acceptodds
Under review as a conference paper at ICLR 2027

Certainty-Guided Feature Separation for Early Decision in Spiking Keyword Spotting

Abstract

Keyword spotting (KWS) requires low latency, low energy consumption, and high recognition accuracy, making efficient early decision-making particularly important for resource-constrained edge devices. Existing early-decision methods for spiking neural networks (SNNs) mainly supervise cumulative outputs across time steps, while feature-level approaches often enforce uniform temporal separation and overlook the progressively increasing discriminative information in speech. In this work, we propose a certainty-guided spiking feature separation framework that dynamically estimates the informativeness of each time step from the SNN's predictive distribution and uses it to weight intermediate feature separation. By applying a certainty-guided feature separation objective combined with a dynamic training schedule, unreliable separation are suppressed during early training and discriminative representation learning are strengthened as the classifier becomes more reliable. Experiments on GSC V2 and MSWC Microset datasets demonstrate that our method improves early decision accuracy while maintaining or improving final recognition accuracy, achieving better performance across multiple time steps. The proposed framework provides a lightweight and inference-free approach to learning temporally adaptive spiking representations, offering a promising direction for low-latency and energy-efficient KWS on edge devices.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Spiking Neural Networks, Keyword Spotting

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

This paper proposes a certainty-guided spiking feature separation framework for early keyword spotting. It derives predictive certainty from normalized entropy and uses it to weight feature separation at each time step, attracting representations toward their class prototypes while enforcing separation from other classes. A weight schedule progressively introduces feature supervision during training. Experiments on GSC V2 and MSWC Microset compare several training objectives and scheduling strategies, showing slightly higher mean accuracy than TET at the reported decision points. However, the main improvements are modest, actual early stopping and energy reduction are not directly evaluated, and several aspects of the loss formulation and implementation require clarification.

Critical strengths

The paper addresses a reasonable concern: uniformly enforcing feature separation when early speech observations contain insufficient evidence may harm representation learning. Using predictive certainty to adjust supervision strength, together with suppressing the auxiliary objective during early training, provides an understandable framework. The auxiliary mechanism is used only during training and requires no additional inference modules. The experiments compare multiple training objectives using the same network architecture and report means and standard deviations across three random seeds, providing stronger evidence than single-run comparisons. Comparisons against uniform separation and alternative weight schedules also offer initial support for the proposed components.

Critical weaknesses

Comment 1: The main improvements are modest, and one performance claim is inaccurate. Compared with TET, Acc@50 improves by only 0.16 percentage points on GSC V2 and 0.12 percentage points on MSWC. Some improvements are comparable to the reported variation across runs and require stronger statistical evidence. Moreover, the text claims the highest accuracy at every evaluated time step on both datasets, but Table 2 reports an MSWC Acc@75 of 90.10%, below the spike-rate baseline at 90.77%. This statement should be corrected, and additional repeated experiments or paired statistical analysis should establish the consistency of the gains.

Comment 2: Accuracy at fixed time points does not establish practical early stopping or energy savings. The evaluation primarily reports classification accuracy at four fixed time steps, without an actual stopping rule, average exit time, or latency comparisons at matched target accuracy. The results therefore support improved prefix classification, but do not directly demonstrate earlier reliable stopping or lower inference energy. Avoiding additional inference modules does not automatically reduce computation. Accuracy versus average decision-time curves, together with corresponding computation or energy estimates, would better substantiate the practical claims.

Comment 3: Predictive certainty is not directly validated as a proxy for informative and reliable observations. Low predictive entropy does not guarantee correctness or establish that the observed speech prefix contains sufficient class information. Although the paper acknowledges this limitation, it does not analyze the relationship between certainty, prediction correctness, and prefix length, or examine confidently incorrect predictions. The authors should also clarify whether gradients are stopped through {\mathrm{D}}{\mathrm{t}}\ . If gradients propagate through it, minimizing {\mathrm{D}}{\mathrm{t}}{\mathcal{L}}_{\mathrm{sep}}\ can also encourage lower certainty rather than solely improving feature separation. Reliability analysis and an ablation of gradient stopping would strengthen the proposed interpretation.

Comment 4: The loss formulation is inconsistent, and important training settings are missing. The weighted separation objective in Section 3.3 includes 1/T, whereas the total objective in Section 3.5 uses an unnormalized temporal sum, changing the auxiliary loss scale by a factor of (T). The text also introduces a maximum schedule weight {\mathrm{\lambda }}{\max}, but the sigmoid formula does not explicitly include this parameter. Under the stated formula, the classification-loss coefficient decreases as λ(e) approaches 1. The implementation should be clarified, and settings for α, margin (m), schedule parameters (\mathrm{k},{\mathrm{e}}{0}), prototype initialization, and network architecture should be provided to make the training behavior assessable and the results reproducible.

Comment 5: Terminology and cross-references need correction. Figure 2(b) is titled "Uncertainty-Guided Feature Separation," whereas the main text, caption, and internal figure label use certainty, and the separation loss is weighted by {\mathrm{D}}{\mathrm{t}}=1-{\mathrm{U}}{\mathrm{t}}. The title should be standardized to "Certainty-Guided Feature Separation." Section 4.3.1 also contains an unresolved "Table ??" reference, which should be corrected in the final version.

Questions for authors

See Weaknesses.

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

This paper proposes a certainty-guided feature separation framework for early-decision keyword spotting with spiking neural networks. The method introduces learnable class prototypes and a feature separation loss that encourages intra-class compactness and inter-class separation. It uses predictive certainty derived from normalized output entropy to adaptively weight feature separation across time steps, while a sigmoid schedule gradually increases its contribution during training. Experiments on GSC V2 and MSWC Microset show improved early-decision accuracy and competitive final accuracy compared with existing methods. Ablation studies further support the effectiveness of certainty weighting and dynamic scheduling, with no additional inference overhead.

Critical strengths
  1. The method adapts feature-level supervision based on predictive certainty and training progress, accounting for varying information across speech prefixes. Its auxiliary components introduce no additional inference overhead.

  2. Across three random seeds, the method consistently outperforms TET at all evaluated time steps, with gains of 0.16–0.59 percentage points on GSC V2 and 0.12–0.47 on MSWC Microset. Ablation studies provide initial evidence for the effectiveness of certainty weighting and dynamic scheduling.

Critical weaknesses
  1. The claimed latency and energy-efficiency benefits lack direct experimental validation. The paper emphasizes low-latency and energy-efficient keyword spotting, but the experiments only report classification accuracy at fixed time steps. While these results demonstrate improved early-prefix recognition, they do not directly establish reductions in actual decision latency or energy consumption. It would be helpful to evaluate practical early-stopping performance, such as average decision time under confidence-based stopping criteria, and provide computational or energy measurements to better support these claims.

  2. The independent contribution of the feature separation loss is not sufficiently evaluated. The current ablation studies compare certainty-weighted separation with uniform separation and examine different scheduling strategies, but do not explicitly evaluate the proposed framework with the entire feature separation loss removed. Although the paper includes comparisons with existing training methods, these do not clearly isolate the contribution of the proposed loss under otherwise identical conditions. An additional ablation retaining only the classification objective would help clarify how much of the performance improvement comes from feature separation itself, independently of certainty weighting and dynamic scheduling.

Questions for authors
  1. The current experiments cover GSC V2 and MSWC Microset. Could the authors also evaluate the method on SC-100, the dataset introduced in the ED-sKWS cited in the article?
  2. The current evaluation is limited to keyword spotting. Could the authors evaluate the proposed framework on another sequential classification task that requires early decisions?
Reviewer_SzYrOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes an SNN training framework for early speech classification. However, its novelty over existing temporal regularization methods is limited. The architecture, implementation details, and scope of spike-driven computation are insufficiently specified, hindering understanding and reproducibility. The writing and figures also require improvement, as the figures lack clear visual organization and do not effectively communicate the key mechanisms.

Critical strengths

None

Critical weaknesses
  1. The method applies established prototype-based feature separation, entropy weighting, and loss scheduling across SNN time steps while retaining standard LIF neurons and surrogate-gradient training. Its novelty beyond generic temporal regularization is limited, with no clear mechanisms tailored to spike discreteness, membrane-potential dynamics, or temporal learning.

  2. The SNN architecture is not clearly specified, limiting reproducibility and assessment of its spike-driven computation. The execution frequency and overhead of probability readout and any entropy-based exit decisions also remain unclear.

  3. The separation loss encourages intra-class compactness and inter-class separation but does not directly optimize exit time or decision reliability.

Questions for authors
  1. What SNN-specific mechanisms distinguish this method from existing prototype learning, confidence weighting, and loss scheduling? Would the same objective yield similar gains in causal ANNs or recurrent networks with matched input prefixes and comparable capacity?

  2. How is the class probability distribution (p_t) computed from the spike outputs? Does this involve floating-point operations? If performed during inference, is the associated overhead included in the efficiency evaluation?

  3. Which modules receive binary spikes, and which process real-valued representations? Does the implementation exploit spike sparsity to skip computation, and do efficiency measurements include acoustic encoding, classification readout, and exit decisions?

  4. Under the same exit policy, can the authors separately ablate prototype separation, certainty weighting, and loss scheduling? Please report accuracy versus observation length, together with average exit time and end-to-end latency at matched accuracy, to establish the practical benefits of early separation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.