paper would provide impactful insights for the community. For one, as mentioned above, it's unclear if the results are sound. In addition, even if the results are sound, the paper only evaluates the role of history on a simple synthetic setting, and an image setting, and across these settings appears to find different results (negligible improvement in the image setting, and some improvement in the synthetic setting). The paper would therefore not practically inform whether or not to use history in another setting. I think the paper could be improved by offering some insight into when to use history and when not.

Questions for authors

See weaknesses.

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

This paper studies trajectory history feedback in DASH-LISTA by separating architectural benefit under matched retraining, checkpoint dependence under frozen-backbone replacement, and older-increment directional effects under equal-norm reflection.

Critical strengths
  1. This paper achieves a clean experimental disentanglement of retraining benefits, checkpoint dependence, and directional effects. Instead of treating post-hoc module removal as joint evidence for both architecture choice and checkpoint simplification, it cleanly separates matched retraining, frozen-backbone replacement, and same-state reflection protocols.

  2. This paper also provides well-controlled synthetic validation under a shared gain cap. With freshly drawn sensing operators, initializations, and test streams, and with all 24 support-bin contributions remaining positive (Table 25), it rules out the alternative explanation that the observed gains simply arise from feedback-scale inflation on the synthetic benchmark.

Critical weaknesses
  1. The analysis distinguishes effects within a trained checkpoint, but I find its connection to finite replacement outcomes underdeveloped. In appendix F.2, the prospective score substantially underestimates all three replacement losses and reduces MAE relative to predicting zero change by only about 4.6%.

  2. Although the fitting budget is extended to 6,400 steps, Current4, TV6, and Box6 all remain scalar rescalings of the current increment(d_k=g_k r_k). I am therefore unsure how much of the remaining image replacement gap reflects historical information rather than the limited expressiveness of these alternatives.

  3. Table 4 does not establish an image reconstruction advantage for FIR over the matched current control, while the persistent replacement cost is approximately 0.026 dB. This is a repeatable effect, but the earlier common-reference analysis in appendix I does not establish different retraining and replacement decisions at a shared quality tolerance, leaving its practical consequence unclear.

I recommend weak rejection. The controlled comparisons are useful, but the limited connection to finite replacement outcomes and the restricted replacement family leave the explanatory and practical contribution insufficient for acceptance.

Questions for authors
  1. Does the image replacement gap persist with a small nonlinear function of the current increment, keeping the backbone frozen and using the same fitting (selection) split and a comparable gain envelope?

  2. For the three checkpoints in appendix F.2, could you show the interpolation loss (R_lambda) between FIR and the fitted current replacement? This would clarify where the finite replacement effect departs from the first-order prediction.

  3. What quality–cost trade-off would make the approximately 0.026 dB replacement loss consequential for retaining history in an existing model? Please distinguish this practical interpretation from the exploratory shared-tolerance calculation in appendix I.

  4. Could the main image replacement and branch-intervention findings be checked on a separate image set with the model and replacement parameters held fixed? Coco comparison in Table 4 evaluates the trained architectures, rather than these replacement and intervention results.

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

Related papers

Discussion (0)

Sign in to comment.