acceptodds
Under review as a conference paper at ICLR 2027

From Spectral Responses to Task Effects: Auditing Frequency Dependence in Spiking Neural Networks

Abstract

Connecting a spectral response in a spiking neural network to task behavior requires measurements beyond band energy. We present a frequency-lineage protocol that separates operator responses, fixed-decoder responses, and predictive interventions. A MaxFormer trace locates the spread of an input-band perturbation at the first LIF, and a three-architecture, three-dataset panel records cross-frequency decoder responses. Matched operator experiments place this observation in context: ordinary ReLU and tanh also generate cross-frequency components, while integration and reset change task performance. In three frozen MaxFormer networks, equal-budget binary patches selected using same- or cross-band targets reduce classification loss, but the tested joint non-source target does not. These patches preserve binary states without strict band isolation; continuous Fourier patches give a different joint response at a larger dose. Separately, learned natural residuals show no consistent predictive advantage across the development panel, and beating a control need not mean beating the baseline. These findings motivate distinct tests for where a frequency perturbation acts, how a decoder responds, and whether an intervention benefits the task. They constrain explanations based on spectral attenuation without identifying information preservation or loss.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spiking Neural Networks, Frequency Analysis, Cross-Frequency Routing, Frequency Lineage, Temporal Representation, Event-Based Vision, Mechanistic Analysis, Representation Probing

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_5nRKOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a frequency-lineage protocol. The main claim is that the decay of certain frequency band does not represent the loss of information while the interaction between different frequency bands does not mean the retain of the information. The experiments test the cross-frequency band interaction under the condition of different non-linear function and the effect on the classification loss led by recovering certain frequency bands on several benchmark datasets and common SNN baselines.

Critical strengths
  1. Interesting research question: The authors investigate the distinction between spectral attenuation and task-relevant information in spiking neural networks, while emphasizing that the decay of certain frequency band activity is not related to the loss of information and the reaction between different frequency band does not mean the retain of information.
  2. Reasonable experimental setting: The frequency-lineage protocol studies the different effects caused by operator responses, fixed decoder sensitivity, and functional interventions, which provides a structured perspective for analyzing frequency activity in SNNs.
Critical weaknesses
  1. Limited novelty: Cross-frequency generation is a trivial and well-studied consequence of nonlinear operations, since the nonlinearity caused saturation will naturally lead to different hamornics. Although the author acknowledge this, the paper does not demonstrate what new understanding of the mechanism is gained from the observation of SNNs and the chosen datasets. The methods novelty shows as the combination of spectral perturbation, decoder probing, and activation patching, which is not convincing of new methodology knowledge.
  2. Unclear scientific contribution: The experiments do not establish whether task-relevant information is preserved, or evicted. Except for the the acknowledged limitation in the paper, the remaining experimental results still provide limited evidence of a new computational mechanism.
    • While Figure 2 shows cross-frequency perturbation effects, it does not demonstrate their modification or improvement to the original classifier.
    • In Figure 3, cross-band-targeted binary patches yield modest CE recovery (+0.0115 nats), whereas joint non-source-band patches yield negative recovery (−0.0036 nats). Besides, these patches are not strictly frequency-isolated, therefore, the observation could not lead to clear causal attribution of the observed improvements to specific frequency components.
  3. Insufficient evaluation of the proposed protocol. The work does not demonstrate whether special perturbation measurements or fixed-decoder responses can predict the effectiveness of the downstream interventions. Therefore, the proposed protocol remains descriptive instead of a feasible mechanism to bring functional benefits. Meanwhile, the advantage of the proposed protocol over there analysis techniques is not transparent.
  4. Limitation in experiemental design. The selected datasets which are commonly used for SNNs classification tasks do not explicitly require temporal frequency information. This problem make the benefits brought by the proposed protocol less direct. Besides, the use of only 16 time steps constrain frequency resolution.
Questions for authors
  1. Scientific contribution: Given to the well-estabilished studies on the nonlinear functions, what new insights are provided beyond conventional spectral perturbation analysis?
  2. Causal validation: Although the authors acknowledge some limitations of the work, such as the observed cross-frequency responses do not create information persistent, decoder sensitivity does not demonstrate original-classifier use, and the interventions do not permit clear frequency-specific causal attribution, my question is what concrete scitific findings can be drawn from the experimental results?
  3. Size of the eval datasets: In DVS128Gesture, only tens of samples are evaluated, could that support the claimed scientific findings?
Reviewer_L7vaOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Most traditional spiking neuron models have a low pass filter implementing somatic integration of inputs (potentially paired with additional low pass or band pass filtering in synaptic transmission). It was recently argued that these filtering properties pose problems for network computation. Recent attempts at using spiking codes in deep network models (Fang et al., 2025; Ye et al., 2026; Feng et al., 2025) overcome this limitation by changing the spectral properties of the spiking units in some way. The aim as far as i can tell is to reason through the spectral filtering properties of such an unit for the purpose of credit assignment (and ultimately optimization).

Novelty and significance: low. The mechanics of achieving that remain opaque and the results seem unimpressive. No meaningful numerical comparison to other proposals in the literature, thin on results, limited to no intuitions to be had and limited conclusions. Not sure who could benefit from reading this paper as it does not seem have to have concrete practical implications for neuromorphic chip design nor for the existing spiking net artificial architectures (as far as I can tell).

Critical strengths

None

Critical weaknesses

Poor problem formulation, inconsistent and haphazard changes in architecture across the narrative, critical details about the experimental setup missing, insufficient justification of metrics, no clear conclusion.

Questions for authors

I am missing some very basic elements of the setup

what is the precise architecture (feedforward? is there recurrence assumed at any point for the time series processing?) ? what is the nature of the task? what is the nature of the inputs and outputs? how does the information get expressed in the spectral domain?
what is the exact form of the spiking neuron dynamics? (LIF/MaxFormer/Relu/tanh are all mentioned at some point) What are the operators mentioned in line 91?

How you are thinking about spatial vs temporal distinctions in the nature of the data? Several datasets mentioned but exact numerical experiments unclear: e.g. nCaltech101 is a spike version of 100 static images converted to a spiking code. How exactly are these converted into time series and what does that imply for the structure of the input spectrum (given that much is known about the spectrum of natural images)

What purpose does the fixed dual-tone audit serve and what does it have to do with ReLU and tanh ?

Given that the numerics use 'short sequences' how much of the spectrum is even computable empirically?

Can you please explain the metrics in more detail? could not follow the text describing the the base and clear distinction

How (it at all) does what is described here relate to perturbative methods of estimating gradients such as REINFORCE?

Reviewer_R2oTOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the relationship between spectral responses and task-relevant effects in spiking neural networks (SNNs). The authors propose a frequency-lineage protocol that combines input-band perturbations, cross-frequency response measurements, fixed-decoder analyses, and predictive interventions. Experiments across several SNN architectures and event-based datasets demonstrate cross-frequency perturbation effects, while the subsequent restoration experiments show limited or inconsistent improvements in task performance.

Critical strengths
  • This paper addresses a relevant question: changes in spectral energy should not be directly interpreted as changes in task-relevant information.
  • The paper distinguishes spectral responses, decoder sensitivity, and predictive benefits, and generally avoids drawing conclusions beyond what the experiments support.
  • The paper includes extensive experiments and control analyses, with careful discussion of their limitations.
Critical weaknesses
  • The motivation in the Abstract and Introduction is not sufficiently clear. While the authors explain why spectral attenuation should not be equated with information loss, it is unclear what specific research gap they aim to address and why the proposed frequency-lineage protocol is needed. This makes it difficult to understand the motivation behind the different experiments.
  • The paper lacks a clear scientific narrative and a well-defined central contribution. Despite the extensive experiments, I found it difficult to identify the main scientific insight of this work. While the paper introduces several analyses, including controlled operator comparisons, perturbation propagation, fixed-decoder responses, oracle activation patching, and learned residual interventions, it is unclear how they fit together and what new understanding emerges from their combination.
  • Although the paper motivates its framework as a way to connect spectral responses to functional consequences, the experiments do not establish whether the proposed spectral or decoder-level measurements can explain or predict the behavior of the original classifier.
  • The main findings are largely expected and provide limited new scientific insights. For example, the generation of cross-frequency components by nonlinear operations is already well understood, and the performance improvements from clean-reference activation patching are not particularly surprising.
Questions for authors
  • What is the central scientific hypothesis or new understanding that the proposed frequency-lineage protocol is intended to establish, beyond the general observation that spectral responses and task information are not equivalent?
  • Can the authors quantitatively relate the cross-frequency response or decoder-response measurements to the effectiveness of original-classifier interventions? For example, can these measurements predict which destination frequency bands are most beneficial to restore?
  • How does the proposed framework provide insights into SNN computation that could not be obtained from existing spectral analyses and causal intervention methods separately?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.