{"id":"9e27bfef-104f-421c-9518-d3945ab7afd2","arxiv_id":"2505.10575","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"SSOCL assigns pseudo-labels to unlabeled EEG streams via clustering of predicted embeddings and trains the model with an entropy-filtered replay buffer, claiming state-of-the-art cross-subject emotion recognition.","lead":"This paper presents SSOCL, a self-supervised continual learning framework that adapts an EEG emotion recognition model to unlabeled streaming data from new subjects. It combines future-embedding prediction, k-means pseudo-labeling, and entropy-based memory replay, reporting strong results on two EEG benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pseudo-labels are never validated; k-means with K fixed to source classes can create spurious clusters in single-class batches, so the reported gains cannot be attributed to the proposed mechanism.","rationale":"Reader's weakest_assumption is the same as the one I find most load-bearing. I considered two other candidates: Eq. (1)'s contrastive loss is not well-formed as written (the denominator sums over predicted embeddings only and includes the self-term, and no normalization is specified), and the temperature used in the main setup (T=10) conflicts with the ablation that calls T=100 optimal. Both are real internal inconsistencies, but they are secondary: a corrected loss or a different temperature would not rescue the method if the pseudo-labels are arbitrary. The pseudo-label chain is the mechanism that converts unlabeled EEG streams into the supervisory signal for Eq. (8), and every reported success metric depends on it. The paper itself flags the single-class-batch failure mode in Section 4.1.1 and then relies on an unsubstantiated claim that the memory buffer will naturally create boundaries; Section 4.2.1 only filters by prediction entropy, which cannot tell a confident wrong label from a confident correct one. Because this gap is central and empirically testable with labels that are already in the datasets, I retain the reader's REJECT verdict.","tokens_in":14443,"tokens_out":10045,"duration_ms":96075,"concrete_test":"Run one control experiment on DEAP (or both datasets): during the streaming loop, record for every batch the pseudo-labels assigned by Eq. 4 and compare them to the ground-truth labels (which exist in the dataset but are not used for training), reporting per-batch accuracy and adjusted Rand index; then rerun the entire SSOCL pipeline with pseudo-labels replaced by random permutations (seeded) while keeping all other components identical. If pseudo-label accuracy is near chance, or if the random-label control matches SSOCL's GenAcc within a few points, the claimed gains do not come from the cluster-mapping mechanism and the central claim is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the pseudo-labels produced by k-means and the cluster-mapping module (Eqs. 3-4) being aligned with true emotion classes, because Eq. 8 trains the classifier on those labels and the memory buffer is curated using them. The paper provides no measurement of pseudo-label accuracy or cluster purity, and its own Section 4.1.1 admits that when the current batch contains a single class, k-means with K fixed to the source-class count creates spurious partitions. Section 4.1.2 makes this worse: the first batch is stored with arbitrary cluster-number pseudo-labels, and if that batch is single-class those arbitrary labels become the anchor for all later mappings via Eq. 4, with no correction step. The entropy filter in Section 4.2.1 removes high-entropy samples but cannot detect whether a confidently predicted label is the right true class. Consequently, the large reported gains (DEAP GenAcc 71.78 vs 40.51; AMIGOS 56.51 vs 37.24) cannot be attributed to the proposed label-free mechanism unless pseudo-labels are shown to be correct; without such evidence, the central claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SSOCL, a bi-level self-supervised continual learning framework for EEG emotion recognition from streaming, unlabeled data. In an inner loop, a feature extractor is adapted to the current batch via a future-prediction contrastive loss; k-means clustering then partitions the batch into K clusters (K equal to the number of source classes), and a cluster-mapping module assigns pseudo-labels by comparing centroids with those of a memory buffer. In the meta-loop, an entropy-based memory enhancement module curates a fixed-size buffer, and the classifier is trained with cross-entropy on the merged pseudo-labeled current and memory data. The framework is evaluated in two transfer settings: PPB-EMO as source to DEAP, and PPB-EMO+DEAP as source to AMIGOS, reporting GenAcc of 71.78% (DEAP) and 56.51% (AMIGOS), with lower forgetting than baselines. The paper claims that SSOCL outperforms existing approaches in this label-free continual learning scenario.","tokens_in":14691,"tokens_out":4947,"duration_ms":44668,"significance":"If the central claims held, this would be a valuable contribution to affective computing: a truly label-free online continual learning method for EEG that handles cross-subject variability without distorting augmentations. The combination of temporal prediction for self-supervision, cluster-based pseudo-labeling, and entropy-based memory curation is original and practically motivated. However, the contribution is currently undermined by (i) an internal contradiction in the reported results on AMIGOS, (ii) a complete absence of validation of the pseudo-label mechanism on which the method depends, and (iii) an unusual contrastive loss with unaddressed degeneracy risks. The paper also has no code or machine-checked artifacts to substantiate reproducibility. The significance is therefore contingent on substantial further evidence, and is not yet established.","major_comments":[{"comment":"The claim in Section 5.5 that 'SSOCL outperforms all baselines across all performance metrics' is directly contradicted by Table 2: on AMIGOS, EWC reaches AdapAcc 72.32% while SSOCL reaches AdapAcc 62.75%. This is not a minor reporting slip—it invalidates the stated superiority claim. The text must be corrected or the experiments re-examined, and any revised claim must account for this counterexample.","section":"Section 5.5, Table 2"},{"comment":"The pseudo-label assignment mechanism is the load-bearing component of the method, yet it is never validated. Section 4.1.1 admits that running k-means with K equal to the source class count on a batch containing a single class creates spurious partitions, and Section 4.1.2 states that the first batch is stored with arbitrary cluster-number pseudo-labels that become anchors for all later mappings via Eq. (4). There is no measurement of pseudo-label accuracy, cluster purity, or mapping reliability anywhere in the paper. Because Eq. (8) trains the classifier on these pseudo-labels and the memory buffer is curated using them, the reported gains cannot be attributed to the proposed label-free mechanism without evidence that the pseudo-labels correspond to the true emotion classes.","section":"Sections 4.1.1, 4.1.2, 5.5"},{"comment":"Equation (1) defines a contrastive loss whose denominator sums over predicted embeddings \\hat z_j, while the numerator is a similarity between a predicted embedding and the actual future embedding. This is not a standard InfoNCE formulation, and it is open to trivial solutions: if the predictor network outputs a near-constant vector for all inputs, the loss becomes uninformative because all predicted embeddings would be similar to each other. The paper offers no analysis, ablation, or theoretical justification for this design; the authors should either validate this loss on a simple diagnostic or replace it with a standard contrastive loss.","section":"Section 4.1.1, Eq. (1)"},{"comment":"The ablation study in Table 3 reports SSOCL with AdapAcc 83.18%, GenAcc 72.96%, and ForAcc −9.92%, while Table 1 reports, for the same configuration on DEAP, AdapAcc 83.76%, GenAcc 71.78%, and ForAcc −15.05%. If these are the same experiment, the numbers should match; if they are from different runs or settings, that must be stated explicitly. The unexplained discrepancy reduces confidence in the reliability of the reported results.","section":"Table 3 vs Table 1"}],"minor_comments":[{"comment":"Equation (4) uses 'arg min_{i,j}' over a cosine-distance value, which yields a single pair rather than a mapping between all clusters of D_t and the memory; the intended assignment procedure should be described precisely (e.g., a one-to-one matching or per-cluster assignment).","section":"Section 4.1.2, Eq. (4)"},{"comment":"The statement '10 steps of training are considered in meta-loop' is ambiguous: clarify whether a step is one optimizer update or one pass over a mini-batch, and report the number of updates per time step and the total number of time steps.","section":"Section 5.4"},{"comment":"The caption uses 'current embeddings zt1' and 'future embeddings ˆzt2', while the main text indexes samples by n; unify the notation throughout the figure and text.","section":"Figure 1, caption"},{"comment":"The standard deviations for SSOCL are substantially larger than those of the baselines (e.g., ±4.23 and ±6.60 on DEAP AdapAcc), yet no significance testing (e.g., paired tests across the five runs) is reported; this should be addressed to substantiate the performance differences.","section":"Tables 1 and 2"},{"comment":"The temperature ablation is evaluated only through UMAP visualizations; the claim that T = 100 is optimal would be much stronger with a quantitative metric such as memory-buffer cluster purity or final accuracy per temperature.","section":"Section 6.1"},{"comment":"References [5] and [6] appear to be the same paper (SimCLR); this duplication should be removed and the remaining citation fixed.","section":"References"}],"recommendation":"reject","confidential_remarks":"The internal contradiction on AMIGOS AdapAcc (Section 5.5 vs Table 2) indicates the results were not carefully cross-checked, and the complete absence of pseudo-label validation is a fundamental gap: the method's entire learning signal comes from unverified labels. Together these issues make the central claim unsupported. A resubmission could be considered if the authors add pseudo-label accuracy/purity measurements, correct the loss formulation or justify it, and resolve the reported inconsistency."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's core idea is worth taking seriously: instead of augmenting EEG signals for self-supervision, SSOCL predicts future embeddings from current ones, then uses k-means pseudo-labels and an entropy-based replay buffer. That is a sensible move for low-SNR physiological data, and the ablation shows both modules help. The experimental design, with two target datasets and relevant baselines, is appropriate, and the fact that the authors include variance and ablations is to their credit.\n\nThe soft spots are substantial. The most glaring is that the claimed universal superiority is false. On AMIGOS, EWC reaches an adaptation accuracy of 72.32% while SSOCL gets 62.75%, yet Section 5.5 says SSOCL outperforms all baselines across all metrics. That is an internal contradiction, and it undercuts the headline result. Second, the pseudo-labeling loop is never validated. The model's embeddings define the clusters, the clusters define the pseudo-labels, and those labels train the classifier. The paper itself notes that a single-class batch will be spuriously partitioned with fixed K, and the first batch's arbitrary cluster numbers become anchors for later mapping. Without reporting cluster purity or pseudo-label accuracy, the reported gains cannot be attributed to the pseudo-label mechanism. Third, the temperature for entropy scaling is inconsistent: the setup uses T=10, while the ablation says T=100 is optimal. That suggests the main results may not use the best configuration. Finally, Eq. 1's contrastive loss is not well-formed as written: the denominator sums over all predicted embeddings including the positive sample's own prediction, which isn't a negative and can make the loss trivially small.\n\nIn sum, the framework is novel and the problem is real, but the evidence as presented does not support the central claim. The good news is that all these issues are repairable. I would send it to review with a request for focused revisions: correct the performance claim, validate pseudo-labels, reconcile the temperature, and clean up the loss. If the authors do that, the paper could make a solid contribution. As is, I would not cite it or rely on its numbers.","headline":"Interesting new combination for EEG continual learning, but the central superiority claim is contradicted by its own Table 2 and the pseudo-label mechanism is never validated; not ready as written.","tokens_in":15212,"tokens_out":4705,"would_cite":false,"duration_ms":42198,"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":"SSOCL, a bi-level self-supervised continual learning framework with a pseudo-labeled dynamic memory buffer, lets an EEG emotion model adapt to unlabeled streaming subject data and generalize across subjects better than supervised and…","keywords":["EEG emotion recognition","online continual learning","self-supervised learning","memory replay buffer","pseudo-labeling","temporal contrastive learning","cross-subject generalization","electroencephalography"],"falsifier":"Take a labeled EEG stream and run SSOCL with labels hidden; compare the pseudo-labels assigned by k-means and cluster mapping to the true labels for each batch, computing cluster purity or pseudo-label accuracy. If those scores are near chance, or if a deliberately single-class batch is split into K spurious clusters without the buffer correcting them, the central claim that the memory buffer stores reliable pseudo-labels collapses.","tokens_in":14244,"feed_emoji":"🧠","tokens_out":7598,"duration_ms":68052,"temperature":0.7,"pith_summary":"SSOCL is a self-supervised online continual learning framework that lets an EEG emotion recognition model keep learning from raw, unlabeled data as new subjects arrive in a stream, without falling into catastrophic forgetting. The paper's central contention is that this is possible using a bi-level architecture: an inner loop adapts a feature extractor to each incoming batch via temporal prediction and assigns pseudo-labels through k-means clustering with centroid-based cluster mapping, while a meta-loop trains the classifier on a dynamically curated memory buffer that retains only low-entropy, high-confidence pseudo-labeled samples. If the claim holds, emotion-recognition systems could be deployed in realistic settings where labeled data is scarce and emotions transition without clear boundaries, matching or exceeding methods that rely on labels. On two benchmark EEG datasets, the paper reports generalization accuracies of 71.78% on DEAP and 56.51% on AMIGOS, substantially above the best supervised baseline, and lower forgetting.","feed_headline":"Label-free EEG learner lifts emotion accuracy to 71.78%","feed_subtitle":"A self-supervised memory-buffer method adapts to new subjects without labels while forgetting less than supervised rivals.","key_machinery":"The central machinery is the bi-level self-supervised loop. In the inner loop, a predictor network h maps each current embedding zt_n = f(xt_n) to a predicted next embedding, and a contrastive loss uses the true next embedding as the positive sample and all other predictions as negatives, forcing the feature extractor to preserve temporal dependencies without augmentation. K-means with K equal to the source class count partitions the adapted batch embeddings into clusters; a cluster-mapping module computes centroids for memory and current-batch clusters and assigns pseudo-labels by minimizing cosine distance between them. In the meta-loop, the model is trained with cross-entropy on the merged pseudo-labeled stream and memory data, and the memory enhancement module computes temperature-scaled softmax probabilities and their entropy, retaining only low-entropy, high-confidence samples in the fixed-size buffer. The memory buffer is therefore not just a replay store but a continuously refined pseudo-labeled set whose class boundaries are maintained across subject shifts.","core_discovery":"On the paper's own terms, the discovery is that label-free online continual learning for EEG emotion recognition can be made to work by pairing two mechanisms: a self-supervised future-embedding prediction loss that adapts the feature extractor to each streamed batch without data augmentation, and a dynamic memory buffer whose contents and pseudo-labels are refined in a bi-level loop. The feature extractor is first adapted on the current batch by predicting the next embedding and applying a contrastive loss; k-means then partitions the batch into K clusters, and a cluster-mapping module aligns those clusters to memory clusters by cosine distance between centroids, assigning pseudo-labels. The meta-loop trains the model with cross-entropy on the merged stream-and-memory data, while a memory enhancement module uses temperature-scaled softmax entropy to keep only low-entropy samples in the fixed-size buffer. The paper presents this as a complete solution to cross-subject variability, noisy subjective labels, and blurred emotion boundaries in continuous physiological data streams, validated by experiments on DEAP and AMIGOS.","pith_inferences":["A natural extension is to measure pseudo-label accuracy directly against the true labels of the streamed subjects; the paper does not report this, and it would reveal how much of the accuracy gain comes from correct cluster assignments versus entropy-based filtering.","Because the same machinery is agnostic to signal type, it is plausible, though not shown, that temporal-prediction self-supervision plus entropy-buffered replay transfers to ECG, EMG, or wearable motion data with similar low-SNR, label-scarce streams.","Varying K from the fixed source-class count would test whether the cluster-mapping module truly resolves spurious partitions when a batch contains one or few classes; if accuracy is insensitive to K, the mapping and buffer curation are doing the corrective work the paper attributes to them.","The large run-to-run standard deviations, up to ±9.39 percentage points on AMIGOS, suggest the advantage over baselines may depend on subject order; reporting per-sequence significance tests would place the headline gain on firmer ground."],"forward_implications":["If SSOCL is right, EEG emotion recognition models can keep adapting in real time to new users without any annotation effort, removing the need for subjective self-report labels that are noisy and costly.","The temporal-prediction objective is designed for low-signal-to-noise EEG and avoids the distortions that augmentation-based contrastive baselines suffer, which the ablation study supports.","The entropy-based memory curation should keep the buffer class-balanced and separable, as the UMAP visualizations show, making replay effective across evolving subjects.","The reported ForAcc values indicate the method forgets less than supervised and self-supervised baselines, so the model remains useful for previously seen subjects after learning new ones."],"supporting_citations":[{"why":"Supplies the AMIGOS EEG dataset used as one of the two continual-learning target domains.","marker":"[8]"},{"why":"Supplies the DEAP EEG dataset used as the main continual-learning target domain.","marker":"[22]"},{"why":"Supplies the PPB-EMO driving EEG dataset used to pre-train the source model before streaming.","marker":"[25]"},{"why":"Defines AMBM, the supervised continual-learning baseline with a memory buffer that SSOCL extends and must outperform.","marker":"[11]"},{"why":"Defines SCALE, the closest self-supervised online continual learning baseline that relies on data augmentation and is compared against.","marker":"[42]"},{"why":"Supplies the temperature-scaled softmax and confidence-based filtering idea that the memory enhancement module adapts for selecting low-noise samples.","marker":"[27]"},{"why":"Provides the InfoNCE-style contrastive loss used as the ablated replacement for the self-supervised module in the ablation study.","marker":"[5]"},{"why":"Documents the low signal-to-noise ratio of EEG, motivating the temporal-prediction objective instead of data augmentation.","marker":"[13]"}],"fun_headline_variants":["Self-supervised EEG model adapts on the fly, no labels, 71.78%","Bi-level memory-buffer learning hits 71.78% EEG emotion accuracy","Label-free continual learning for EEG emotion recognition reaches 71.78%","Dynamic memory buffer trains EEG emotion model without labels","Adapting EEG emotion recognition to new subjects with zero labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The weakest assumption is that k-means clustering, run with K fixed to the source class count on every incoming batch, yields pseudo-labels that correspond to the true emotion categories in that batch, even when the batch contains only one class.","fun_headline_variants_meta":{"raw":{"variants":["Self-supervised EEG model adapts on the fly, no labels, 71.78%","Bi-level memory-buffer learning hits 71.78% EEG emotion accuracy","Label-free continual learning for EEG emotion recognition reaches 71.78%","Dynamic memory buffer trains EEG emotion model without labels","Adapting EEG emotion recognition to new subjects with zero labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000808,"raw_usage":{"total_tokens":3548,"prompt_tokens":945,"completion_tokens":2603,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":2509}},"tokens_in":561,"tokens_out":2603,"duration_ms":17874,"temperature":1.0,"reasoning_tokens":2509,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:42:00.647730+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a labeled EEG stream and run SSOCL with labels hidden; compare the pseudo-labels assigned by k-means and cluster mapping to the true labels for each batch, computing cluster purity or pseudo-label accuracy. If those scores are near chance, or if a deliberately single-class batch is split into K spurious clusters without the buffer correcting them, the central claim that the memory buffer stores reliable pseudo-labels collapses.","supporting_citations":[{"cited_title":"Amigos: A dataset for affect, personality and mood research on individuals and 22 groups","cited_arxiv_id":null,"evidence_quote":"Supplies the AMIGOS EEG dataset used as one of the two continual-learning target domains."},{"cited_title":"Supervised contrastive learning","cited_arxiv_id":null,"evidence_quote":"Supplies the DEAP EEG dataset used as the main continual-learning target domain."},{"cited_title":"A multimodal psychological, physiological and behavioural dataset for human emotions in driving tasks","cited_arxiv_id":null,"evidence_quote":"Supplies the PPB-EMO driving EEG dataset used to pre-train the source model before streaming."},{"cited_title":"Online continual decoding of streaming eeg signal with a balanced and informative memory buffer","cited_arxiv_id":null,"evidence_quote":"Defines AMBM, the supervised continual-learning baseline with a memory buffer that SSOCL extends and must outperform."},{"cited_title":"Scale: Online self- supervised lifelong learning without prior knowledge","cited_arxiv_id":null,"evidence_quote":"Defines SCALE, the closest self-supervised online continual learning baseline that relies on data augmentation and is compared against."},{"cited_title":"Enhancing the reliability of out-of- distribution image detection in neural networks, in: International Con- ference on Learning Representations","cited_arxiv_id":null,"evidence_quote":"Supplies the temperature-scaled softmax and confidence-based filtering idea that the memory enhancement module adapts for selecting low-noise samples."},{"cited_title":"A simple framework for contrastive learning of visual representations, in: Pro- ceedings of the 37th International Conference on Machine Learning, pp","cited_arxiv_id":null,"evidence_quote":"Provides the InfoNCE-style contrastive loss used as the ablated replacement for the self-supervised module in the ablation study."},{"cited_title":"Self-supervised models are continual learners","cited_arxiv_id":null,"evidence_quote":"Documents the low signal-to-noise ratio of EEG, motivating the temporal-prediction objective instead of data augmentation."}],"review_version":1}