Pith. sign in

REVIEW 5 major objections 6 minor 12 references

EEG emotion models can be made readable by turning neural signals into identity-free facial emojis, and the same training also lifts EEG-only accuracy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

EEG can be mapped to identity-anonymized facial emojis that act as a semantic regularizer, yielding SOTA EEG-only emotion accuracy and human-interpretable animations.

T0 review reviewed 2026-07-12 challenge →

load-bearing objection Solid multi-task EEG–emoji system that delivers real accuracy gains and a usable privacy-preserving visualization; the “brain’s emotional evolution” framing is stronger than the motor-proxy evidence fully supports. the 5 major comments →

arxiv 2607.02912 v1 pith:XZS5N5WM submitted 2026-07-03 cs.CV cs.HCcs.MM

See the Emotion: A Facial Emoji Proxy Modeling for EEG Emotion Recognition

Jingjing Hu , Guo Dan , Haofan Cheng , Ying Zeng , Zhan Si , Jinxing Zhou , Meng Wang This is my paper
classification cs.CV cs.HCcs.MM
keywords EEG emotion recognitionfacial emoji proxycross-modal generationexplainable BCIprivacy-preserving visualizationFMENetsemantic regularizer
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Standard EEG emotion classifiers map scalp voltages to labels but give no intuitive picture of what the model has seen. This paper reframes that opacity as a generation problem: learn to reconstruct short sequences of binary, identity-stripped facial emojis from EEG alone. The joint training forces the network to keep only those neural patterns that also predict facial geometry, so the same encoder both classifies better and can animate an emoji face at inference time. On two synchronized EEG–face benchmarks the resulting system sets the highest EEG-only recognition scores while producing animations that non-experts can read at roughly three-quarters accuracy. The practical claim is that a privacy-preserving facial proxy is enough to turn an opaque neural code into a visible emotional trajectory.

Core claim

Facial Emoji Proxy Modeling—pairing an expression-aware EEG backbone (FMENet) with a frozen emoji decoder (FELB)—turns EEG-to-emoji reconstruction into a semantic regularizer that simultaneously raises EEG-only emotion accuracy to state-of-the-art levels and yields identity-anonymized facial animations that let a user watch the emotion unfold directly from neural signals.

What carries the argument

Facial Emoji Learning Branch (FELB): a multi-task head that projects EEG features into a pre-trained emoji VAE latent space, reconstructs binary facial landmarks, and feeds the latent code back into the emotion classifier so that classification is conditioned on the behavioral manifold.

Load-bearing premise

The paper assumes that the coupling between EEG patterns and facial muscle geometry is strong and stable enough that emoji reconstruction both regularizes the emotion classifier and faithfully visualizes the brain’s emotional evolution rather than merely a correlated motor proxy.

What would settle it

On a held-out subject set, mask the motor-cortex and prefrontal channels that the model’s own attention highlights; if emoji reconstruction quality and human-readable emotion accuracy collapse while classification accuracy stays high, the claimed neural–facial semantic bridge is broken.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes Facial Emoji Proxy Modeling for EEG emotion recognition: an EEG encoder (FMENet: Expression-Relevant Spatial Merger + Multi-Scale Temporal Capturer) is trained multi-task with a Facial Emoji Learning Branch (FELB) that maps EEG to a frozen VAE latent space of identity-anonymized binary facial-landmark emojis. At inference only EEG is used, yielding both an emotion label and a dynamic emoji sequence. On EAV the full model reports 77.13% accuracy / 77.02% F1 (EEG-only inference), competitive with multimodal baselines; on MMER it reports 69.73% / 64.28%. Ablations, backbone compatibility, reconstruction metrics (SSIM ≈0.91 on EAV), a small user study (n=10, 76% 5-way), region-masking, and zero-shot transfer to SEED (~60% 3-way via an emoji ResNet) are provided. The central claim is that emoji reconstruction is both a semantic regularizer that improves EEG-only accuracy and a privacy-preserving visualization of the brain’s emotional evolution.

Significance. If the accuracy and reconstruction results hold under broader scrutiny, the work offers a useful paradigm shift for EEG interpretability: from saliency maps to identity-stripped behavioral proxies that non-experts can read. Strengths include a clear multi-stage training recipe (Algorithm 1), extensive backbone comparisons (Tables 1–2, 6–7), architectural ablations (Table 3), hyperparameter grids (Tables 9–10), efficiency numbers (Tables 11–12), cross-subject splits (Table 8), a user study, zero-shot SEED transfer, and an unusually thorough ethics appendix. Code is promised. The privacy-by-design emoji abstraction is a genuine practical contribution for affective BCI. The significance is therefore real for affective computing and explainable BCI, provided the authors carefully bound what the emojis actually visualize (facial-motor geometry correlated with labels, not a direct readout of subjective affect).

major comments (5)
  1. [§1, abstract, §4.7, Fig. 4, Tab. 5, §5] The load-bearing interpretability claim—that generated emojis give a “transparent window into the brain’s emotional evolution” (abstract, §1, §5)—is stronger than the evidence. Fig. 4 and Table 5 show ESM attention and SSIM drops concentrated on motor cortex for mouth-corner actions and prefrontal for brow-related negatives; this is consistent with decoding expression-related motor/premotor patterns that co-vary with emotion labels, not necessarily internal affective state. Frame-level Emo. Acc. (Table 4) and the user study (Fig. 5) evaluate face-derived geometry with face-trained classifiers/humans. Please reframe claims throughout as privacy-preserving EEG→facial-geometry visualization that is emotion-consistent, and discuss the motor-proxy alternative explicitly (e.g., in §4.7 and §5).
  2. [§4.2–4.3, Tab. 4, App. E, Eq. 5] Semantic-fidelity evaluation is partially circular. Training uses synchronized face video to supervise emoji reconstruction (Eq. 5, FELB); “Emo. Acc.” then scores reconstructed emojis with a ResNet-18 trained on the same emoji corpus against frame-level pseudo-labels from an external face emotion model (§4.2–4.3, App. E, Tabs. 14–15). Emotion classification itself uses video-level labels and is not circular, but the claim of “semantically faithful” visualization is. Report at least one face-independent check (e.g., correlation of emoji trajectories with EEG spectral markers known a priori, or human ratings against video-level labels only) and state clearly that Emo. Acc. measures geometric–label consistency, not independent affective ground truth.
  3. [§4.5, Tab. 1, Tab. 8, App. C.3] SOTA claims are driven by subject-dependent splits; cross-subject performance is much weaker and under-emphasized in the main narrative. Table 1 / FMENet+FELB: 77.13% Acc (subject-dependent EAV); Table 8 leave-subject-out: only 39.05% Acc / 38.17% F1 on EAV (and ~69% on the small MMER hold-out). Please put cross-subject numbers in the main results section, qualify “state-of-the-art among EEG-only models” as subject-dependent unless cross-subject SOTA is demonstrated, and discuss inter-subject variability (already visible in Tabs. 6–7).
  4. [§4.1, MMER protocol] MMER subject selection introduces selection bias that is not stress-tested. §4.1 retains only 14 of the available subjects with ≥95% landmark detection success. This favors easy-to-track faces and may inflate both reconstruction quality and multi-task gains. Provide results (or a sensitivity analysis) on a less filtered subject set, or at minimum report how many subjects/videos were excluded and whether accuracy/SSIM degrade when lower-quality landmarks are included.
  5. [Tab. 2, §4.6, Figs. 11–12, App. B/C.2] Table 2 shows that attaching FELB degrades accuracy for nearly every non-FMENet backbone (often by 1–3 points) while only FMENet gains slightly (76.96→77.13 on EAV). Combined with the video-level vs. frame-level label mismatch documented in Figs. 11–12 and App. B, this suggests the multi-task objective is fragile and architecture-specific rather than a general “semantic regularizer.” Please analyze when and why FELB helps vs. hurts (e.g., conflict between frame-level emoji dynamics and video-level labels), and temper language that presents FELB as a drop-in regularizer for arbitrary EEG encoders.
minor comments (6)
  1. [Tab. 3, §3.2.2] Notation inconsistency: Multi-Scale Temporal Capturer is MTC in §3.2.2 / Fig. 3 but “MEC module” in Table 3; unify.
  2. [Figs. 2, 8, 13–14; App. D] Typos / naming: “EA V” spacing throughout; “Clam” in Fig. 2; “Orignal Face” in Figs. 13–14; “incomprehensiblet” in Fig. 8 caption; “EV A” in App. D.
  3. [§4.7, Fig. 5] User study (n=10, Fig. 5) is useful but under-specified: report clip selection, inter-rater agreement, and whether participants saw ground-truth faces. Move key protocol details into the main text.
  4. [§4.8, App. D] SEED zero-shot uses a 28-channel common subset and a face-trained ResNet for scoring (App. D, Tab. 13). State channel mapping and evaluation protocol more clearly in the main §4.8 so readers do not over-read the ~60% figure.
  5. [§4.4, Tab. 10] λ grid (Tab. 10) and architectural ablations (Tab. 9) are thorough; a short sentence in §4.4 on why EAV prefers λ=5 and MMER λ=0.2 (active vs. passive elicitation) would help non-specialists.
  6. [§2.2] Related work on EEG–face coupling (Soleymani et al., Du et al.) is cited; a brief comparison to prior EEG-to-face or AU regression work would better position the emoji abstraction choice.

Circularity Check

0 steps flagged

No definitional circularity; mild partial dependence of semantic-fidelity metrics on face-derived emojis used only at training/eval, not for the EEG-only accuracy claim.

full rationale

This is an empirical multi-task ML paper, not a first-principles derivation. Emotion labels Y supervise L_cls independently of the emoji reconstruction branch (Eq. 5); FMENet+FELB accuracy (Tabs. 1–3, 6–8) is measured against those video-level labels under EEG-only inference and is therefore not forced by the face targets. Emoji reconstruction quality (MSE/PSNR/SSIM in Tab. 4) and Emo. Acc. are evaluated against ground-truth emojis / a ResNet trained on the same emoji corpus, which is the ordinary supervised reconstruction objective rather than a self-definitional loop or fitted-input-as-prediction. Frame-level pseudo-labels are explicitly withheld from training and model selection (§4.2, App. B). No uniqueness theorems, self-citation chains, or ansatz smuggling appear; related-work citations are external. The only mild circularity-adjacent element is rhetorical over-claim that generated emojis visualize “the brain’s emotional evolution” when the regularizer is facial-motor geometry; this does not reduce any reported number by construction. Score 1 reflects that single non-load-bearing evaluation dependence; central SOTA accuracy claim remains independent.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 3 invented entities

The central claim rests on the empirical neural–facial association, a set of architectural free parameters chosen by grid search, and three invented modules (FMENet, FELB, Facial Emoji Proxy) whose value is demonstrated only inside this paper’s experiments.

free parameters (5)
  • multi-task weight λ = 5 (EAV), 0.2 (MMER)
    Chosen by grid search over [0,10]; set to 5 (EAV) and 0.2 (MMER). Directly balances classification vs. reconstruction and therefore affects both accuracy and emoji fidelity claims.
  • spatial merger heads O = 8 / 6
    Number of expression-relevant spatial groups; set to 8 (EAV) / 6 (MMER) after ablation (Tab. 9).
  • MTC depth K and dilation period D* = K=5/3, D*=5
    Temporal stack depth and dilation schedule; set to K=5/3, D*=5 after ablation.
  • emoji VAE latent dim L = 64 / 32
    Latent size of the frozen emoji decoder; 64 (EAV) / 32 (MMER).
  • emoji temporal interval Δt = 1.0 s / 0.5 s
    Seconds of EEG summarized by each emoji token; chosen as 1.0 s (EAV) / 0.5 s (MMER).
axioms (3)
  • domain assumption Affective states are systematically expressed in coordinated facial muscle activity that is partially decodable from EEG (neural–facial association).
    Stated in §1 and used to justify the entire cross-modal generation objective; supported by citations but not re-proven here.
  • ad hoc to paper Binary landmark rasterizations (56×56) preserve expression geometry while removing identity.
    Design choice in §3.3.1; privacy and topology claims rest on it.
  • ad hoc to paper A frozen VAE decoder trained on all emojis supplies a stable behavioral manifold that EEG can be aligned to.
    Stage-2 pre-training assumption (§3.3.2); reconstruction quality metrics depend on it.
invented entities (3)
  • Facial Emoji Proxy no independent evidence
    purpose: Identity-anonymized visual target that turns EEG explainability into a generation task.
    Core modeling object; no independent existence outside this framework.
  • FMENet (Expression-Relevant Spatial Merger + Multi-Scale Temporal Capturer) no independent evidence
    purpose: EEG backbone specialized for expression-aligned features.
    New architecture whose superiority is shown only on the paper’s benchmarks.
  • Facial Emoji Learning Branch (FELB) no independent evidence
    purpose: Multi-task head that maps EEG latents into the emoji VAE space and regularizes classification.
    Invented regularizer; value demonstrated only inside the multi-task experiments.

reviewed 2026-07-12 · how reviews work

0 comments
Cite this review

Pith. "Pith review of See the Emotion: A Facial Emoji Proxy Modeling for EEG Emotion Recognition." pith.science (2026). https://pith.science/paper/XZS5N5WM

@misc{pith2026260702912,
  author       = {Pith},
  title        = {Pith review of: See the Emotion: A Facial Emoji Proxy Modeling for EEG Emotion Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XZS5N5WM}},
  note         = {Machine review of arXiv:2607.02912}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Despite the high accuracy of EEG-based emotion recognition, existing models remain opaque "black boxes", lacking semantic grounding between abstract neural features and human-interpretable states. In this paper, we reframe EEG explainability as a cross-modal generation task, shifting the paradigm from feature attribution to behavioral visualization. We introduce Facial Emoji Proxy Modeling, a novel framework that translates high-dimensional EEG signals into identity-anonymized facial emojis. Guided by the neuroscientific inspiration of neural-facial association, this approach grounds neural representations in the manifold of observable facial dynamics. Technically, our framework integrates FMENet, a specialized backbone modeling expression-relevant spatial synergies, and the Facial Emoji Learning Branch (FELB), which treats emoji reconstruction as a structured semantic regularizer. Extensive experiments on EAV and MMER benchmarks demonstrate that our method achieves state-of-the-art accuracy among EEG-only models. Crucially, it generates semantically faithful facial animations that provide a transparent, privacy-preserving window into the brain's emotional evolution, effectively allowing users to "see the emotion" directly from neural signals. Code is available at https://github.com/xian-sh/SeeEmotion

Figures

Figures reproduced from arXiv: 2607.02912 by Guo Dan, Haofan Cheng, Jingjing Hu, Jinxing Zhou, Meng Wang, Ying Zeng, Zhan Si.

Figure 1
Figure 1. Figure 1: Facial Emoji Proxy Modeling. (a) Task goal: translate incomprehensible EEG signals into facial emojis that provide more clearly emotion-related cues. (b) Proposed pipeline: an EEG-based facial emoji proxy model that jointly predicts emotion labels and generates identity-anonymized emojis. Data scenarios see [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the Facial Emoji Proxy Modeling framework. (a) Facial Emoji Learning Branch: 1) facial videos are abstracted into binary, identity-anonymized Emojis and 2) learn a VAE-based emoji manifold (Eϕ–Dϕ), 3) training coupled with emotion recognition task. (b) Cross-modal multi-task learning: the shared FMENet backbone (Eθ) encodes EEG into Z, driving two branches: ① Generative Branch maps to the emoji… view at source ↗
Figure 3
Figure 3. Figure 3: FMENet backbone, It encodes neural dynamics via two specialized modules: (1) Expression-Relevant Spatial Merger (ESM): discrete electrode coordinates are embedded by Fourier po￾sition embeddings to enable geometry-aware attention that aggre￾gates channels into functionally coherent groups. (2) Multi-Scale Temporal Capturer (MTC): a stack of dilated 1D convolutions with exponentially expanding receptive fie… view at source ↗
Figure 4
Figure 4. Figure 4: EEG spatial-attention (Eq. 2) drivers for emoji actions [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 7
Figure 7. Figure 7: Model on EAV is used for SEED (No Face data). ally plausible, with most predictions falling into the overall Negative category. More qualitative cases are in App. E. 4.8. Zero-shot Generalization to EEG-only Dataset Since our framework requires no facial input at inference time, it can be directly applied to EEG-only datasets. We take the model trained on EAV (without any fine-tuning) and deploy it on SEED… view at source ↗
Figure 8
Figure 8. Figure 8: Data collection scenarios and research goal. (a) EAV dataset (Lee et al., 2024): active online communication with a cue-based conversational system, where 30-channel EEG and frontal facial videos are recorded while subjects engage in emotional dialogues. (b) MMER dataset (Yang et al., 2024): passive viewing of emotion-eliciting film clips, with synchronized EEG and facial videos and post-trial self-reports… view at source ↗
Figure 9
Figure 9. Figure 9: File organization of the EAV dataset. The data in the green area is newly extracted by us. All data will be open-sourced. MMER/ Images/ 1_face_imgs/ video_00_0.0s.jpg # Size: 3x56x56 video_00_0.5s.jpg …… video_31_19.5s.jpg 5_face_imgs/ Emojis/ # The file structure is the same as Images/. # Size of each emoji image: 1x56x56 EEG/ 1_eeg_20s.pkl # Size: 32x20x18x300 # Size: #Video, #Time, #Channels, #Sampling … view at source ↗
Figure 12
Figure 12. Figure 12: Frame-level emotion annotations created to describe the emotional dynamics of the reconstructed face emojis. though such modeling remains necessary for constructing authentic emotional dynamics. Data Organization. The processed datasets maintain consistent structures: EAV is organized by subject with separate pickle files for EEG (subject XX eeg.pkl), facial images (subject XX vis.pkl), and emojis (emoji … view at source ↗
Figure 13
Figure 13. Figure 13: Extended visualization of representative examples from the EAV dataset. (1) EEG. Label: Positive; Subject:11, Video:7 (4) Facial Emojis (Reconstructed), Predicted Emotions Happy(0.80) Happy(0.86) Happy(0.93) Happy(0.94) (3) Facial Emojis (Ground Truth) MSE:0.0167 PSNR:17.8 SSIM:0.860 MSE:0.0175 PSNR:17.6 SSIM:0.858 MSE:0.0145 PSNR:18.4 SSIM:0.878 MSE:0.0166 PSNR:17.8 SSIM:0.867 Happy Happy Happy Happy (3)… view at source ↗
Figure 14
Figure 14. Figure 14: Extended visualization of representative examples from the MMER dataset. high similarity in arousal between fear and surprise), we merged semantically similar categories into five meta-classes: neutral, happy, sad, angry+disgust, and fear+surprise. The model was validated using a large-scale test set comprising 126,000 frames from the EAV dataset (our largest benchmark), thereby ensuring robust performanc… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

12 extracted references · 1 canonical work pages

  1. [1]

    Bouazizi, S

    doi: 10.1109/TII.2022.3197419. Bouazizi, S. and Ltifi, H. Explainable grouped deep echo state network for eeg-based emotion recognition.Soft Computing, pp. 1–18,

    doi:10.1109/tii.2022.3197419 2022

  2. [2]

    Feng, L., Cheng, C., Zhao, M., Deng, H., and Zhang, Y

    URL https://arxiv.org/abs/ 2404.08472. Feng, L., Cheng, C., Zhao, M., Deng, H., and Zhang, Y . Eeg-based emotion recognition using spatial-temporal graph convolutional lstm with attention mechanism.IEEE Journal of Biomedical and Health Informatics, 26(11): 5406–5417,

    Pith/arXiv arXiv

  3. [3]

    T., and Wang, N

    Kang, Z., Li, Y ., Gong, S., Zeng, W., Yan, H., Bian, L., Zhang, Z., Siok, W. T., and Wang, N. Hypergraph multi- modal learning for eeg-based emotion recognition in con- versation.arXiv preprint arXiv:2502.21154,

    Pith/arXiv arXiv

  4. [4]

    Lai-Tan, N., Gu, X., Philiastides, M

    URL https://arxiv.org/ abs/1312.6114. Lai-Tan, N., Gu, X., Philiastides, M. G., and Deligianni, F. Cross-subject and cross-montage eeg transfer learning with individual tangent space alignment. InWomen in Machine Learning Workshop@ NeurIPS 2025,

    Pith/arXiv arXiv 2025

  5. [5]

    doi: https://doi.org/10.1016/j.neuroimage.2023.120209

    ISSN 1053-8119. doi: https://doi.org/10.1016/j.neuroimage.2023.120209. URL https://www.sciencedirect.com/ science/article/pii/S1053811923003609. Mohsan, M. M., Akram, M. U., Rasool, G., Alghamdi, N. S., Baqai, M. A. A., and Abbas, M. Vision transformer and language model based radiology report generation.IEEE Access, 11:1814–1824,

    work page doi:10.1016/j.neuroimage.2023.120209 2023

  6. [6]

    doi: 10.1002/hbm.23730

    ISSN 1097-0193. doi: 10.1002/hbm.23730. URL http: //dx.doi.org/10.1002/hbm.23730. 10 See the Emotion: A Facial Emoji Proxy Modeling for EEG Emotion Recognition Scotti, P., Banerjee, A., Goode, J., Shabalin, S., Nguyen, A., Dempster, A., Verlinde, N., Yundler, E., Weisberg, D., Norman, K., et al. Reconstructing the mind’s eye: fmri- to-image with contrasti...

    doi:10.1002/hbm.23730

  7. [7]

    sub01_train_sample0_frame0

    B.1.2. DATASTATISTICS ANDANNOTATION EA V Datasetcontains 42 subjects with balanced distribution across five emotion categories: Neutral, Anger, Happiness, Sadness, and Calmness. The original video-level emotion annotation distribution is shown in Fig. 11(a), demonstrating well-balanced class representation. After processing, we obtained approximately 400K...

    2024

  8. [8]

    15 See the Emotion: A Facial Emoji Proxy Modeling for EEG Emotion Recognition Table 6.Supplement to Tab.2 in the main paper

    is a time series lightweight adaptive network that divides input sequences into patches and employs transformer-like architecture for temporal representation learning. 15 See the Emotion: A Facial Emoji Proxy Modeling for EEG Emotion Recognition Table 6.Supplement to Tab.2 in the main paper. Performance (%) of different backbones onEA Vdataset. Subject Sy...

    arXiv 1953

  9. [9]

    Consistent with the default settings in § 4.4, the best-performing weights are λ=5 for EA V and λ=0.2 for MMER, respectively

    The search is performed over a unified gridλ∈[0,10] for both datasets. Consistent with the default settings in § 4.4, the best-performing weights are λ=5 for EA V and λ=0.2 for MMER, respectively. Importantly, the performance curves exhibit broad plateaus (within3.5% relative variation) rather than sharp peaks, suggesting that FMENet is reasonably robust ...

    2023

  10. [10]

    The introduced FELB module incurs negligible overhead (0.7K parameters and less than 0.001 GLOPS) and the 15.83 MB frozen V AE decoder only requires initialisation once offline

    The computational cost of our model is within the acceptable range for practical deployment. The introduced FELB module incurs negligible overhead (0.7K parameters and less than 0.001 GLOPS) and the 15.83 MB frozen V AE decoder only requires initialisation once offline. On a single A5000 GPU, the end-to-end inference latency for producing both the emotion...

    1983

  11. [11]

    in our generalisation experiments, to ensure compatibility between the EV A/MMER and SEED datasets. Taking the generalization validation from the EA V dataset to the SEED dataset as an example, we identified 28 electrodes that are physically shared between the two datasets based on the International 10–20 System (Lee et al., 2024; Zheng & Lu, 2015). The m...

    2024

  12. [12]

    Table 14.Per-class Precision, Recall, and F1 of the ResNet-18 classifier on the EA V test set (126k frames)

    These supplementary visualizations encompass a wider range of emotional states and subject variations, consistently showing high-quality emoji generation with semantically meaningful emotional labels. Table 14.Per-class Precision, Recall, and F1 of the ResNet-18 classifier on the EA V test set (126k frames). Class Support Precision Recall F1 Neutral 25,65...

    1976

This paper was first reviewed by grok-4.5 on July 12, 2026.