acceptodds
Withdrawn from ICLR 2027

Spiking Learned ISTA for Inverse Problems with Finite-Time Convergence Analysis

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Deep unfolding offers an interpretable framework for learned sparse recovery. Learned Iterative Shrinkage-Thresholding Algorithm (LISTA) unfolds sparse-recovery iterations into a trainable reconstruction network. However, it represents the recovered coefficients as real-valued activations processed by dense multiply-accumulates. By contrast, a spiking implementation can represent them through sparse signed spike trains and replace dense coefficient multiplications with event-triggered accumulations. We therefore propose Spiking LISTA (S-LISTA), which replaces LISTA soft-thresholding with a Signed Soft-Threshold Integrate-and-Fire (SST-IF) unit that uses paired non-leaky integrate-and-fire neurons for the positive and negative branches and coordinate- and stage-specific scales to balance firing-rate resolution and saturation. We prove a finite-time approximation bound for SST-IF, with error proportional to the scale divided by the number of timesteps, and propagate this error through a fixed-depth unfolded network. The resulting bound shows that the spiking error of S-LISTA decreases as the number of timesteps increases. We evaluate S-LISTA on synthetic support recovery, natural-image compressive sensing, compressed-sensing MRI, and low-precision deployment. S-LISTA preserves the support-recovery behavior of its non-spiking counterpart, and after fine-tuning, it closely matches the corresponding backbone on Set11, BSD68, and MRI while remaining competitive with unfolded baselines. Under low-precision deployment, S-LISTA remains stable under reduced weight precision and avoids the sharp failures observed in its non-spiking counterpart under aggressive post-training quantization. Its measured spike activity further suggests that event-driven affine execution can reduce synaptic arithmetic.

Primary area:
optimization
Keywords:
spiking neural networks, inverse problems, compressed sensing, iterative shrinkage-thresholding algorithms, deep unfolding

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

The paper proposes S-LISTA, replacing LISTA’s soft-thresholding with signed integrate-and-fire neurons, and derives finite-timestep approximation bounds for a fixed-depth network. Experiments cover sparse recovery, image compressive sensing, MRI, and low-precision deployment.

Critical strengths

The signed spiking operator provides a principled connection between sparse recovery and event-driven computation, supported by explicit error analysis. After fine-tuning, S-LISTA closely matches its non-spiking backbone and shows improved robustness to aggressive post-training quantization.

Critical weaknesses
  • The theoretical guarantee assumes constrained weights and non-saturation, whereas the trained models use a different parameterization and allow clipping.
  • Energy savings rely on an operation-count proxy; actual hardware energy and latency are not measured.
  • Reconstruction accuracy remains below stronger unfolding baselines, and the quantization benefit is not isolated from the effects of clipping and additional fine-tuning.
Questions for authors

See the weaknesses.

Reviewer_3C7ZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes S-LISTA, a spiking implementation of LISTA using paired integrate-and-fire neurons for soft-thresholding. It establishes finite-timestep approximation bounds under specific assumptions. Experiments demonstrate competitive reconstruction accuracy relative to its non-spiking backbone and improved robustness in some low-precision settings.

Critical strengths
  1. S-LISTA preserves the optimization-inspired structure of LISTA and provides a straightforward spiking implementation of its soft-thresholding operator.

  2. The analysis separates LISTA recovery error, measurement noise, and finite-timestep spiking error, with constructive choices of scales and timesteps for a fixed-depth guarantee.

  3. Experiments cover synthetic support recovery, natural-image reconstruction, MRI, and quantization. Fine-tuning substantially reduces conversion losses, while some aggressive quantization settings reveal improved robustness.

Critical weaknesses
  1. The method largely repackages a straightforward conversion of LISTA using standard IF neurons.
    SST-IF is obtained by decomposing soft-thresholding into two shifted ReLUs and replacing each branch with a conventional IF firing-rate representation. Its signed output comes from subtracting the branches, its dead zone comes from shifting the inputs by the threshold, and its range and resolution come from scaling by MM. The underlying IF dynamics are unchanged. Likewise, Equation (5) retains the existing LISTA-CP residual update and only replaces its activation. The paper therefore needs to establish what substantive technical advance goes beyond this direct conversion. Coordinate-specific scales and fine-tuning provide useful implementation choices, but their independent contributions are not sufficiently isolated.

  2. The central efficiency motivation remains experimentally unverified.
    The implementation decodes firing rates between affine stages, while the reported energy proxy estimates a future implementation operating directly on spike streams. It excludes memory access, neuron-state updates, comparisons, routing, and control overhead. The bound $N_{\mathrm{AC}}\le T N _ {\ mathrm {MAC}} $ also does not establish computational savings.

  3. The recovery guarantee does not cover the model responsible for the reported results.
    The analysis relies on coherence-minimizing weights, a prescribed threshold schedule, and non-saturation. The trained image models instead use tied weights W(k)=βGW^{(k)}=\beta G and percentile-based scales, with Table 4 reporting 2.7% saturation after fine-tuning. Thus, both the parameter construction and the operating regime differ from those required by the recovery theorem.

  4. The quantization experiments do not isolate a benefit attributable to spiking.
    The paper associates improved robustness with the bounded SST-IF output, but bounded activations can also be implemented in a conventional ANN.

Questions for authors

See weakness

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

This paper proposes Spiking LISTA (S-LISTA), a spiking extension of the Learned Iterative Shrinkage-Thresholding Algorithm for sparse recovery and inverse problems. The main contribution is a Signed Soft-Threshold Integrate-and-Fire (SST-IF) operator, which uses two non-leaky integrate-and-fire neurons to approximate the positive and negative branches of soft-thresholding. The authors claim an (O(1/T)) finite-timestep approximation bound and extend it to a fixed-depth unfolded network. Experiments on synthetic sparse recovery, image compressive sensing, and compressed-sensing MRI show that S-LISTA can largely preserve the reconstruction performance of its ANN counterpart after fine-tuning. The paper also investigates low-precision quantization and estimates potential synaptic arithmetic savings using measured spike activity.

Critical strengths
  1. The paper addresses an interesting problem of incorporating spiking computation into deep unfolding networks for inverse problems. The signed nature of soft-thresholding makes this setting more challenging than conventional ReLU-based ANN-to-SNN conversion.
  2. The finite-timestep approximation analysis is clearly presented, and the authors explicitly acknowledge important limitations, including the fixed-depth nature of the guarantee and the difference between theoretical assumptions and practical model parameterization.
  3. The experiments cover several reconstruction settings, including synthetic support recovery, natural-image compressive sensing, and MRI. The results demonstrate that finite-timestep spiking conversion with fine-tuning can retain much of the accuracy of the corresponding non-spiking backbone.
Critical weaknesses
  1. The proposed SST-IF operator is constructed by decomposing soft-thresholding into two ReLU branches and approximating each branch using the firing rate of a non-leaky IF neuron. While this construction is reasonable, the underlying ingredients are well established in ANN-to-SNN conversion. More importantly, for a constant input and the deterministic IF dynamics considered in this paper, the firing rate can be expressed directly as a clipped, quantized function of the input. Consequently, SST-IF is functionally equivalent to a particular clipped and uniformly quantized soft-thresholding operator at the level of its decoded outputs. This raises a fundamental question: to what extent do the reported accuracy preservation and quantization robustness arise from spiking computation, rather than activation clipping, quantization, and subsequent fine-tuning? The current experiments do not adequately isolate these factors. I would expect comparisons against non-spiking LISTA variants incorporating activation clipping, low-bit activation quantization, and quantization-aware fine-tuning, with matched quantization resolution and training budgets. Without these comparisons, it is difficult to establish whether the proposed spiking formulation provides benefits beyond those achievable through conventional low-precision ANN implementations.

  2. A major motivation of the paper is that spike-based representations can replace dense multiplications with event-triggered accumulations. However, the implementation described in the paper decodes spike rates between unfolding stages rather than implementing the proposed spike-stream affine operations end to end. The efficiency evaluation in Section 4.3.2 is therefore based on a theoretical synaptic arithmetic proxy rather than measured execution cost. The reported proxy excludes important sources of overhead, including spike generation, membrane-state updates, memory access, routing, and control operations. These costs may be significant, particularly when a fixed number of timesteps is required for each reconstruction.

  3. The finite-depth approximation result relies on a specific LISTA parameter construction based on generalized mutual coherence, together with sparsity and non-saturation assumptions. However, the practical image-reconstruction models use the tied parameterization (W^{(k)}=\beta G), with learned sampling and dictionary parameters. The experimentally calibrated scales also allow a nonzero fraction of activations to saturate, including approximately 2.7% after fine-tuning. Although the authors acknowledge these differences, they significantly limit the applicability of the theoretical results to the actual models being evaluated. In addition, the result characterizes approximation to LISTA at a fixed unfolding depth as the number of timesteps increases. It should not be interpreted as a general convergence guarantee for the trained S-LISTA network as its depth increases. I would encourage the authors to provide empirical measurements of layerwise approximation errors, stability factors, and their relationship to the theoretical bounds. This would help establish whether the analysis provides meaningful insights into the behavior of practical trained models.

  4. The low-precision experiments are interesting, but the conclusions regarding S-LISTA's robustness should be qualified. For example, in Table 3, the NVFP4-aware trained A-LISTA achieves an NMSE of 0.0432, outperforming the corresponding NVFP4-aware trained S-LISTA models, whose NMSE ranges from 0.0475 to 0.0715 across the evaluated timestep settings. Thus, the apparent advantage of S-LISTA is primarily observed when comparing post-training quantization of existing FP32 checkpoints, rather than under comparable quantization-aware training. Similarly, the MRI quantization results do not show uniformly smaller quantization degradation for S-LISTA across sampling ratios and formats. These observations suggest that the benefits may depend strongly on the quantization protocol and the specific baseline. A controlled comparison with clipped or quantization-aware ANN models is needed before attributing the observed robustness specifically to the spiking architecture.

Questions for authors
  1. Since SST-IF with deterministic non-leaky IF neurons and static inputs can be represented by a clipped quantization function, how does S-LISTA compare against an ANN-LISTA model using the equivalent quantized soft-thresholding operator, particularly in inference latency and energy?
  2. Can the authors provide hardware-aware efficiency results that include spike generation, membrane updates, memory access, and routing overhead? How does the proposed implementation compare against an optimized LISTA implementation that retains the factorized residual update rather than explicitly multiplying by the dense matrix (H)?
  3. How closely do the layerwise approximation errors and stability factors of the trained image and MRI models follow the theoretical analysis? Can the authors quantify the effect of saturation on the reconstruction error?
  4. Why does NVFP4-aware trained A-LISTA outperform NVFP4-aware trained S-LISTA in the synthetic experiments? What evidence supports attributing the improved post-training quantization behavior to spiking computation rather than clipping or quantization-aware fine-tuning?

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

Related papers

Discussion (0)

Sign in to comment.