acceptodds
Under review as a conference paper at ICLR 2027

Transform Coordinates Separate Vulnerability from Compensatory Capacity under Deployment Perturbations

Abstract

Local sensitivity identifies functional directions vulnerable to deployment perturbations without necessarily revealing which can compensate for the induced error. A three-parameter Iwasawa representation of the linear canonical transform (LCT), together with a real-valued quadrature of its complex kernel, yields a token-mixing matrix whose derivatives with respect to these parameters can be obtained analytically. Analog in-memory computing (AIMC) supplies a controlled post-training perturbation, followed by hardware-aware fine-tuning to measure adaptation. These differentials capture pronounced anisotropy across coordinates, with additional variation remaining unexplained. After mapping, fractional order retains the highest sensitivity yet exhibits the smallest displacement during adaptation, while the relative contributions of scale and shear change with depth. A local response estimate produces candidate-subset rankings that align more closely with restricted recovery than sensitivity-based selection, and comparison with unrestricted fine-tuning shows that the LCT parameters account for only part of the available recovery. Transform coordinates thereby separate vulnerability from compensatory capacity and provide an interpretable basis for prioritizing post-deployment adaptation.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
post-deployment adaptation, parameter sensitivity, hardware-aware fine-tuning, analog in-memory computing, linear canonical transform, selective fine-tuning

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_6G2oOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates parameter sensitivity and adaptation under deployment perturbations in neural networks using linear canonical transform (LCT) token mixing. The authors analyze three transform coordinates and examine their sensitivity, parameter movement, and recoverability after mapping to analog in-memory computing (AIMC) hardware. A regularized local response score is introduced to select parameter subsets for restricted fine-tuning. The experiments suggest that parameter sensitivity does not necessarily correspond to compensatory capacity.

Critical strengths

The authors provide theoretical derivations for the LCT parameterization and conduct several experiments to investigate parameter sensitivity, movement, and restricted recovery.

Critical weaknesses

My main concern is the lack of a clear and focused contribution. The paper combines LCT parameterization, hardware perturbation analysis, sensitivity estimation, and restricted fine-tuning, but the connections between these components are not sufficiently explained. I find it difficult to understand what the paper is ultimately trying to achieve and to distinguish its methodological contributions from the various empirical observations. The overall presentation is confusing, and the paper fails to clearly establish what fundamental new insights or algorithmic contributions it provides.

The experimental evaluation is also rather limited, focusing on a specific LCT-based architecture. Moreover, the results show that most recoverable capacity lies outside the LCT parameter space, which raises further questions about the practical significance and generalizability of the proposed approach.

Overall, I believe the paper requires substantial restructuring to clarify its research objectives, technical contributions, and the connections between its theoretical analysis and experimental findings.

Questions for authors

See weaknesses above.

Reviewer_n9gvOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper shows that the parameters that are sensitive to hardware-induced errors are not always the best ones that practitioners should target to fix during finetuning. The paper also demonstrates a method for response-based criterion to improve the identification of such parameters to reduce error. Experimental results show that applying the criterion achieves this goal, albeit in very specific settings.

Critical strengths
    1. The paper tackles an interesting problem that involves both hardware and software component of finetuning. Its insights show that parameters most vulnerable to hardware-induced errors are not necessarily the parameters whose adaptation best restores performance. This may be challenge conventional intuition.
    1. The proposed approach is mathematically backed and appears to be supported empirically. The score correlates with realized recovery (p = 0.705), with a reported 95% interval of [0.607,0.798][0.607, 0.798]), providing empirical support for the idea that recovery potential can be estimated more effectively than by sensitivity alone.
Critical weaknesses
    1. The experiments are performed on CIFAR and ImageNet with limited configuration variability. This restriction raises doubt on the generalizability of the paper's findings.
    1. The results also lack convincing baselines along with interpretations. Although the response score correlates with realized recovery (rho = 0.705), the paper could more clearly establish how much it improves practical recovery over sensitivity-based baselines and whether the improvement justifies its computational cost.
Questions for authors
    1. Could the authors quantify the improvement of response-based selection over pre-mapping sensitivity, post-mapping sensitivity, and random selection in terms of absolute loss reduction or accuracy recovery, rather than ranking correlation alone?
    1. On robustness, can the authors provide evidence that the framework captures a general property of deployment perturbations beyond this specific configuration?
    1. To what extent do the findings depend on the chosen datasets, model architecture, and training configuration? A suit of sensitivity tests would help?
    1. Is it possible to clarify what are the assumptions or empirical conditions necessary for this ranking advantage to hold, and where they may fail or need more attention?
Reviewer_8m9POpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies post-deployment adaptation of LCT-parameterized token mixing under an AIMC simulation. It stores Iwasawa parameters (θ,s,k)(\theta,s,k), while optimization and coordinate-wise analysis use p=(θ,log⁡s,asinh⁡k)p=(\theta,\log s,\operatorname{asinh} k). A real-valued quadrature of the LCT kernel yields the token-mixing matrix, for which the authors derive analytic quadrature derivatives. They map FP32/FP16 mixed-precision models to AIHWKit with 8-bit ADC and 6-bit DAC, then separate sensitivity MiiM_{ii}, forcing ∣gi∣|g_i|, realized movement ∣Δpi∣|\Delta p_i|, and restricted recovery. The main claims are: quadrature derivatives capture the dominant between-axis sensitivity anisotropy but not the depth-dependent residual; fractional order remains the most sensitive axis yet contributes the least realized movement during hardware-aware fine-tuning; scale and shear contributions vary with depth; a regularized local response score QM(S)Q_M(S) ranks restricted subsets more reliably than sensitivity and random controls at the matched one-third coordinate budget in most CIFAR comparisons; and the LCT subspace contains only part of the recoverable loss, with channel-branch adaptation recovering the largest share.

Critical strengths
  • S1. The conceptual separation is clear and consequential. The paper distinguishes vulnerability, forcing, movement, and recoverability in interpretable transform coordinates. The finding that fractional order is highly sensitive but realizes little movement shows that sensitivity is not a reliable proxy for compensatory movement or restricted capacity in this setting.

  • S2. The deployment setup is controlled and diagnostically useful. AIMC mapping plus hardware-aware fine-tuning defines a reproducible simulated perturbation, while matched random controls, label-convention checks, parameterization robustness, and the separation of CstructC_{struct} and RcR_c strengthen the restricted-selection analysis.

Critical weaknesses
  • W1. The generality is narrower than the framing suggests. The evidence is tied to LCT token mixing, one fixed AIHWKit RPU configuration, twelve small-scale CIFAR models, and two Tiny ImageNet models, one per quadrature, used only descriptively. Tiny ImageNet does not evaluate subset selection. In addition, CIFAR-10 and CIFAR-100 differ in both task and depth, so cross-model capacity differences do not identify a pure depth effect.

  • W2. The practical selection value is not fully stress-tested. The main matched-budget comparison is against pre-mapping sensitivity, post-mapping sensitivity, stratified random selection, and axis- or axis-and-block-composition-matched random controls. It does not include a stronger matched-budget activation-, gradient-, curvature-, or hardware-aware selection baseline. The practical benefit of retaining the full coupling matrix is also modest and not uniformly positive: most of the ranking gain is already captured by QdiagQ_{diag}, while QfullQ_{full} mainly suppresses unusually poor selections. This comparison is based on only twelve CIFAR models, and Tiny ImageNet does not evaluate subset selection, so the practical advantage of QM(S)Q_M(S) remains under-supported beyond the tested CIFAR setup.

Questions for authors
  • Q1. Under other ADC/DAC precisions, RPU configurations, noise levels, or readout settings, do the qualitative orderings remain stable, especially the combination of high fractional-order sensitivity and low realized movement? Does response-based subset ranking retain its advantage at a matched budget?

  • Q2. At the same released-coordinate budget, how does QM(S)Q_M(S) compare with matched activation-, gradient-, curvature-, or hardware-aware subset-selection baselines? If OPTQ-, AdaRound-, or RAOQ-style methods are used as comparisons, can they be adapted to the same released-coordinate objective and budget?

  • Q3. Because coordinates saturated by the clamp have vanishing derivatives (gi=Mii=0g_i=M_{ii}=0), does an active-set or KKT treatment of boundary coordinates change the reported subset rankings? In particular, would identity-regularized QM(S)Q_{\mathcal{M}}(S) need to be modified so that saturated coordinates are either excluded a priori or handled as constrained variables?

Comments on the paper

1 comment

Thank you for the follow-up and the additional results. The responses have clarified several of my earlier concerns.

1. Corrected random controlsThe corrected results address my concern about the random controls. The response criteria still outperform the composition-matched controls in 20/24 comparisons under both matching schemes. Although some margins have decreased, I agree that the conclusion in Section 5.3 remains supported.

2. Comparison with the gradient baselineThe explanation of the CIFAR-10 real-quadrature result is helpful. However, the advantage over gradient-based selection is less clear in the other settings. On CIFAR-10 (imaginary), QfullQ_{\mathrm{full}} achieves 0.934 compared with 0.928 for the gradient baseline. On CIFAR-100 (imaginary), QdiagQ_{\mathrm{diag}} achieves 0.583 compared with 0.584. These results are quite close, and I agree that the benefit is setting-dependent.

Finally, the additional depth and hardware experiments are useful, but the findings remain limited to LCT models under simulated AIMC conditions. I suggest keeping this scope explicit in the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.