{"id":"f8666c14-59e9-4902-bff8-65ed6f95f5c1","arxiv_id":"2412.14848","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A fully onboard, calibration-free hybrid electrooculography system on smart glasses classifies up to 10 eye-movement classes at 81% accuracy using a 79 kB quantized neural network.","lead":"ElectraSight is a pair of smart glasses that tracks eye movements using electrical signals from the skin and from contactless electrodes, with all processing done on the glasses. A tiny neural network classifies 10 eye movements with 81% accuracy and runs on a few milliwatts, aiming at comfortable, battery-friendly eye tracking.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 81%/92% accuracy may come from an acquisition-level split rather than the claimed subject-independent evaluation; the split description is contradictory and the user-independence claim depends on it.","rationale":"The reader's conditional verdict is appropriate, and I agree that the engineering contribution is credible. The reader highlighted both label-pipeline risk and split ambiguity; I single out the split ambiguity as the more decisive and falsifiable issue. A heuristic labeling pipeline, while imperfect, is at least anchored to an external eye tracker; a leaky split directly invalidates the headline user-independent accuracy regardless of label quality. The contradiction between Section VII-A's per-acquisition 80/20 split and Section VIII's claim of evaluation on unseen subjects is a concrete, checkable inconsistency. A strict subject-disjoint evaluation is the minimum test. If the split is already subject-disjoint, the concern is resolved and the reported numbers stand; if not, the calibration-free and user-agnostic claims are unsupported. I therefore keep the reader's CONDITIONAL verdict, conditional on clarification or release of the split metadata and a subject-disjoint re-evaluation.","tokens_in":22703,"tokens_out":3883,"duration_ms":34307,"concrete_test":"Obtain the exact split used for the final 10-class and 6-class models, or release subject-level train/test assignment. Then rerun the same training procedure with a strict leave-one-subject-out split over the 20 subjects, reporting per-subject accuracy and mean ± std. If the subject-disjoint 10-class accuracy is within about 3 points of 81%, the calibration-free claim holds; if it falls materially below 81% (e.g., below 75%), the reported accuracy is inflated by same-subject leakage and the central claim needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section VII-A states: \"Each acquisition is randomly assigned to either a training set (80% of all acquisitions) or a test set (20% of all acquisitions).\" Section VIII then claims the model is \"evaluated on subjects not encountered during training.\" These describe incompatible evaluation protocols. If the final 81%/92% numbers come from an acquisition-level split, windows from the same subject can appear in both training and test sets. Since EOG morphology and electrode placement are subject-specific, even per-window standardization does not eliminate this leakage. The abstract's \"not requiring any calibration or user-specific fine-tuning\" is only supported by a subject-disjoint split, not by a random acquisition split. No split ratio, subject count, or subject IDs are given for the final model, so the reader cannot tell which protocol produced Figures 11a/11b, Table V, or the latency results. This is the load-bearing weakness: every headline accuracy and the user-independence claim rests on the split definition.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents ElectraSight, a smart-glasses eye-tracking system that uses hybrid contact and contactless electrooculography (hEOG) with fully onboard processing on a GAP9 RISC-V coprocessor. The authors describe the hardware design, electrode placement, impedance characterization, a data-acquisition protocol with 20 subjects, and a labeling pipeline that combines logger timestamps with Pupil Labs Neon gaze data. They report a tinyML CNN that classifies 10 eye-movement classes with 81% accuracy and 6 basic classes with 92% accuracy, runs in 301 µs per inference, fits in 79 kB, and consumes 46 µJ per inference, with a total system power of 8.85 mW and an estimated three-day battery life. The paper also includes an ablation study of channel types, a window-size sweep, a quantization study, and a rolling-window latency analysis.","tokens_in":22741,"tokens_out":3893,"duration_ms":25542,"significance":"If the headline numbers are reproducible, ElectraSight is a meaningful step toward practical, non-invasive, calibration-free eye tracking for wearables. The work has several genuine strengths: the ground truth comes from an independent commercial eye tracker, the hardware is based on open-source components and is thoroughly characterized, the dataset includes 20 subjects with a defined protocol, and the quantization and deployment results are presented with concrete memory and energy figures. The main significance lies in demonstrating that hybrid contact/contactless EOG can achieve competitive classification accuracy with fully onboard tinyML inference at sub-milliwatt processing power. However, the evaluation-protocol ambiguity described below directly affects whether the central accuracy claims and the user-independence claim are supported.","major_comments":[{"comment":"The evaluation protocol for the final model is contradictory and underspecified. Section VII-A states that \"Each acquisition is randomly assigned to either a training set (80% of all acquisitions) or a test set (20% of all acquisitions)\" for the ablation study, whereas Section VIII claims that the 81% and 92% accuracies are \"evaluated on subjects not encountered during training.\" These are incompatible protocols: a random acquisition-level split can place windows from the same subject in both training and test sets, and EOG morphology and electrode placement are subject-specific, so such leakage could inflate accuracy. The paper gives no subject count, subject IDs, or split ratio for the final model, so the reader cannot determine whether Figures 11a/11b, Table V, Table VI, and the latency results come from a subject-disjoint split. This is load-bearing because the abstract's claim of \"not requiring any calibration or user-specific fine-tuning\" is only supported by subject-independent evaluation. Please specify the exact split used for the final model, report subject-level or repeated-split statistics, and if the 81%/92% numbers came from the acquisition-level split, re-run the evaluation on a subject-disjoint split and restate the claims with the new results.","section":"§VII-A and §VIII"},{"comment":"The ground-truth labeling depends on manually selected thresholds on the gaze derivative (Th_up: Elev' > 3, Th_down: Elev' < -3, Th_right: Az' > 2.5, Th_left: Az' < -2.5, Th_straight: -1 < Elev' and Az' < 1), chosen \"based on manual inspection of the gaze derivative data.\" The same thresholds determine the movement timestamps used in the latency analysis of Section VIII-C. The paper itself notes that logger labels are 4.6 times less accurate than eye-tracker labels for rapid movements, which shows that the reported accuracy is sensitive to the labeling pipeline. However, no sensitivity analysis, threshold validation, or comparison against an alternative labeling rule is provided. Since every confusion matrix and latency figure is derived from labels produced by Eq. (2), the authors should justify the thresholds, quantify their stability (e.g., by perturbing them and recomputing accuracy), or provide an independent validation of the resulting labels.","section":"§VI-A, Eq. (2)"},{"comment":"The abstract reports \"46 mJ for the tinyML inference,\" while Table VI and Section IX-B report 46 µJ. This is a factor of 1000 discrepancy in a headline energy claim. The 46 µJ figure is consistent with the stated average power of 153 mW and 301 µs execution (153 mW × 301 µs ≈ 46 µJ), so the abstract appears to contain a unit error, but as written it overstates energy consumption by three orders of magnitude. Please correct the abstract and check all other unit presentations (e.g., \"8.85 mW\" total power, \"7.75 mW\" acquisition) for consistency.","section":"Abstract and Table VI"},{"comment":"All accuracy values are reported from a single train/test split with no confidence intervals or repeated-split variability. In Table V, the differences between window sizes (e.g., 84% at 1000 ms vs. 81% at 416 ms vs. 73% at 312 ms) are presented as if they are meaningful, but with a single split and no error bars it is impossible to judge whether these differences are statistically significant. Similarly, Table VI states that accuracy \"remains constant down to 4-bit\" (81% for 8-bit and float16, 80% for 4-bit), but the 1-point gap could be noise. Given the otherwise careful hardware and deployment work, the authors should report confidence intervals, multiple splits (e.g., k-fold or repeated random splits), or per-subject accuracy distributions for at least the final model and the key ablation comparisons.","section":"Table V and Table VI"}],"minor_comments":[{"comment":"There is a typo in the acknowledgment: \"the anonymous people involved in the datasat collection\" should read \"dataset collection.\"","section":"Acknowledgments"},{"comment":"The sentence \"An affordable system based on near-infrared (NIR) LED and phototransistors is presented in [24]\" appears to cite the wrong reference. Reference [24] is an EOG processor paper (Das et al.), while the NIR system described in that paragraph corresponds to [28] (Guo et al., \"A wearable non-contact optical system...\"). Please fix the citation.","section":"§II-A3"},{"comment":"The latency definition is stated somewhat ambiguously: \"The latency for a movement is calculated if the movement is correctly predicted, and is equal to the difference between the timestamp of the end of the correctly predicted window, and the timestamp of the movement in the ground truth.\" Please clarify whether the end timestamp is the end of the window in which the movement is first detected or the end of the first window that yields the correct class; the subsequent sentence \"only 1% of the test samples are never correctly predicted\" should also specify how \"never\" is defined over the finite set of rolling windows.","section":"§VIII-C"},{"comment":"The text says \"Live inference, as explained in Section VII, requires overlapping windows,\" but Section VII covers EOG signal evaluation and the ablation study; the rolling-window live inference procedure is described in Section VIII-C. Please reference the correct section.","section":"§IX-B"},{"comment":"The caption says \"with N standing for Negative movements,\" but the figure labels such as \"N-R\" and \"N-UL\" are not explicitly defined in the text. Please explain the notation (presumably movements back to the center) in the caption or in Section VI-A.","section":"Figure 7 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and represents solid systems work, but the split contradiction is a serious correctness issue that needs to be resolved before publication. The authors should be asked to clearly separate the ablation-study protocol (Section VII-A) from the final-model protocol (Section VIII) and to provide subject-disjoint results if the user-independence claim is to stand. The unit error in the abstract is easily fixed but should not survive. The self-citations to the group's own VitalCore/GAP9 work are frequent but appropriate given the hardware lineage; no citation-ethics concern beyond the usual expectation of full disclosure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"ElectraSight is a real piece of engineering: a fully onboard, low-power EOG eye-movement classifier in glasses using hybrid contact/contactless electrodes, with a 20-subject dataset, a 4-bit quantized CNN on GAP9, and external ground truth from Pupil Labs Neon. That part is solid and worth reading. The ablation study showing that neither contact-only nor contactless-only channels work well (48% and 50% relative accuracy for all classes) is a good piece of evidence for the hybrid design. The impedance characterization is careful, and the GAP9 quantization table is honest: accuracy holds at 4-bit, collapses at 2-bit.\n\nThe soft spot is the evaluation protocol. Section VII-A says acquisitions are randomly assigned 80/20 to train/test. Section VIII claims the model is evaluated 'on subjects not encountered during training.' Those are incompatible. If the final 81%/92% numbers come from the acquisition-level split, then the same subject's data can appear in both training and test, and the abstract's 'no calibration or user-specific fine-tuning' claim is not supported. The paper doesn't give the split ratio, subject count, or subject IDs for the final model, so the reader can't tell which protocol produced Figures 11a/11b, Table V, or the latency results. This is load-bearing, not cosmetic.\n\nMinor issues: the abstract says 46 mJ per inference but the body and Table VI say 46 µJ; there are class-count inconsistencies (10 vs 9 vs 6); no confidence intervals or repeated splits; and no code, dataset, or model weights released, so reproducibility is limited. The self-citations (VitalCore, GAP9-related work) are standard for a hardware-group paper and don't generate the accuracy numbers.\n\nBottom line: the hardware and dataset are genuinely useful, and the confusion matrices look plausible, but the headline calibration-free/user-independent claim hinges on the split ambiguity. This deserves peer review—a good referee will catch it—but the authors need to clarify the split, provide a subject-disjoint evaluation, and release the artifacts. I'd want to see the revised version before relying on the 81%/92% numbers.","headline":"A credible low-power EOG glasses system with a real hardware contribution, but the headline 'no calibration/user-specific fine-tuning' claim rests on a contradictory train/test split description that needs to be resolved before the accuracies are taken at face value.","tokens_in":23460,"tokens_out":2204,"would_cite":true,"duration_ms":16586,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ElectraSight claims that smart glasses with hybrid contact and contactless EOG electrodes can classify eye movements fully onboard, with 81% accuracy on ten classes and 92% on six, without calibration or per-user fine-tuning.","keywords":["smart glasses","electrooculography (EOG)","QVar charge variation sensing","tinyML","eye movement classification","onboard processing","low-power wearable","hybrid contact and contactless electrodes"],"falsifier":"A subject-holdout re-evaluation using independently annotated labels—for example, frame-by-frame video labeling of the same eye movements by a human annotator instead of gaze-derivative thresholds—would settle the claim. If the 81% ten-class and 92% six-class accuracies do not survive re-labeling, or if accuracy drops sharply when test subjects are completely excluded from training rather than sampled by acquisition split, the generalization claim would be refuted.","tokens_in":22337,"feed_emoji":"👓","tokens_out":7158,"duration_ms":48256,"temperature":0.7,"pith_summary":"ElectraSight aims to show that eye tracking can be done entirely on a pair of smart glasses, without cameras, without wet electrodes, and without per-user calibration. The paper claims that a hybrid electrooculography (hEOG) setup—two contact channels on the nose pads and temples plus three contactless channels around the eye—picks up the eye's corneo-retinal potential well enough for a tiny 79 kB neural network to classify ten eye movements at 81% accuracy, and 92% for six basic movements. If true, this would make continuous, private, low-power eye tracking practical for everyday glasses, at 8.85 mW average power and a three-day battery life on a 175 mAh cell. The central payoff is a plug-and-play wearable that turns eye movement into a control or health signal without the energy cost and privacy exposure of camera-based trackers.","feed_headline":"No camera, no calibration: glasses read eye movements at 81% accuracy","feed_subtitle":"Hybrid contact and contactless EOG runs fully onboard for three days on a 175 mAh battery, with no per-user training.","key_machinery":"The load-bearing mechanism is the hybrid EOG (hEOG) front end: two galvanic contact channels at the nose pads and temples plus three contactless channels around the eye, all read by QVar sensors whose input impedance (up to 2.4 GΩ) keeps the attenuated contactless signal above the sensor noise floor. The signal chain is per-window standardization followed by a Savitzky–Golay filter, then a 1D-CNN with four convolutional layers (64 filters, kernel size 7) and transposed-convolution layers, trained on 416 ms windows and quantized from float32 to 4-bit (79 kB) without accuracy loss down to 4-bit. The quantized model executes on the GAP9 cluster at 370 MHz in 301 microseconds, enabling rolling-window inference with a 42 ms slide at 90% overlap.","core_discovery":"The paper's central claim is that a fully onboard, calibration-free eye movement classifier can be built from hybrid contact and contactless electrooculography. Using charge-variation (QVar) sensors that measure quasi-electrostatic potential changes through a gigaohm-input differential front end, five channels capture the corneo-retinal potential through the glasses frame. A 1D convolutional network with 151,447 parameters, quantized to 4-bit precision and fitting in 79 kB, runs on a RISC-V GAP9 coprocessor at 301 microseconds per inference, achieving 81% accuracy on ten movement classes and 92% on six basic ones. The system also reports that 90% of movements are detected within 60 ms of onset and that the whole glasses draw 8.85 mW, allowing over three days of continuous use on a 175 mAh battery. Accuracy is measured against labels derived from a camera-based ground-truth eye tracker, using thresholds on the gaze derivative to mark movement timing.","pith_inferences":["The accuracy claims depend on the eye-tracker-derived labels; if the gaze-derivative thresholds mistime fast movements (the paper reports logger labels are 4.6 times less accurate for rapid movements), the confusion matrices and latencies would shift, and an independent manual annotation study on the same recordings would quantify this.","The reported 92% and 81% accuracies might transfer to everyday use only if the controlled screen-following protocol generalizes to natural gaze behavior; a field test with unstructured viewing would be a direct extension.","Because head motion and facial muscle artifacts couple into charge-variation channels, fusing the QVar signal with the onboard IMU could reject motion artifacts and possibly push accuracy beyond the current all-channels model.","The hybrid-channel ablation suggests contactless channels are indispensable for corner movements (relative accuracy drops from 91% to 50% without them), so exploring additional contactless electrode placements might improve the ten-class result."],"forward_implications":["Eye movement can become a hands-free control signal for AR interfaces, menu navigation, and assistive communication, generated locally on the glasses with no video leaving the device.","The 8.85 mW total consumption projects to more than three days of continuous operation on a 175 mAh battery, making all-day wearable eye tracking feasible.","Because the model is subject-agnostic and calibration-free, a user can put on the glasses and immediately get blink and gaze-direction classification, unlike systems that require per-user training.","With 90% of movements detected within 60 ms, the classification latency is short enough for real-time interaction and notification-style responses.","The 416 ms window with 90% overlap supports detection of several movements per second, matching the typical rate of human saccades."],"supporting_citations":[{"why":"Supplies the wet-electrode EOG baseline that reaches high accuracy but consumes 1446 mW, used as the main power-and-invasiveness comparison.","marker":"[24]"},{"why":"Supplies a fully onboard dry-electrode EOG/EEG glasses baseline that requires per-user training, compared against the calibration-free claim.","marker":"[18]"},{"why":"Supplies an electrostatic-interface eye tracking baseline with 97% accuracy on nine categories, compared for contactless sensing and industrialization.","marker":"[27]"},{"why":"Supplies an acoustic gaze tracking baseline with 95.4 mW power and susceptibility to external noise, compared for power and reliability.","marker":"[26]"},{"why":"Supplies an event-camera low-power eye tracking baseline that is not fully onboard and uses non-commercial components, compared for integration.","marker":"[20]"},{"why":"Supplies the standard corneo-retinal potential amplitude and bandwidth used to justify the QVar sensor's sensitivity requirements.","marker":"[39]"},{"why":"Supplies prior QVar biopotential measurements including limited EOG, which this paper extends to full eye movement classification.","marker":"[37]"},{"why":"Supplies the open-source VitalCore hardware platform on which the custom glasses electronics are built.","marker":"[40]"}],"fun_headline_variants":["No cameras, no calibration: glasses read eyes at 81%","Smart glasses eye tracking: 81% accuracy, zero setup","Non-invasive eye tracking: 3-day battery, no calibration","Onboard AI predicts eye movements with 81% accuracy","Eye-tracking glasses: 10 classes, 81% accuracy, no per-user tuning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracies and latencies stand or fall with the ground-truth labeling: labels come from manually chosen thresholds on the derivative of a camera-based reference eye tracker, and the paper itself notes that the logger-based labels are 4.6 times less accurate for rapid movements, so any bias in those thresholds would shift every accuracy and timing figure.","fun_headline_variants_meta":{"raw":{"variants":["No cameras, no calibration: glasses read eyes at 81%","Smart glasses eye tracking: 81% accuracy, zero setup","Non-invasive eye tracking: 3-day battery, no calibration","Onboard AI predicts eye movements with 81% accuracy","Eye-tracking glasses: 10 classes, 81% accuracy, no per-user tuning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000323,"raw_usage":{"total_tokens":1867,"prompt_tokens":1053,"completion_tokens":814,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":669,"completion_tokens_details":{"reasoning_tokens":722}},"tokens_in":669,"tokens_out":814,"duration_ms":6098,"temperature":1.0,"reasoning_tokens":722,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:51:33.891860+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A subject-holdout re-evaluation using independently annotated labels—for example, frame-by-frame video labeling of the same eye movements by a human annotator instead of gaze-derivative thresholds—would settle the claim. If the 81% ten-class and 92% six-class accuracies do not survive re-labeling, or if accuracy drops sharply when test subjects are completely excluded from training rather than sampled by acquisition split, the generalization claim would be refuted.","supporting_citations":[{"cited_title":"EOG Signal Processor: An SVM-based Multiclass Classifier to Detect Eye Movements,","cited_arxiv_id":null,"evidence_quote":"Supplies the wet-electrode EOG baseline that reaches high accuracy but consumes 1446 mW, used as the main power-and-invasiveness comparison."},{"cited_title":"GAPses: Versatile smart glasses for comfortable and fully-dry acquisition and parallel ultra-low-power processing of EEG and EOG,","cited_arxiv_id":null,"evidence_quote":"Supplies a fully onboard dry-electrode EOG/EEG glasses baseline that requires per-user training, compared against the calibration-free claim."},{"cited_title":"Eye tracking and eye expression decoding based on transparent, flexible and ultra-persistent electrostatic interface,","cited_arxiv_id":null,"evidence_quote":"Supplies an electrostatic-interface eye tracking baseline with 97% accuracy on nine categories, compared for contactless sensing and industrialization."},{"cited_title":"Retina : Low-Power Eye Tracking with Event Camera and Spiking Hardware","cited_arxiv_id":"2312.00425","evidence_quote":"Supplies an event-camera low-power eye tracking baseline that is not fully onboard and uses non-commercial components, compared for integration."},{"cited_title":"ISCEV Standard for Clinical Electro-oculography (EOG) 2006,","cited_arxiv_id":null,"evidence_quote":"Supplies the standard corneo-retinal potential amplitude and bandwidth used to justify the QVar sensor's sensitivity requirements."},{"cited_title":"Long-Term Polygraphic Monitoring through MEMS and Charge Transfer for Low-Power Wearable Applications,","cited_arxiv_id":null,"evidence_quote":"Supplies prior QVar biopotential measurements including limited EOG, which this paper extends to full eye movement classification."},{"cited_title":"TinyssimoRadar: In-Ear Hand Gesture Recognition with Ultra-Low Power mmWave Radars,","cited_arxiv_id":null,"evidence_quote":"Supplies the open-source VitalCore hardware platform on which the custom glasses electronics are built."}],"review_version":1}