acceptodds
Under review as a conference paper at ICLR 2027

Budget-Aware Step Selection for Forward-Only Recovery of Analog Accelerators

Abstract

Recovering a drifted analog accelerator with forward-only optimization costs queries both to update its parameters and to choose those updates. We study step selection under this combined cost. In a controlled testbed, measured single-step descent boundaries exceed whole-run failure boundaries by a geometric mean of : initial progress alone is a poor guide to a full recovery run. We isolate accumulated readout noise, derive a finite-horizon displacement bound, and use its step scale to select a rate from donor-device measurements without searching on the target device. The resulting model captures observed noise and probe scaling, but overpredicts failure boundaries by in geometric mean; its budget dependence remains unresolved. On held-out simulated crossbar targets, a selector with a fine rate grid and donors matched by device condition uses fewer operational forward evaluations than target search, with a percentage-point accuracy deficit. A transferred condition-specific fixed setting is more accurate at lower incremental cost, showing that avoiding target search need not require donor measurements. Calibration uses clean simulator readout, and the bound concerns noise-only displacement, not device reliability. These results support evaluating recovery methods by both accuracy and configuration cost, with fixed transfer as an explicit comparator. An anonymous project page is available at https://dordlark158.github.io/budget-aware-step-selection/

Primary area:
optimization
Keywords:
zeroth-order optimization, analog in-memory computing, forward-only adaptation, step-size selection, query efficiency, photonic computing.

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_edbfOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses a drifted analog accelerator after deployment with forward-only optimization recovery, that costs queries to update its parameters and to choose those updates. Specifically, it proposes a budget-aware step-size selection for forward-only recovery of degraded analog accelerators due to costly forward evaluations, including configuration search, calibration, validation, and recovery updates. The targeted research question here is to answer what accuracy-cost trade-off the shared calibration can achieve. To do this, this paper first investigates the forward-only recovery and its costs by training a network offline, simulating degradation, and adapting the parameters. Then, it employs a recovery step that can provide meaningful progress without accumulating excessive noise over the run. It leverages a two-point zeroth-order (ZO) optimization to estimate a gradient from forward evaluations, derives a finite-horizon bound on parameter displacement, and converts the scale into a donor-calibrated rule. In each run, the final parameter state replaces the drifted baseline only if it improves mean accuracy over five independent noisy validations; otherwise, the policy returns the baseline. All evaluations are based on simulations on an MNIST classification task and analog devices (e.g., ReRAM).

Critical strengths

The configuration cost is treated as part of the optimization problem, by considering forward evaluations used for calibration, candidate search, recovery, monitoring, and validation. It is useful as it shows the overhead of recovery algorithms for identifying the appropriate configuration.

Critical weaknesses

• The problem statement, motivation, and scope of study (e.g., the types of analog devices) are not well established in the beginning of the introduction section. Instead, the paper directly mentions the research question on accuracy-cost trade-off while skipping essential definitions of several techniques/terminologies (e.g., recovery, shared calibration, donor, etc.), and why they matter in a chain of arguments. This writing style makes it difficult to follow the discussion in the early part of the paper. • The paper studies devices that expose writable parameters and forward scalar-loss evaluations, but no backward pass for computing gradients. However, the scope of analog accelerators should be defined more explicitly. For instance, the accelerators are targeted for neural network workloads and the analog devices technologies such as non-volatile memories (NVMs). • All evaluations are based on simulations on a small MNIST classification task, limited types of analog devices (e.g., ReRAM) and analog crossbar architectures, and non-practical network model architectures. This condition is insufficient to draw conclusions about the effectiveness of the proposed techniques. Moreover, the details of simulators, analog crossbar architectures, analog device models, and device drift models are not clearly described and characterized. These are important to understand the experimental setup from hardware perspective, and how it is connected with the overall targeted evaluations.

Questions for authors

Besides addressing the above mentioned weaknesses, address the following: • It is suggested to provide evaluations based on bigger/practical datasets (e.g., ImageNet-1K), several analog devices (e.g., ReRAM, PCM, MRAM, and optical), practical analog crossbar architectures, and practical network model architectures (e.g., VGG-16, ResNets, etc.). • It is suggested to clearly describe and characterize the details of simulators, analog device models, and device drift models, because they are important to understand the experimental setup, especially from hardware perspective, and how it is connected with the overall evaluations. • It is also suggested to clarify the generality and scalability of the proposed solution across different design choices, including crossbar architectures, analog devices technologies (e.g., ReRAM, PCM, MRAM, and optical), network models, and larger datasets (e.g., ImageNet-1K and CIFAR-100). • The costs should also be presented in terms of latency, memory, and power/energy overheads to clearly show the practical requirements to conduct the proposed solution.

Reviewer_3narOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to select recovery step sizes for analog accelerators affected by parameter drift when only forward loss queries are available and the query budget is limited. The authors distinguish single-step expected descent from whole-run failure, derive a finite-horizon bound on noise-only parameter displacement, and use donor-device measurements of noise and a radius proxy to select step sizes without per-target configuration search. On 36 simulated crossbar targets disjoint from the donors, the proposed fine-grid, condition-matched strategy reduces operational forward evaluations by 93.62% relative to per-target search, with a 0.399 percentage-point deficit in the primary test accuracy, measured with transient readout noise disabled and static damage retained. Transferring previously acquired condition-specific fixed settings achieves higher accuracy at lower current operational cost. The paper emphasizes that configuration acquisition costs should be included in the evaluation of recovery algorithms.

Critical strengths

The paper accounts for calibration, search, recovery, monitoring, and validation costs, including those of failed candidates. The main experiment uses disjoint donor and target devices, identical damaged initial states across policies for each target, and transfer policies frozen before target recovery, strengthening the credibility of the comparison. The authors also clearly state the scope of the theoretical analysis and openly report results where fixed-setting transfer outperforms the proposed method, providing a balanced account of the accuracy-cost trade-off and the method’s limitations.

Critical weaknesses
  1. The fine-grid, condition-matched strategy achieves a primary test accuracy of 95.075%, below the 95.391% achieved by transferring previously acquired condition-specific fixed settings. It also incurs higher current operational costs, that is, approximately 1.217 million versus 1.154 million forward evaluations. Pooling all donors produces the same selections as the fixed default while adding calibration costs. Although the historical acquisition costs of existing settings are not fully accounted for, preventing a definitive full-lifecycle comparison, the current results still do not establish when donor measurements provide decision-relevant information beyond fixed settings or condition labels. This limits the justification for the proposed method and the understanding of its applicability.
  2. The proposed rule uses η∝B−1/2\eta \propto B^{-1/2}, whereas the directly fitted budget exponent is -0.282, with a reported interval of [-0.43, -0.13]. Given the single seed and coarse boundary grid, these measurements do not adequately establish the budget-scaling relationship. The main 36-target experiment also uses only a 30,000-query optimization budget and does not demonstrate that adjusting the step size with the budget improves on keeping it fixed. Furthermore, the theory considers only noise-driven displacement, the radius proxy is not equivalent to an actual failure boundary, and its calibration relies on clean simulator readout. The current evidence therefore primarily supports an accuracy-cost trade-off in fixed-budget simulations, rather than establishing effective step selection across varying budgets and device conditions.
Questions for authors
  1. Please distinguish between scenarios where a fixed configuration is already available for transfer and those where a configuration must be acquired. Using the existing target results, could you clarify under which base models, noise levels, or drift conditions donor measurements are beneficial? Can the contribution of the measurements themselves be separated from those of condition matching and candidate-grid refinement?
  2. Could you use existing cross-budget results, or provide a small number of budget comparisons within the current experimental setup, to compare scaling the step size according to the proposed rule with keeping it fixed? The key issue is whether the budget term improves recovery decisions and how the exponent discrepancy affects the selected step sizes and recovery accuracy.
  3. What conclusions are supported by the existing noisy-radius diagnostics? Please use the available results to explain how calibration errors affect step selection, recovery accuracy, and query costs, and clarify which capabilities still depend on clean simulator-based calibration.
Reviewer_w6vBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies budget-aware step selection for forward-only recovery of analog accelerators. It derives a finite-horizon noise-displacement bound and uses measurements from donor devices to select a learning rate. On 36 simulated crossbar targets, the proposed selector reduces operational forward evaluations by 93.62% relative to target search, with a 0.39 percentage point accuracy deficit. A transferred fixed setting achieves better accuracy at lower current operational cost (Sec. 4.3).

Critical strengths
  1. The paper studies configuration cost with disjoint donor and target devices and explicit query accounting (Sec. 2.1, Sec. 2.2). The controlled study shows significantly fewer operational forward evaluations than target search while reporting the associated 0.39 percentage point accuracy loss. The comparison is quite informative than reporting metrics beyond the optimization budget.

  2. The experiments identify a substantial difference between initial descent and whole-run failure. The measured direction-count and readout-noise exponents also support parts of the proposed noise scaling. These finding substantiate the motivation for considering the complete recovery horizon

Critical weaknesses
  1. The donor-based selector does not outperform transferred cohort-specific fixed settings, which achieve higher accuracy at lower current operational cost (Sec. 4.3:373-378). The coarse/substrate selector also reproduces the fixed default's outcomes while adding calibration calls. It is therefore unclear when the donor measurements justify their additional cost.

  2. The predicted budget dependence remains unresolved, with direct measurements giving an exponent of −0.282 rather than the predicted −0.5 (Sec. 4.2:320-323). Calibration also requires clean simulator readout, which has not been demonstrated on physical devices. These issues limit the evidence for applying the selection rule across budgets and to physical hardware.

Questions for authors
  1. When would donor-based selection outperform cohort-specific fixed transfer after accounting for the acquisition cost of both approaches?

  2. Can calibration work using only noisy forward evaluations, and how would that affect the selected rates?

  3. How does selector performance vary with recovery budget, given the discrepancy in budget scaling?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.