acceptodds
Under review as a conference paper at ICLR 2027

Learning Event Ownership via Finite-State Action-Path Marginalization for Irreversible Command Decoding from Two-Channel EOG

Abstract

Continuous facial-gesture control from few-channel electrooculography (EOG) requires distinguishing both action classes and event identities: incoming evidence may continue an existing event or begin a new one. Frame-wise classifiers with fixed post-trigger cooldowns do not explicitly model this distinction and can suppress closely spaced commands. We introduce Action-Constrained Event Decoding (ACED), which separates action classification from event ownership. A learned ownership policy selects among accumulating evidence, emitting a command, and re-arming, replacing fixed cooldown timers. For ambiguous action prefixes, supervision targets an intermediate class and defers the final label until the command grammar permits disambiguation. Training exactly marginalizes over all supervision-compatible paths satisfying an event grammar that requires one emission per supervised event and excludes premature triggers. At inference, the decoder operates causally, never revises past outputs, and guarantees at most one emission per inferred event identity. Experiments show that ACED recovers closely spaced commands suppressed by cooldown baselines and reduces false-trigger rates. The same decoding logic transfers to public datasets with different action taxonomies through task-specific retraining. We further quantify the effects of cross-wear signal drift and address it with robust training recipes.

Primary area:
applications to neuroscience & cognitive science
Keywords:
electrooculography (EOG), electrophysiology, biomedical signal processing, event detection, time series

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_f5xyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates causal and irreversible command decoding from two-channel electrooculography (EOG), focusing on the challenge of distinguishing new gesture events from continuing evidence of previously detected events. The authors propose Action-Constrained Event Decoding (ACED), which explicitly models event ownership through two approaches: ACED-P, a learnable finite-state policy trained by exact marginalization over supervision-compatible action paths, and ACED-D, a rule-based decoder that defers ambiguous double/triple-blink classification until additional evidence becomes available.

The method is evaluated on a dataset containing 5 participants, 30 recording sessions, and 1,442 annotated gesture windows, using leave-one-wear-out (LOWO), leave-one-subject-out (LOSO), and public-dataset transfer experiments. Under identical classifier outputs, ACED-D improves the LOWO first-correct fraction from 0.641 to 0.715 compared with hysteresis, while ACED-P13 reduces false emissions at a comparable detection rate. The results demonstrate trade-offs between command recognition accuracy, false-trigger suppression, and decision latency, although neither proposed decoder consistently outperforms the baselines across all metrics.

Critical strengths
  1. Meaningful problem formulation: The paper explicitly addresses event ownership in continuous EOG command decoding, distinguishing new gestures from ongoing events. The proposed deferred-resolution mechanism improves consecutive-pair recognition from 0.303 to 0.478 compared with hysteresis on identical classifier outputs, demonstrating the practical relevance of this problem formulation.
  2. Rigorous and transparent experimental evaluation: The paper compares decoders using identical classifier outputs and includes LOWO/LOSO evaluations, false-trigger analysis, and event-time versus commit-time measurements. This controlled evaluation and the disclosure of unsuccessful results provide valuable insights into the trade-offs of causal command decoding, strengthening the credibility of the empirical analysis.
Critical weaknesses
  1. Limited evidence for the core methodological contribution. Although the paper emphasizes learnable event ownership through finite-state action-path marginalization, the best performance is achieved by the hand-designed ACED-D rather than the learned ACED-P. In Table 1, ACED-P18 even underperforms the hysteresis baseline, raising concerns about the effectiveness and necessity of the proposed learning approach.
  2. Inconsistent evaluation of irreversible command decoding. The primary results are evaluated using event time rather than actual command commitment time. As reported in Section 5.3, ACED-D's first-correct fraction drops from 0.715 to 0.542 when evaluated at commitment time, substantially weakening its claimed advantage for real-time command execution.
Questions for authors
  1. Evaluation Protocol: Why is 30-fold leave-one-wear-out (LOWO) selected as the primary evaluation protocol rather than using fewer folds with multiple held-out recordings? Given the substantial performance degradation under LOSO, which evaluation setting better reflects practical deployment?
  2. Comparison with Lightweight Neural Networks: Why does the paper not compare against lightweight temporal neural networks commonly used for continuous EMG/EEG signal recognition, such as sliding-window-based 1D-CNNs, TCNs, or GRUs? These methods can directly model temporal dependencies and produce continuous predictions. Could the authors clarify what advantages the proposed event ownership mechanism provides over such approaches?
  3. Effectiveness of the Core Learning Method: The main methodological contribution is finite-state action-path marginalization, yet the learned ACED-P performs worse than the hand-designed ACED-D, and ACED-P18 underperforms hysteresis in Table 1. What evidence demonstrates that the proposed learning-based formulation provides advantages over simpler rule-based decoding?
  4. Event Time versus Commitment Time: Section 5.3 reports that ACED-D's first-correct fraction decreases from 0.715 to 0.542 when evaluated at commitment time. Since commands are irreversible, why is event-time accuracy selected as the primary evaluation metric?
  5. Potential Evaluation Bias: Section 5.6 acknowledges 17 development iterations involving the outer LOWO evaluation results. Could the authors clarify how these iterations influenced the final decoder design and how potential adaptive overfitting was addressed?
  6. Cross-Subject Generalization: Under LOSO evaluation, ACED-D's first-correct fraction drops to 0.522, and outside emissions exceed the predefined guardrail. How do the authors interpret these results regarding the practical applicability of the proposed framework?
  7. Dataset Collection and Scale: The proposed dataset contains only five participants, 30 recording sessions, and approximately 2.54 hours of valid EOG signals. What were the main challenges in collecting and annotating the data, and why was the dataset limited to such a small number of participants? How might this limited scale affect the reliability and generalizability of the conclusions?
Reviewer_NMQ2OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper addresses a novel problem in few-channel electrooculography (EOG)-based facial-gesture control: deciding whether incoming evidence belongs to an existing event or marks the start of a new one. To this end, the authors propose Action-Constrained Event Decoding (ACED), which makes this ownership decision explicitly. The authors also collect a dataset from 5 participants performing blinking and jaw-clenching tasks. The experiments show that ACED captures closely spaced commands while lowering the false trigger rate, and they also evaluate the approach on two public datasets.

Critical strengths

Clear problem motivation and formulation: The paper addresses a novel and underexplored problem in event-level decoding and formulates it explicitly rather than relying solely on fixed timeouts.

Critical weaknesses

Limited evaluation and generalizability: The results are mostly ablations of the decoder’s own rules. In the main experiments, all reference decoders are rule-based variants of the proposed approach, with no comparison to related threshold-based methods for distinguishing single and multiple blinks, such as interval-based grouping. In addition, the analyses use data from only 5 participants. Although the paper provides results on the OPPORTUNITY and PhysioMotion datasets, these do not include cases where consecutive movements must be grouped, such as deciding whether blinks form a double or triple blink. This makes it difficult to judge whether the added complexity is needed and how well the results generalize.

Questions for authors
  • How were outliers handled during data preprocessing, such as unintentional blinks or trials where participants did not follow the instructions?
  • ACED-D relies on many fixed parameters. How were these values selected, and was any of the data used for this also used for testing?
  • What is ACED-simple in Section 5.5, and how does it differ from ACED-D and ACED-P?
  • Although the authors state that there is no directly comparable prior work, rule and threshold based approaches for distinguishing single and multiple blinks are closely related. How does the proposed approach compare to a simple interval-based grouping baseline, where individual blinks within a fixed time window are counted and labeled as double or triple?
Reviewer_2Xx1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript proposes Action-Constrained Event Decoding (ACED), a framework for continuous facial-gesture command recognition using two-channel electrooculography (EOG) signals. The central idea is to explicitly model event ownership, distinguishing whether incoming evidence belongs to an existing gesture or indicates a new command. The framework introduces two decoding strategies: ACED-P, a learned finite-state ownership policy trained through exact marginalization over supervision-compatible action paths, and ACED-D, a rule-based decoder that delays ambiguous blink classification until additional causal evidence becomes available.

Experiments on a newly collected EOG dataset and two public datasets demonstrate improvements in selected event-recognition metrics over fixed-cooldown and hysteresis-based decoders. The paper also investigates output timing, false activations, cross-wear generalization, and decoder development history.

The problem is practically relevant, and explicitly modeling event ownership is an interesting direction. However, the manuscript has several fundamental limitations concerning the relationship between the proposed learning framework and the reported improvements, the interpretation of irreversible decoding, the reliability of the evaluation protocol, and the generalizability of the findings. These issues substantially weaken the evidence supporting the proposed contribution.

Critical strengths

see Questions For Authors

Critical weaknesses

see Questions For Authors

Questions for authors

The manuscript addresses an important problem in continuous EOG command decoding: deciding whether new evidence belongs to an existing gesture or a new event. Separating event ownership from classification is an interesting idea, especially for closely spaced commands and ambiguous blink patterns. Comparisons on identical per-fold logits and the detailed output-timing audit are useful strengths. However, several issues weaken the central claims.

  1. Limited Evidence for the Learned Ownership Mechanism

The principal contribution is an ownership policy trained by exact marginalization over finite-state action paths. Yet the best result is from the largely rule-based ACED-D (first-correct 0.715), while ACED-P13 and ACED-P18 achieve 0.644 and 0.621, compared with 0.641 for hysteresis (Table 1). The main improvement therefore seems attributable to hand-designed deferred resolution, not learned path marginalization. What does the learned policy add over a tuned deterministic state machine with equivalent states? Why does ACED-P18 underperform both ACED-D and hysteresis? A controlled comparison is needed to isolate the learning contribution.

  1. Event Time Versus Actual Command Availability

The distinction between event time τ and commit time λ is helpful, but raises a fundamental concern for irreversible command decoding. Section 5.3 reports that scoring ACED-D at λ instead of τ reduces first-correct from 0.715 to 0.542 and raises outside emissions from 899 to 1236. Since blink labels can remain unavailable at τ, why is τ-based accuracy the primary metric? Please compare all decoders using the time when their final commands become actionable, with decision latency and false activations. This difference materially changes the interpretation of the main result.

  1. High False-Activation Rates and Unmatched Operating Points

ACED-D produces 899 outside emissions, or 490 per background hour, versus 752 for hysteresis (Table 1). This appears high for a practical command interface. Although ACED-P13 reduces the count to 649, Section 5.2 reports that utility-matched hysteresis achieves 632, indicating a substantial calibration effect. What application requirement motivates the guardrail of 900? Please report recall-versus-false-activation tradeoffs under comparable budgets, ideally with actionable decision latency. The practical advantage is unclear without such matching.

  1. Potential Bias from Repeated Outer Evaluation

The nested LOWO protocol and sealed predictions improve transparency, but Section 5.6 discloses 17 outer development boards used during ACED-D refinement. Repeated feedback from outer-fold results can produce optimistic estimates despite nested threshold selection. Which rules changed after examining these results? Did the baselines receive comparable design refinement? Can the fixed final decoder be tested on independent recordings not consulted during development?

  1. Confounded Ablation Comparisons

The ablation study is useful but does not fully isolate the proposed mechanisms. Appendix B states that ACED-P13 uses App-cue constraints whereas ACED-P18 uses manual-window constraints; P13 also allows optional inference-time blink deferral. Thus, their performance difference cannot be ascribed solely to pending states. Please compare ownership, pending resolution, and path marginalization while holding supervision, scorer outputs, inference settings, and threshold selection fixed.

  1. Weak Cross-Subject Generalization and Limited Event Diversity

The dataset includes five participants and scripted cues. Under LOSO, ACED-D drops to 0.522 first-correct and produces 1044 outside emissions, exceeding the paper's guardrail (Table 4). Same-class consecutive gestures are also sparse, although recognizing repeated instances of the same action is an especially direct test of ownership. Can the authors evaluate same-class repeats, irregular or shorter gaps, and self-paced operation on unseen participants? Current conclusions should be limited to the evaluated acquisition conditions.

  1. Transfer Evidence Needs Clearer Controls

Testing PhysioMotion and OPPORTUNITY broadens the evaluation, but the results are mixed. On PhysioMotion, ACED covers 33.9% of blocks under the 10/h budget, versus 65.9% for CNN1D; its coverage advantage emerges under the 60/h budget (83.8%). On OPPORTUNITY, a task-adapted scorer and ACED-simple are used. Which decoder components are unchanged, and are scorer quality and operating points matched across baselines? These details are necessary to attribute transfer gains to the ownership mechanism.

Overall, the problem formulation and careful diagnostic reporting are promising, but the evidence does not yet demonstrate that learned path marginalization drives the principal improvements or that the gains persist when commands are evaluated at actionable times.

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

This paper proposes Action-Constrained Event Decoding (ACED) for continuous gesture recognition from two-channel EOG signals. The authors introduce a learned event ownership policy using finite-state action-path marginalization (ACED-P) and a deferred resolution mechanism for ambiguous blink patterns (ACED-D). The methods aim to reduce duplicate commands, recover closely spaced events, and improve command reliability.

Critical strengths
  1. The paper addresses an interesting and practical problem: distinguishing new events from continuing evidence in continuous signal decoding.
  2. The evaluation is carefully designed, with identical classifier outputs across decoders and clear ablations of different decoding mechanisms.
  3. The deferred resolution mechanism demonstrates meaningful improvements in event recognition, particularly for closely spaced commands. The authors also provide a useful analysis of event time versus commit time.
Critical weaknesses

The strongest accuracy improvement comes from ACED-D, which relies on hand-designed deferred resolution rules. ACED-P13 only marginally improves first-correct accuracy over hysteresis (0.644 vs. 0.641), while utility-matched hysteresis achieves a similar outside-emission count. Furthermore, ACED-P18 performs worse than both methods. These results do not clearly establish the benefit of finite-state path marginalization. A comparison with hysteresis under the same deferred resolution mechanism would help isolate the contribution of learned ownership.

Questions for authors

Could the authors compare the learned ownership policy with a well-calibrated hysteresis decoder under the same deferred resolution mechanism? This would help clarify whether finite-state path marginalization provides benefits beyond the hand-designed decoding rules.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.