Online Inference of Human Intention as a Latent Control State from Single-Trial EEG

• Many BCIs decode externally imposed signals, while intention inference must distinguish stimuli that are merely perceived from those evaluated as goal-relevant.The introduction also notes that existing object-based approaches remain tied to object-category structure.Source: §I · p. 2
• Single-trial P3b responses vary substantially, while fuzzy learning represents uncertain neural states through graded memberships rather than hard assignments.The introduction presents fuzzy modeling as a way to accommodate temporal variability and overlapping neural responses.Source: §I · p. 2
• The paradigm cues an abstract goal, varies object identity independently, and asks participants to evaluate each stimulus internally for goal relevance.This design aims to make neural differences primarily reflect internal evaluation rather than stimulus properties or engineered mediating signals.Source: §I · p. 2
• Fuzz-PLI learns compact spatiotemporal prototypes for intention-relevant and irrelevant evaluations, assigning graded single-trial memberships with interpretable rules and prototypes.Its TSK-inspired softmax-weighted firing strength is presented as improving prototype contributions under neural variability.Source: §I · p. 2
• Offline within-subject decoding achieved 93.22%±3.21% accuracy, while an online setting validated real-time feasibility with average decoding accuracy of 70.11% ± 10.87%.Source: §I · p. 3
• The findings support inferring intention as a latent evaluative state, shifting the BCI focus from what users see toward what they intend the system to act upon.Source: §I · p. 3

A. Task Definition

• Intention is a category-agnostic evaluation of whether a stimulus is relevant to the observer’s current goal, not recognition of a specific object class.Each stimulus is labeled as intention-relevant (matched) or intention-irrelevant (mismatched).Source: A. Task Definition · p. 3
• Each trial is represented by a multichannel EEG time series and labeled as matched or mismatched for a subject-specific intention estimator.The estimator maps the trial to an estimated posterior probability of intention relevance.Source: A. Task Definition · p. 3
• The paradigm cues a target category before presenting objects, which participants judge for a match to that category.The figure depicts offline trials without system decisions and online trials with a displayed predicted-target or predicted-non-target frame.Source: Figure 3 · p. 4
• A binary intention decision is obtained by thresholding the estimated probability, with τ set to 0.5 unless otherwise specified.The output is interpreted as the posterior probability that the trial was evaluated as intention-relevant.Source: A. Task Definition · p. 4
• Model parameters are learned by minimizing empirical risk over the dataset using a designed training loss.The supplied passages do not specify the loss form here.Source: A. Task Definition · p. 4

B. Category-Cued Human Intention Paradigm

• The stimulus set spans four categories: animals, flowers, food, and vehicles.Five representative object classes were selected within each category.Source: 1) Image Dataset · p. 4
• Ten unique images were sampled for each of the five object classes in each category.The images were drawn from Caltech-256, following prior work.Source: 1) Image Dataset · p. 4
• The resulting balanced stimulus set contains 200 images and was designed to control low-level visual variability while preserving semantic structure.The total is 4 categories × 5 objects × 10 images.Source: 1) Image Dataset · p. 4

2) Trial Structure:

• Object identity and location are randomized independently, while target relevance is determined dynamically by the current category cue.Intention is evaluated for each stimulus, not at the sequence level, allowing inference independent of sequence structure.Source: 2) Trial Structure · p. 4

3) Task Instructions:

• Participants covertly judged whether each object was goal-relevant and were instructed to maintain fixation and minimize eye and muscle activity.No overt motor responses were made.Source: 3) Task Instructions · pp. 4–5
• The accompanying figure depicts stronger posterior P3b responses for matched than mismatched objects.It shows representative midline parietal-occipital ERPs, scalp topographies, and matched-minus-mismatched difference maps.Source: Figure 4 · p. 5

C. Participants

• Ninety-seven healthy adults participated offline and nine completed online; six took part in both experiments.Participants reported normal or corrected-to-normal vision and no neurological or psychiatric disorder history.Source: C. Participants · p. 5

D. EEG Signal Preprocessing

• EEG was recorded with a 128-channel system at 1000 Hz and filtered from 0.5 to 40 Hz before stimulus-locked epoching.Epochs spanned −200 to 800 ms, were baseline-corrected over −200 to 0 ms, and were re-referenced to the common average.Source: D. EEG Signal Preprocessing · p. 5

E. Neuroscientific Analysis

• Offline EEG underwent artifact rejection, including ICA ocular-component removal and exclusion of trials exceeding ±100 µV amplitude or variance thresholds.Residual muscle artifacts, abrupt discontinuities, and abnormal channel-wise variance were also grounds for exclusion.Source: E. Neuroscientific Analysis · p. 5
• ERP amplitudes were quantified at PO3, PO4, and PPOz from 300–600 ms, with scalp maps and matched-minus-mismatched differences examined across late windows.Paired t-tests at each electrode and window used FDR correction for multiple comparisons.Source: E. Neuroscientific Analysis · p. 5

F. Model Structure

• Fuzz-PLI uses the same convolutional encoder, interpretable fuzzy prototype module, and lightweight classification head for offline analysis and online validation.The architecture supports real-time inference without architectural modification.Source: F. Model Structure · p. 6

1) Latent Feature Extraction:

• Each EEG epoch is mapped to a compact structured latent representation, with latent units serving as antecedents for fuzzy inference.The EEGNet-inspired encoder uses temporal, depthwise spatial, and separable convolution blocks; its parameters are shared across trials within a subject and jointly optimized with fuzzy inference.Source: 1) Latent Feature Extraction · p. 6

2) Fuzzy Prototype Modulation in Fuzzy Prototype-Based Latent Intention (Fuzz-PLI):

• Fuzz-PLI performs rule-conditioned modulation of latent representations, using vector-valued fuzzy consequents rather than conventional scalar consequents.This design supports graded, interpretable intention reasoning and selectively amplifies intention-consistent neural patterns.Source: 2) Fuzzy Prototype Modulation in Fuzzy… · pp. 6–7
• Fuzzy rules operate over distributed latent representations, while their prototype centers and widths are parameterized in raw EEG and passed through the shared encoder.Gaussian antecedent memberships quantify compatibility between trial-specific latent patterns and rule prototypes.Source: 2) Fuzzy Prototype Modulation in Fuzzy… · p. 6
• Fuzzy modulation begins after a warm-up stage, and an inter-prototype regularizer penalizes average pairwise cosine similarity to encourage distinct rules.The warm-up trains the encoder and classification head without fuzzy inference to avoid early domination by unstable rules.Source: 2) Fuzzy Prototype Modulation in Fuzzy… · p. 7

3) Training and Evaluation Protocol:

• For each participant, 400 trials were split 80%/20% into balanced training and validation sets, repeated across five random seeds.Source: 3) Training and Evaluation Protocol · p. 7
• The final objective combined binary crossentropy classification loss with prototype-diversity regularization.Source: 3) Training and Evaluation Protocol · p. 7
• Evaluation used within-subject classification accuracy, with λdiv = 0.3 and group-level statistics computed across participants.Source: 3) Training and Evaluation Protocol · p. 7

G. Baseline Models

• EEGNet uses depthwise and separable convolutions for compact spatiotemporal modeling.Source: G. Baseline Models · p. 7
• EEG Conformer combines convolutional front-ends with self-attention to capture long-range dependencies.Source: G. Baseline Models · p. 7
• EEGformer uses a hierarchical transformer architecture for regional and temporal feature modeling.Source: G. Baseline Models · p. 7

N STM

• EEGNet uses depthwise and separable convolutions, EEG Conformer combines convolutional layers with self-attention, and EEGformer uses hierarchical regional and temporal modeling.Baseline inputs, preprocessing, and training configurations were matched across models.Source: N STM · p. 7

A. P3b Responses to Matched and Mismatched Object

• Matched stimuli elicited stronger posterior P3b activity than mismatched stimuli, with a significant difference in the 300–600 ms window and a widespread centro-parietal positivity.The maximal paired contrast was t(96) = 6.59, p < 0.001, Cohen’s d = 0.39; activity was bilateral and spatially extended.Source: A. P3b Responses to Matched and… · p. 7

B. Model Comparison

• Low-frequency bands yielded consistently higher decoding performance, while higher-frequency bands showed marked degradation.Source: Figure 6 · p. 8
• Accuracy plateaued with a small number of fuzzy rules, with no substantial benefit from increasing the rule count beyond that range.Source: Figure 7 · p. 8
• 93.22% ± 3.21% accuracy: Fuzz-PLI significantly outperformed EEGNet (85.190% ± 6.390%), EEGConformer (91.47% ± 4.39%), and EEGformer (91.17% ± 4.49%).Fuzz-PLI had 1.46M parameters; paired tests found significant differences against all three baselines.Source: B. Model Comparison · pp. 7–8

C. Ablation Study

• 0.5–8 Hz filtering yielded 93.22% ± 3.21% accuracy; higher-frequency bands significantly degraded performance, while broader low-frequency ranges did not differ significantly.Source: C. Ablation Study · p. 8
• Accuracy plateaued with at least two fuzzy rules, with differences below 0.53%; performance peaked at six rules.The results support a parsimonious fuzzy structure; the study’s overall ablation conclusion also highlights low-frequency modeling and soft rule aggregation.Source: C. Ablation Study · pp. 8–9
• 93.22% ± 3.21% versus 88.84% ± 5.42% accuracy: softmax-weighted firing strength outperformed max-based aggregation under the same six-rule, 0.5–8 Hz settings.The difference was significant, t(96) = 9.516, p < 0.001.Source: C. Ablation Study · p. 8
• 93.22% ± 3.21% versus 88.84% ± 5.42% accuracy: removing the fuzzy prototype module with the backbone fixed reduced performance.The difference was significant, t(96) = 9.516, p < 0.001, Cohen’s d = 0.983.Source: C. Ablation Study · p. 9

A. Experimental Setup

• Online evaluation used nine participants, training on each participant’s first session and testing the second without further tuning or recalibration.Each session contained 400 trials; only Fuzz-PLI generated real-time feedback, while baseline comparisons were evaluated offline.Source: A. Experimental Setup · p. 9

B. Online System

• Online EEG was buffered continuously and segmented into short trigger-marked epochs for real-time processing by the trained Fuzz-PLI model.Epochs spanned approximately -0.2 to 0.8 seconds.Source: B. Online System · p. 9

C. Results

• 70.11% ± 10.87% online decoding accuracy; Fuzz-PLI outperformed EEGNet, EEGConformer, and EEGFormer.Fig. 8 displays participant means, one-standard-deviation error bars, individual results, and paired participant lines.Source: Figure 8 · p. 9, C. Results · p. 9
• TSK-style inference achieved 69.17% ± 10.19% online accuracy, significantly exceeding the max-based Mamdani-style strategy.The comparison used identical experimental settings.Source: C. Results · p. 9
• Online degradation is framed as a challenge of real-time latent-state inference under distribution shift and uncertainty, rather than a deficiency of a particular architecture.Fuzz-PLI nevertheless retained superior performance to the baselines under online conditions.Source: C. Results · p. 9

A. P3b as an Observable of a Latent Intention-Related State

• P3b is treated as an observable of goal-dependent evaluation, not intention itself, reflecting whether sensory evidence is committed to action-relevant processing.The discussion distinguishes post-perceptual evaluation and intention realization from awareness or intention initiation.Source: A. P3b as an Observable of a Latent… · pp. 9–10

B. Interpretability of the Fuzzy Prototype-Based Model

• Aligned fuzzy rule prototypes showed substantial cross-subject consistency despite within-session training, suggesting shared evaluative neural patterns.The authors interpret this as evidence against prototypes merely reflecting idiosyncratic noise or session-specific artifacts.Source: B. Interpretability of the Fuzzy… · p. 10
• Different prototypes captured distinct spatiotemporal patterns, including a parieto-occipital POOz response at 200–400 ms resembling canonical P3b.Other rules represented complementary temporal profiles and spatial distributions, partitioning neural activity into multiple functionally meaningful components.Source: B. Interpretability of the Fuzzy… · p. 10

C. System-Level Interpretation and Cybernetic Implications

• The paper frames intention as a latent state with posterior P3b as its neural observation channel, and interprets compact prototypes as capturing low-dimensional intention dynamics.It also reports that online validation supports real-time deployment and that Fuzz-PLI outperforms deep baselines.Source: C. System-Level Interpretation and… · p. 10

1) Online versus Offline Performance:

• Online decoding accuracy was lower than offline performance, but Fuzz-PLI maintained superior performance to all baseline approaches.The discussion links the setting difference to offline non-causal preprocessing and artifact removal versus online latency and causal constraints.Source: 1) Online versus Offline Performance · pp. 10–11

2) System-level Sources of Online Performance Degradation:

• The online-offline gap is attributed to session non-stationarity, latency and marker-timing jitter, and possible overfitting to session-specific patterns.The authors state that these factors collectively account for the observed degradation.Source: 2) System-level Sources of Online… · p. 11

E. Limitations and Future Directions

• The study is limited to within-subject inference; cross-subject adaptation and long-term deployment remain future work.Other stated directions include continuous latent-process modeling and adaptive or hierarchical rule learning.Source: E. Limitations and Future Directions · p. 11