acceptodds
Under review as a conference paper at ICLR 2027

Robust Resonate‑and‑Fire Neurons for Real-Time End‑to‑End EEG Emotion Recognition

Abstract

Real-world applications of EEG analysis call for energy-efficient, low latency operation to enable scalable, real-time inference on streamed signals. Spiking Neural Networks (SNNs), through sparse, event-driven processing, can satisfy these demands, yet their benefits are rarely realised in practice as implementations stick to pipelines that feature dense, windowed computation. This includes SNNs for EEG emotion recognition, where the use of windowed spectral features hampers sparse computation, inflates latency, and coarsens temporal resolution. More elaborate neuron models can address these limitations: the Resonate-and-Fire (RF) model captures spectral features via oscillating membrane dynamics, but remains under-explored due to parameter sensitivity and instability. Existing RF formulations mitigate these issues at the cost of added complexity; we instead introduce a robust RF model, bolstered by data-driven initialisation, that enables reliable and efficient integration into SNNs, as evidenced by reduced computational cost and spike rates compared to other RF models. Leveraging our RF neurons to extract spectral features directly from raw EEG, we further present **SS-EEGNet**, a Spiking and Streaming EEGNet-based model for online end-to-end emotion recognition in real-time. Our approach enables causal, low-latency inference while preserving sparsity, achieving competitive accuracy alongside substantial efficiency gains.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spiking Neural Networks, Resonate-and-Fire Neurons, EEG Emotion Recognition, Neuromorphic Computing, Streaming Inference

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_2hzbOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

A RF-neuron-based SNN for efficient time & frequency domain learning of features, primarily from EEG data, also benchmarked on standard neuromorphic datasets. The architecture provides sparser outputs by setting a high, parameterised threshold (for the RF neuron) and using exact Euler discretisation instead of a widely adopted forward discretisation. For spatial filtering, 2D convolutions and pooling operations have been utilised.

Critical strengths
  1. Interesting threshold scaling reparameterization for sparser outputs.
  2. A thorough literature review on reset mechanisms to familiarise readers with the literature on RF neurons.
  3. Clearly state the data preprocessing and architecture hyperparameters for reproducibility and a complete understanding of the experimental results observed.
Critical weaknesses
  1. Marginal gains over baselines for neuromorphic datasets: The authors argue the importance of exact discretisation over the first-order baselines; however, the results do not agree, as RF methods (Higuchi et. al. 2024) outperform their methods, and the exact method marginally surpasses their method (mainly due to the proposed threshold scaling). Is there any limitation of the introduced method for recurrent tasks?
  2. The authors proposed SS-EEGNet from the prior method EEGNet. However, the authors failed to isolate the performance benefit of utilising a frequency-based RF method, which they argue is their primary contribution.
Questions for authors
  1. Why does the method not outperform prior methods (both RF and LIF-based) using the proposed discretisation and threshold strategy?
  2. Please assess the performance of your method on the long-range arena (LRA) datasets as well, in addition to EEG-specific datasets.
  3. Similarly, on the EEG tasks, please evaluate the performance of the already available recent SSM-based models too, such as PSpikeSSM, SpikingSSM, SH2RFSSM, etc.
  4. What is the benefit of a frequency-based feature extractor in RF neurons, and how exactly does it help in emotion recognition? (Shen et al., 2020) (as cited) utilises 4-D frequency-band-based features instead of the 3-D tensors for EEG-based emotion recognition. Can you isolate the performance benefits of such bands and explain how their performance differs from that of RF neurons?
  5. Maybe run ablations to see whether RF-based learnable frequency methods outperform such curated frequency-based features.
Reviewer_DyRDOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents a robust resonate-and-fire (RF) neuron designed to reduce parameter sensitivity, improve numerical stability, and control spike sparsity while preserving simple dynamics. It combines exact discretization, resonance-based threshold scaling, and data-driven initialization, with reset and refractory mechanisms for firing control. Building on these components, the authors introduce SS-EEGNet, which extracts spectral features directly from raw EEG signals for end-to-end emotion recognition at each timestep. Experiments on several neuromorphic benchmarks and the DEAP and SEED-VII datasets assess classification accuracy, parameter count, and spike sparsity.

Critical strengths
  1. The paper derives an exact discrete-time update for RF neurons and a firing threshold based on the resonance response, providing an interpretable basis for threshold selection.

  2. Building on this formulation, SS-EEGNet uses RF neurons to process raw EEG signals sample by sample and extract spectral features. SS-EEGNet preserves temporal information in its neuronal states, reducing reliance on window-based spectral analysis.

Critical weaknesses
  1. Insufficient validation of real-time streaming. Section 4 describes predictions based on full-sample membrane-potential aggregation, and the conclusion acknowledges that streaming deployment remains unvalidated. Evaluation of online decision-making under continuous inputs and device-level processing latency is needed to support the real-time claims.

  2. Limited evidence of end-to-end energy savings. The SATA-sim evaluation focuses on individual neurons and a synthetic two-layer network. Energy evaluation of the complete SS-EEGNet pipeline on EEG workloads, including state-update and memory-access costs, is needed to establish system-level efficiency gains.

  3. Unclear identification of the proposed methods. Table 1 uses boldface for both selected method names and best-performing results without clearly distinguishing these conventions. Explicit labels for the proposed methods, ablation variants, and baselines would make the comparisons easier to interpret.

Limited methodological novelty. Prior work has already explored RF-based streaming inference and related threshold initialization strategies [1, 2]. The paper does not clearly establish how its formulation advances beyond these approaches, making the methodological contribution appear incremental.

[1] Balanced Resonate-and-Fire Neurons[C], ICML 2024.

[2] Lightweight Streaming Keyword Spotting System Based on Resonate-and-Fire Neurons[J], IEEE TCDS 2025.

Questions for authors
  1. Is the threshold in Eq. (6) fixed after initialization or updated as the trainable parameters β\beta and ω\omega change? If updated, are gradients propagated through its dependence on these parameters?

  2. How does the proposed threshold formulation compare with tuned fixed thresholds or learnable thresholds under identical architectures, initialization, and training budgets?

  3. Section 5.1 describes DEAP as a four-class task, whereas Table 2 reports separate arousal, valence, and dominance results. Could the authors clarify the task definition and how these metrics relate to the four-class setting?

Reviewer_6unhOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a Resonate-and-Fire neuron with exact discrete-time dynamics, threshold scaling, and data-informed initialization. It incorporates recurrent RF neurons into SS-EEGNet, a spiking model intended to process EEG sequentially for emotion recognition. The neuron is evaluated on S/PS-MNIST, N-MNIST, and SHD; the EEG model is evaluated on subject-dependent DEAP and SEED-VII tasks. The paper also reports an early-decision analysis and simulator-based energy estimates.

Critical strengths

The analysis of RF neuron parametrization is a useful technical contribution. Exact discretization and threshold scaling provide a more principled basis for initialization and sparse firing. Processing EEG one sample at a time is also a potentially useful direction for reducing decision latency.

Critical weaknesses
  • The EEG evaluation is insufficient for the paper’s central claim. All reported EEG results are subject-dependent; no leave-one-subject-out (LOSO) or other held-out-subject evaluation is provided. Table 2 includes too few competitive EEG baselines. At a minimum, the authors should compare with established methods such as EEG-Conformer and TSception under matched inputs, labels, preprocessing, trial-level splits, and metrics. Without these comparisons, it is difficult to assess the model’s competitiveness or its generalization beyond a calibrated individual.

  • The real-time EEG and efficiency claims lack direct system-level validation. The MNIST and SHD experiments are useful for assessing the proposed neuron, but cannot validate an EEG emotion-recognition system. The EEG experiments use segmented recordings, while the time-to-decision analysis is retrospective; the paper acknowledges that genuine streaming deployment remains unvalidated. Energy is estimated using a simulator rather than measured for the complete EEG pipeline on a target device. A continuous EEG acquisition experiment with online predictions would provide more direct evidence. In addition, SS-EEGNet changes EEGNet’s input representation and temporal processing at the same time, without ablations that isolate the effects of those changes.

Questions for authors
  • Was any LOSO or other held-out-subject evaluation performed on DEAP or SEED-VII? Please specify whether the reported splits were made at the trial or window level, and whether all windows from a trial were kept in the same partition.

  • Can the authors strengthen the EEG evaluation by comparing SS-EEGNet with established EEG methods, at least EEG-Conformer and TSception? These baselines should be evaluated using the same data splits, preprocessing, and metrics, with mean performance and variation across subjects reported.

  • Can the authors evaluate SS-EEGNet on a continuous EEG stream, preferably from a live acquisition setup, and report when predictions become available, whether they remain stable, and the end-to-end latency including signal acquisition and preprocessing?

Comments on the paper

1 comment
Area_Chair_tVcROpenReview ↗

Please Engage in the Author–Reviewer Discussion

Dear Reviewers,

As the author–reviewer discussion period is approaching its deadline (November 18, AoE), I would like to kindly remind you to review the authors' rebuttals and actively engage in the discussion as soon as possible.

Please address any outstanding questions or concerns and update your assessments and ratings if appropriate.

Thank you very much for your time and efforts!

Best regards, Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.