{"id":"1a3ead2e-6b80-4059-88b3-ced7276d89ad","arxiv_id":"2507.14537","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Using a pre-trained EEG decoder and temporal masking, the authors find concept-specific activation windows and prototypical temporal clusters in THINGS-EEG data.","lead":"This paper analyzes EEG recordings from people viewing object images and asks whether different object properties, such as 'plant' or 'furniture', are represented at different moments after the image appears. It uses a trained decoder plus a time-masking trick to produce per-concept activation timelines and clusters across ten subjects.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Temporal masking curves may reflect encoder artifacts rather than concept-specific neural timing, since no random-mask or permutation baseline is reported.","rationale":"The reader's weakest assumption and my concern are the same core issue: the masking curves are read directly as neural relevance without a control for decoder/encoder artifacts. My analysis sharpens this by noting that feeding a contiguous zero block to a pretrained waveform encoder is an out-of-distribution input, so the curve shape may reflect the encoder's mechanical response to missing data. The held-out-category design and the public THINGS-EEG dataset are genuine strengths, but they do not address this artifact. A permutation null is a concrete, decisive check: if the temporal peaks survive when EEG-concept pairing is destroyed, the paper's headline interpretation fails; if they vanish, the conditional acceptance can be upgraded. Since the reader already flagged this gap as the reason for CONDITIONAL, no verdict adjustment is needed.","tokens_in":8918,"tokens_out":6075,"duration_ms":81647,"concrete_test":"Retrain the ridge mapping on the same EEG embeddings but with the pairing between EEG embeddings and concept embeddings randomly permuted in the training set, destroying any true EEG-concept correspondence, then repeat the full temporal masking analysis of Section IV-C on the held-out test set. If the averaged (original minus masked) correlation curve still exhibits the two early peaks around mask start 0-25 and 30-60, or if concept-specific curves still separate into the reported clusters, the temporal specificity claim is a pipeline artifact rather than neural encoding. A flat or near-zero curve in this null would support the authors' interpretation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — that different object properties have temporally specific EEG representations (Section IV-C) — rests entirely on treating the drop in Pearson correlation after masking a 50-timepoint (200 ms) contiguous window as evidence that the masked window carried that concept's information. This inference is not secured. The EEG encoder f was trained on complete, unmasked signals; feeding it a masked signal X~ with a contiguous zero block (Eq. 3) produces inputs far outside its training distribution. The resulting change in the embedding, and hence in rho (Eq. 6), can be driven by the encoder's sensitivity to an artificial zero block, boundary effects, or positional discontinuities rather than by removal of concept-related neural activity. No control is reported: no random-masking baseline, no permuted-label null, and no check that the ridge mapping W transfers semantic content rather than low-level signal statistics. Consequently, the two 'concentrated visual activation' periods at mask start 0-25 and 30-60 (Section IV-C), and the subsequent concept ordering and clustering, could be properties of the encoder-decoder pipeline rather than of the brain. This is the load-bearing link between the measured curves and the claimed neuroscience result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework to uncover the temporal EEG representation of low-dimensional object properties (behavior-derived concepts) by combining a pre-trained EEG encoder, a pre-trained concept encoder, ridge regression mapping, and a temporal masking scheme. A contiguous 200 ms window of the EEG signal is zeroed, the affected embedding is mapped into concept space, and the Pearson correlation between the original and masked predicted concept embeddings is used as a measure of how much that time window contributes to the concept. The authors report that different concepts show temporally specific activation patterns, identify two periods of concentrated visual activation, and use dynamic time warping plus hierarchical clustering to group concepts into clusters they interpret as early visual, multi-feature integration, and late semantic processing. The conclusions are drawn from the THINGS-EEG dataset with held-out categories.","tokens_in":9096,"tokens_out":4553,"duration_ms":58315,"significance":"If the central interpretation is valid, the paper would provide a practical pipeline for localizing concept processing in EEG time, with possible value for BCI and cognitive neuroscience. The authors use a publicly available dataset, held-out categories for the ridge regression, and a transparent set of equations. The pre-trained encoders and the use of behavior-derived concept embeddings are useful ingredients, and the DTW-based clustering is a reasonable exploratory tool. However, the significance is heavily conditional: the temporal curves are deterministic functions of the fitted ridge mapping and the fixed encoder, and the leap from masking-induced correlation drops to neural relevance is not yet secured. The strengths of the manuscript are the clear experimental setup and the explicit equations, but the missing control analyses and independent validation currently prevent the central neuroscience claim from being supported.","major_comments":[{"comment":"The sentence 'Higher values indicate weaker relevance between the masked and original embeddings' is backwards. A higher Pearson correlation between the original and masked predicted concept embeddings means the two embeddings are more similar, so the masked segment had less impact on the prediction. The correct reading is that lower rho indicates a greater effect of masking. Since the temporal activation curves in Section IV and the clustering in Section IV-D all rely on this sign convention, the authors should correct Eq. (6)'s interpretation and re-verify that the plotted quantities ('difference in Pearson correlation values' in Fig. 4(b)) have the intended direction.","section":"Section III-B, Eq. (6)"},{"comment":"Zero-masking a contiguous block of EEG samples produces inputs that lie far outside the training distribution of the pre-trained EEG encoder f, which was trained on complete, unmasked signals. The observed drops in rho could therefore reflect the encoder's sensitivity to an artificial zero block, boundary discontinuities, or a general distribution shift, rather than the removal of concept-related neural activity. The manuscript reports no control condition: no random masking of equal-length windows, no permutation of concept labels, no noise-masking baseline, and no comparison with masking non-contiguous timepoints. Without such a control, the load-bearing link between the masking curves and 'temporally specific activation patterns' is not established.","section":"Section III-B (Eq. 3) and Section IV-B"},{"comment":"The temporal activation curves are computed from the fitted ridge mapping W and the fixed encoders via Eqs. (5)-(6), and the clustering in Section IV-D operates on these same curves. Consequently, the two 'concentrated visual activation' periods and the resulting concept clusters are properties of the authors' encoder-regression pipeline as much as of the brain; the interpretation that they reflect 'the hierarchical nature of neural processing' is not independently validated. The authors should provide a null distribution (e.g., ridge fits with shuffled labels or random concept-embedding assignments) or a direct neural validation (e.g., time-resolved decoding of concept properties from raw EEG) to show that the temporal structure survives when the mapping is unrelated to the semantic labels.","section":"Section IV-C and IV-D"},{"comment":"The claim that 'different concepts exhibit temporally specific activation patterns' is supported only by qualitative examples: three selected subjects, three visual stimuli, and the top-5 concepts for those images. No quantitative summary across the ten subjects and 200 test stimuli is given, no effect sizes or confidence intervals are reported, and no statistical test establishes that the ordering of activation times is consistent across subjects. The abstract and conclusion make general claims that require a systematic analysis, such as the distribution of activation latencies per concept across all subjects or a permutation test against a null of no temporal specificity.","section":"Section IV-B"},{"comment":"The clustering analysis uses a fixed number of clusters (K=5) without justification or stability assessment, and the labels 'early visual processing,' 'multi-feature integration,' and 'late semantic representation' are assigned post hoc. Moreover, the abstract's claim that this is 'the first' attempt to identify prototypical temporal characteristics of concepts in EEG is difficult to reconcile with reference [15], which the authors themselves cite for ridge regression and which studies the temporal dynamics of multidimensional object properties in the brain. The authors should compare their approach and findings with [15] explicitly, and provide a quantitative evaluation of cluster structure (e.g., silhouette scores or bootstrap stability) before drawing conclusions about brain organization.","section":"Section IV-D and Abstract"}],"minor_comments":[{"comment":"Please clarify the notation used for the 'difference' plotted in Fig. 4(b): the text mentions 'the difference in Pearson correlation values (original minus masked),' but Eq. (6) already defines a correlation between original and masked embeddings. Using a quantity such as Δρ = 1 − ρ would make the direction unambiguous.","section":"Section III-B and Fig. 4"},{"comment":"The caption says 'calculate the difference in predicted concept activation values before and after masking,' while the text and Eq. (6) describe a Pearson correlation between embeddings. These formulations should be aligned.","section":"Fig. 3 caption"},{"comment":"It is not fully clear whether the EEG encoder f is fine-tuned on the THINGS-EEG training split or used as a frozen pre-trained feature extractor. The phrase 'was trained according to the parameters specified in [22]' needs to be explicit about whether the weights are frozen or adapted for this dataset.","section":"Section IV-A"},{"comment":"There is a typo in 'there is a high probability of of being clustered together,' and the claim that 'clusters A through C have stronger activation values' would benefit from a definition of 'activation value' in terms of the correlation drop.","section":"Section IV-D"},{"comment":"The statement 'We are the first to attempt to identify the specificity and prototypical temporal characteristics of concepts within temporal distributions' is a strong novelty claim. Given the existing work on temporal dynamics of object properties in EEG (e.g., reference [15]), the authors should either provide a clear differentiation or soften the claim.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper is a conference-style contribution (accepted at IJCNN2025) and the journal version needs to substantially strengthen the empirical validation. The missing control analyses are not merely cosmetic: without them, the central temporal-specificity finding could be an artifact of the masking procedure and the fitted mapping. I would encourage the editor to request a revision that adds a random-masking/permutation baseline and an independent neural validation before further consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper does something genuinely new—temporal masking applied to EEG concept decoding on THINGS-EEG—and the held-out-category setup for ridge regression is a real positive. But the core claim, that different concepts have distinct temporal activation patterns reflecting neural processing, is not yet supported because the masking procedure is never tested against a null baseline.\n\nThe stress-test note gets this right. The pretrained EEG encoder was trained on complete signals; feeding it a zeroed 200 ms block is out of distribution. The resulting drop in correlation could be driven by boundary effects or the encoder's sensitivity to missing input, not by loss of concept-relevant neural activity. No random-mask control, permutation test, or even error bars are reported. That is the load-bearing gap between the curves and the neuroscience conclusion.\n\nWhat the paper does well: the equations are simple and internally consistent, the pipeline is reproducible in principle, and the specific empirical curves and clustering results on THINGS-EEG are not in the prior literature. The decision to measure the impact of masking via Pearson correlation between original and masked predicted embeddings is reasonable, though the text's description of the metric is confusing—it says higher rho means weaker relevance, then plots a \"difference in Pearson correlation values\" without pinning down the operational definition. That needs fixing. The reliance on the authors' own pretrained models (refs. 21, 22) is a minor concern, not fatal, but independent validation would strengthen the interpretation.\n\nIf the authors add a random-mask baseline and show the concept-specific timing survives it, this becomes a useful contribution. As it stands, it's a plausible pipeline with an unverified interpretation. I'd send it to peer review, because the method is coherent and the question matters, but I'd ask for controls before accepting the strong temporal-schema claims. It's also a good reading-group example of why decoder sensitivity is not identical to neural relevance.","headline":"A cleanly described masking pipeline for EEG concept timing, but the central temporal-specificity claim is under-supported without a null baseline.","tokens_in":9649,"tokens_out":2346,"would_cite":false,"duration_ms":28869,"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":"The paper argues that EEG temporal masking reveals when and in what order the brain encodes low-dimensional object properties, identifying early visual, integrative, and late semantic stages.","keywords":["EEG","temporal masking","concept representation","low-dimensional object properties","neural decoding","dynamic time warping","hierarchical clustering","brain-computer interface"],"falsifier":"Train the identical pipeline on the same EEG signals with image identities randomly permuted before fitting the concept-mapping regression; if the temporal masking curves still show the same concept-specific ordering and clusters, the curves are driven by encoder or data statistics rather than by concept representation.","tokens_in":8671,"feed_emoji":"🧠","tokens_out":7980,"duration_ms":84609,"temperature":0.7,"pith_summary":"This paper tries to establish that the brain's encoding of low-dimensional object properties can be tracked in time with EEG. The authors mask successive 50-timepoint windows in the EEG signal, re-encode the masked signal, and measure how much the predicted concept embedding changes; a large change marks the moment that concept is being processed. They report that different object concepts show distinct temporal activation curves, that these curves follow a sequential order, and that clustering them yields stages resembling early visual processing, multi-feature integration, and late semantic representation. If correct, this gives EEG a role in concept localization that complements fMRI's spatial maps.","feed_headline":"EEG masks reveal when each object concept activates","feed_subtitle":"Each concept has its own window of peak neural activity, from early visual features to late semantics.","key_machinery":"The load-bearing object is the temporal masking curve: for each concept and each start time $t_k$, a 50-timepoint segment of the multichannel EEG is zeroed, the masked signal is passed through the EEG encoder, a ridge regression predicts the concept embedding, and the Pearson correlation $\\rho(t_k, L)$ between original and predicted embeddings quantifies how much that time window matters. A dip in $\\rho$ marks a critical window for the concept. Dynamic time warping then measures the shape similarity of these curves across concepts, and hierarchical clustering groups the 42 concepts into prototypical temporal schemas.","core_discovery":"On the paper's own terms, the central discovery is that low-dimensional object properties are not uniformly distributed across the EEG timecourse. Each concept has a characteristic window or windows during which masking the signal most degrades the predicted concept embedding, so concepts can be ordered chronologically. The paper further reports that these temporal profiles cluster into a small number of prototypical shapes, which it interprets as hierarchical neural processing stages. Across the ten subjects, a two-phase pattern (an early window near mask starts 0 to 25 and a later window near 30 to 60) is consistent, although individual peak times vary.","pith_inferences":["Editorial inference: applying the same masking to shuffled EEG-image pairings would test whether the concept-specific order survives when image identities no longer match the signals; if it does, the curves may reflect encoder artifacts rather than neural representations.","Editorial inference: the claim that masking sensitivity equals neural relevance could be checked against independent neural measures, such as event-related potential latencies or source-imaging estimates for the same time windows.","Editorial inference: if the clustering is stable across subjects, concepts sharing a temporal cluster may share processing resources, predicting that behavioral similarity judgments should correlate with temporal-profile distances.","Editorial inference: the masking curves could transfer to naturalistic video stimuli, where concept activation could be timed against continuous behavior instead of fixed trial onsets."],"forward_implications":["EEG can be used to order concept processing chronologically, adding temporal specificity that fMRI spatial maps lack.","The identified early visual, integrative, and late semantic windows imply a staged neural hierarchy for object properties.","Concept-specific timing differences mean EEG decoding models should use time-resolved features rather than treating the whole epoch as one block.","Inter-subject temporal variability suggests that personalized time windows may matter for brain-computer interface decoding.","New concepts can be scored by their temporal profile through the same masking pipeline without retraining, as long as they lie in the concept space."],"supporting_citations":[{"why":"Supplies the pre-trained concept encoder that maps image embeddings into the 42-dimensional behavioral concept space used as the decoding target.","marker":"[21]"},{"why":"Supplies the EEG encoder architecture and pretrained weights that turn multichannel EEG into the latent embedding manipulated by masking.","marker":"[22]"},{"why":"Establishes the behavior-derived low-dimensional object properties and the regression-style mapping from brain signals to concept activation that this work adapts.","marker":"[15]"},{"why":"Provides the paired EEG-image dataset (1,854 concepts, rapid serial visual presentation) on which the model is trained and tested.","marker":"[24]"},{"why":"Provides the behavior-based low-dimensional object representation space that defines the 42 concept properties.","marker":"[17]"},{"why":"Supplies the dynamic time warping algorithm used to compare the shapes of temporal masking curves before clustering.","marker":"[23]"}],"fun_headline_variants":["EEG masks reveal object concepts' temporal windows","Each concept has its own EEG activation time","EEG timing shows how object concepts unfold","Temporal fingerprints of object concepts in EEG"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a drop in the correlation between unmasked and masked concept embeddings reflects the brain's own representation of that concept at the masked time window, rather than an artifact of the particular encoding models or regression mapping used.","fun_headline_variants_meta":{"raw":{"variants":["EEG masks reveal object concepts' temporal windows","Each concept has its own EEG activation time","EEG timing shows how object concepts unfold","Temporal fingerprints of object concepts in EEG"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000181,"raw_usage":{"total_tokens":1253,"prompt_tokens":835,"completion_tokens":418,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":451,"completion_tokens_details":{"reasoning_tokens":362}},"tokens_in":451,"tokens_out":418,"duration_ms":5926,"temperature":1.0,"reasoning_tokens":362,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:53:37.183991+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the identical pipeline on the same EEG signals with image identities randomly permuted before fitting the concept-mapping regression; if the temporal masking curves still show the same concept-specific ordering and clusters, the curves are driven by encoder or data statistics rather than by concept representation.","supporting_citations":[{"cited_title":"Cocog: controllable visual stimuli generation based on human concept representations,","cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained concept encoder that maps image embeddings into the 42-dimensional behavioral concept space used as the decoding target."},{"cited_title":"Dynamic representation of multidimensional object properties in the human brain,","cited_arxiv_id":null,"evidence_quote":"Establishes the behavior-derived low-dimensional object properties and the regression-style mapping from brain signals to concept activation that this work adapts."},{"cited_title":"Human eeg recordings for 1,854 concepts presented in rapid serial visual presentation streams,","cited_arxiv_id":null,"evidence_quote":"Provides the paired EEG-image dataset (1,854 concepts, rapid serial visual presentation) on which the model is trained and tested."},{"cited_title":"Dynamic programming algorithm optimization for spoken word recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies the dynamic time warping algorithm used to compare the shapes of temporal masking curves before clustering."}],"review_version":1}