{"id":"b7d819bc-9b0c-4728-bc3e-660d5ee3649e","arxiv_id":"2412.15560","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Training EEG encoders with an auxiliary InfoNCE loss that predicts a co-trained music encoder's representation improves 10-song EEG identification accuracy on the NMED-T dataset.","lead":"The paper trains an EEG model to identify which of 10 songs a person is listening to, using a second neural network trained on the music itself as a teaching signal. The method reports accuracy gains on a public EEG dataset, but the target 'ANN representation' is trained jointly on the same labels rather than taken from a fixed pretrained network.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'ANN representation' target is randomly initialized and co-trained on the same 10 labels (Eqs. 1-3), so the reported gain may be audio-EEG distillation or regularization rather than knowledge from an independent pretrained ANN.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the supervisory 'ANN representation' is not a fixed, externally pretrained representation but a randomly initialized encoder co-trained on the same song labels and the same data split. This matters because the paper's motivation and title promise a test of whether ANN representations can complement EEG; if the target is co-trained, the result is compatible with a much weaker explanation, namely that an auxiliary audio-to-EEG contrastive objective regularizes or stabilizes training. The proposed frozen-encoder experiment would settle this by holding the target fixed and external. I agree with the CONDITIONAL verdict: the reported accuracy gains, especially in the 2D CNN comparison, could survive a proper control, but the current paper does not provide that control. Secondary issues such as the post hoc delay selection and the lack of subject-independent evaluation are real but less central to the causal attribution claim; they would become more important if the frozen-encoder experiment succeeds.","tokens_in":16208,"tokens_out":9007,"duration_ms":89265,"concrete_test":"Run the exact best 2D-CNN pipeline (delay 200 ms, lambda=0.05, same seeds and split) with three music-encoder conditions: (A) a frozen SampleCNN pretrained on an external large music corpus, (B) a frozen randomly initialized SampleCNN, and (C) the paper's co-trained encoder. If (A) preserves the roughly 0.077 average gain over the lambda=0 baseline (0.624 vs 0.547) or exceeds (C), the ANN-representation claim is supported. If only (C) beats baseline, the improvement is an artifact of co-training the target on the same labels, not evidence for ANN-to-brain complementation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ANN representations provide complementary information to noisy EEG. In Eqs. (1)-(3), however, the music encoder that defines the 'ANN representation' is a randomly initialized CNN, structurally identical to the EEG encoder, and is trained jointly on the same 10 song labels via LclsM; the PredANN term only applies stop-gradient to the music embedding. The target is therefore a co-trained, moving embedding, not an independently validated ANN representation of the kind used in brain-encoding similarity studies. With this design, the lambda=0 vs lambda=0.05 comparison cannot separate 'knowledge transferred from a pretrained ANN' from (i) a cross-modal distillation regularizer that pulls EEG toward an audio representation learned on the same training split, (ii) stabilization of optimization by a second view, or (iii) label-driven drift of the target during training. The improvement may be genuine and practically useful, but the paper's mechanistic claim is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an auxiliary 'PredANN' loss that trains an EEG encoder to predict embeddings produced by a music CNN, while both branches also solve a 10-song classification task. On the NMED-T dataset, the method reports average accuracy of 0.624 over three seeds for 3-second clips with a 2D CNN, versus 0.547 for the classification-only baseline, and 0.783 at 7 seconds with the mean scoring method. The paper interprets these gains as evidence that ANN representations can complement noisy EEG signals for music identification.","tokens_in":16406,"tokens_out":4334,"duration_ms":38532,"significance":"If the improvement were shown to derive from a fixed, independently trained ANN representation, this would be a novel way to exploit brain-ANN similarity for neural decoding. The empirical recipe may still be useful as a cross-modal distillation technique, and the authors use a public dataset, describe preprocessing in detail, and report McNemar tests. However, the current design does not separate the hypothesized mechanism from generic regularization or co-training drift, so the significance claim is conditional on a stronger experimental control.","major_comments":[{"comment":"The load-bearing claim is that training an EEG model to predict ANN representations improves recognition. But the 'ANN representation' is not a fixed pretrained model: the music encoder is randomly initialized and trained jointly on the same 10 song labels through LclsM, with only stop-gradient applied to the PredANN term. Therefore the comparison between lambda=0 and lambda=0.05 cannot separate knowledge transferred from an independent pretrained ANN from (i) cross-modal distillation toward an audio representation learned on the same training split, (ii) optimization stabilization from a second view, or (iii) label-driven drift of the target during training. To support the central claim, the authors should either use a frozen, pretrained music-identification ANN as the target, or re-scope the claim to co-trained cross-modal distillation and include a control with a randomly initialized, untrained target.","section":"Methods, Eqs. (1)-(3)"},{"comment":"The 1D CNN baseline collapses to chance-level accuracy of 0.100 on seeds 0 and 1, while reaching 0.486 and 0.474 on seeds 2 and 42. This bimodal behavior indicates optimization failure in the baseline, not a typical performance level, so the reported average of 0.324 is dominated by two failed runs. The claim that PredANN makes learning robust may be true, but the comparison is misleading as presented. The authors should report results conditioned on successful baseline training, fix the baseline's training procedure, or use multiple restarts and report the full distribution of outcomes.","section":"Results: Robustness Testing, Table 1"},{"comment":"The headline improvement of the 2D CNN over its own classification-only baseline is not statistically significant for seed 1 (p=0.367), one of the three seeds. The average gain (0.624 vs 0.547) therefore rests on inconsistent seed-level differences. The authors should report per-seed confidence intervals, increase the number of seeds, or otherwise demonstrate that the effect is not driven by a subset of initializations.","section":"Results: 2D CNN vs 1D CNN, Table 3"},{"comment":"The 200 ms delay was selected after scanning 80, 160, 320, and 640 ms, then adding 240 and 480 ms, and finally testing 200 ms, all on the same evaluation data. No correction for multiple comparisons is applied, and no separate validation set is used for delay selection. The reported p-values for 200 ms versus 0 ms (seed 0 p<0.001, seed 1 p=0.0043) are post-selection and therefore overstate significance. The delay should be treated as a selected hyperparameter with a held-out validation split, or the full search and a multiple-testing correction should be reported.","section":"Results: Incorporating Time-delay, Table 2 and Figure 2"}],"minor_comments":[{"comment":"The phrase 'ANN representations' is used throughout to describe the target, but the music encoder is randomly initialized and co-trained on the same labels. Please clarify this in the abstract and introduction so that readers are not misled about the nature of the supervisory signal.","section":"Abstract and Introduction"},{"comment":"The indices in Eq. (2) run from 0 to B, but a mini-batch of size B typically has indices 0 to B-1; please fix the indexing and clarify the batch-size convention.","section":"Methods, Eq. (2)"},{"comment":"The captions of Tables 1 and 3 should identify which rows are baseline models and which are proposed models; currently the reader must infer this from the text.","section":"Results, Table 1 and Table 3"},{"comment":"Please report the optimizer, learning rate, weight decay, batch size, and the meaning of 'stride of 200' in terms of samples or time steps; these details are needed for reproducibility.","section":"Methods: Model Training and Evaluation"},{"comment":"The comparison with Avramidis et al. (Table 4) is not a controlled comparison because the architecture, delay, loss weights, and evaluation protocol differ. Presenting this as an indicative benchmark rather than a head-to-head comparison would be more accurate.","section":"Results: Previous Study Comparison"},{"comment":"Reference 42 is cited as a PsyPost article, which is not a peer-reviewed source; please replace it with the underlying primary literature if available.","section":"Discussion"}],"recommendation":"major_revision","confidential_remarks":"The paper's framing and abstract promise a demonstration that ANN representations complement brain recordings for decoding. Given that the target is a co-trained, randomly initialized music encoder, the experiments cannot yet distinguish the claimed mechanism from generic cross-modal distillation or regularization. A revision that either uses a frozen pretrained music model or explicitly re-scopes the claim to co-trained distillation would be necessary before I could recommend acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper is worth a serious look, but the central mechanistic claim is not yet supported. What's new: they invert the usual brain-encoding setup—instead of regressing ANN features to cortical responses, they train an EEG classifier to predict features from a music encoder, with a classification loss on both branches and stop-gradient on the music side. That specific recipe, applied to 10-song identification on NMED-T, is not something I've seen in the prior work they cite. Good things: public dataset, multiple seeds, baseline and prior-method comparisons, delay sweep, longer-window evaluation, per-song and per-subject analyses. They also honestly report non-significant seeds, which is more than many papers do.\n\nThe main soft spot is the target. The 'ANN representation' comes from a randomly initialized CNN, structurally identical to the EEG encoder, trained jointly on the same 10 labels. Stop-gradient keeps EEG gradients from shaping the music encoder, but the music encoder still drifts with the labels during training. So the comparison lambda=0 vs lambda=0.05 can't separate 'knowledge from a pretrained ANN' from cross-modal distillation or optimization stabilization. The gain may be real and practically useful—the 2D results are decent—but the paper's story about ANN/brain similarity doesn't follow from these experiments.\n\nOther issues: the 200 ms delay is chosen after scanning a grid of delays, and the McNemar tests are applied to the same data without correction, so the p-values are over-claimed. On 1D seeds 0 and 1 the baseline collapses to 0.100, making those comparisons less informative. The 2D improvement over baseline is not significant for seed 1. And the data split is stratified by song, not by subject, so no subject-independent evaluation is reported; with EEG, that typically inflates accuracy. Code is promised only after acceptance.\n\nThe math itself is simple and clean; no red flags in the loss equations. Citation pattern looks fine, with appropriate credit to CLIP/ConVIRT and Avramidis et al.\n\nWho this is for: people working on EEG decoding, music BCI, or cross-modal contrastive learning. It deserves a serious referee, but acceptance should hinge on code release, a frozen pretrained target, and a subject-independent split. I'd engage with it.","headline":"A genuinely interesting training trick for EEG music identification, but the paper oversells the 'ANN representation' mechanism: the target is a co-trained audio encoder, not a fixed pretrained one.","tokens_in":16948,"tokens_out":2534,"would_cite":true,"duration_ms":23082,"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":"Training an EEG recognition model to predict ANN representations of music substantially improves music identification accuracy from noisy brain recordings.","keywords":["music identification","EEG decoding","contrastive learning","ANN representations","neural decoding","brain-computer interface","NMED-T dataset","auditory perception"],"falsifier":"Train the same model but replace the jointly-trained music encoder with a fixed, pretrained music model that is not updated by the PredANN loss, then compare accuracy; if accuracy does not exceed the classification-only baseline, the reported benefit is not due to predicting ANN representations.","tokens_in":16005,"feed_emoji":"🎵","tokens_out":7105,"duration_ms":55873,"temperature":0.7,"pith_summary":"The paper proposes a new training strategy for decoding music from EEG: instead of classifying brain recordings directly, the model is also trained to predict the feature representation that an artificial neural network (ANN) produces for the same music. The authors show that adding this auxiliary contrastive loss (called PredANN) to the classification objective substantially improves music identification accuracy, raising the best 2D CNN from a 0.547 classification-only baseline to 0.624 average accuracy on 3-second clips, and to 0.783 at 7 seconds with mean scoring. The improvement is robust across random seeds, is strongest when the EEG is delayed by about 200 ms relative to the music, and is larger with a 2D CNN than a 1D CNN. The paper interprets this as evidence that ANN representations can complement noisy, non-invasive brain recordings, and argues the framework could benefit brain-computer interfaces and the study of music cognition.","feed_headline":"ANN-guided training lifts EEG song identification to 78%","feed_subtitle":"Predicting music-ANN features from EEG beats classification-only training and keeps improving to 7-second clips.","key_machinery":"The load-bearing mechanism is the PredANN loss: an InfoNCE-style contrastive loss (Eq. 2) that pulls the EEG encoder's feature embedding toward the music encoder's feature embedding for the same song while repelling embeddings from different songs, with a stop-gradient operation that prevents the music encoder from being updated by this loss. The model also jointly trains both branches on the 10-way song classification task (Eq. 1), and the total loss is the sum of the two classification losses plus the weighted PredANN loss (Eq. 3). The stop-gradient is the key design choice: it keeps the ANN representations discriminative for the target task rather than letting EEG noise corrupt them, and the paper shows that removing it degrades accuracy. The same architecture is used for both encoders, a 2D CNN based on prior work, with the contrastive head branching off after the main classification head.","core_discovery":"The central discovery is that a recognition model for brain recordings can be trained to predict the representations of an artificial neural network processing the same auditory stimulus, and this reverses the usual direction of ANN-brain alignment studies: instead of regressing from ANN to cortical activity, the authors regress from EEG to ANN features as a supervisory signal. Concretely, the model optimizes the sum of an EEG classification loss, a music classification loss, and a weighted InfoNCE-style contrastive loss between the two modalities, with a stop-gradient applied to the music branch so that the ANN encoder is not distorted by EEG noise. On the NMED-T dataset, the best configuration (2D CNN, PredANN weight 0.05, 200 ms delay) attains 0.624 average accuracy over three seeds on 3-second clips versus 0.547 for the classification-only baseline, and 0.783 at 7-second evaluation with the mean scoring rule. The authors further report that the improvement is statistically significant for most seeds, that the optimal 200 ms delay matches known auditory response latencies, and that longer evaluation windows monotonically improve accuracy without retraining.","pith_inferences":["Because the music encoder is randomly initialized and co-trained rather than a fixed pretrained ANN, the paper's claim that 'ANN knowledge' complements EEG is not fully established; a direct test would be to fix a pretrained music-encoding network and see whether the same gain persists.","If the benefit is mainly a regularizing or optimization effect rather than genuine ANN-brain alignment, the method may transfer to any paired stimulus-recording domain, such as speech, video, or imagined music, where a co-trained 'teacher' representation is available.","The 200 ms peak could serve as a calibration sanity check for other EEG decoding pipelines: a pipeline that shows no such latency dependence may be learning spurious features.","A testable extension is to vary the strength of the stop-gradient, for example by applying it only every other batch, to map how much ANN-branch plasticity hurts EEG accuracy."],"forward_implications":["Increasing the EEG evaluation length from 3 to 7 seconds via overlapping windows improves accuracy (from 0.716 to 0.783 with mean scoring) without any additional training, supporting real-time sliding-window decoding.","The 200 ms delay between music onset and EEG input consistently improves accuracy, aligning with known auditory response latencies and suggesting that temporal alignment matters for EEG decoding.","The stop-gradient operation on the music branch is essential: removing it drops average accuracy from 0.624 to 0.497, supporting the claim that preserving ANN discriminative power is key.","The proposed point-to-point contrastive alignment substantially outperforms the set-to-set gradient-reversal domain adaptation of a prior study (0.624 vs about 0.159 average) on the same 10-class task.","Song and subject analyses show that accuracy depends on both stimulus distinctiveness (songs with electronic or unusual features are easier) and individual differences, which the authors interpret as reflecting neural response salience."],"supporting_citations":[{"why":"Supplies the NMED-T EEG dataset of 20 subjects listening to 10 songs that all experiments are run on.","marker":"31"},{"why":"Provides the InfoNCE/CPC formulation that the PredANN contrastive loss is based on.","marker":"28"},{"why":"Establishes the multi-modal contrastive learning approach (CLIP) that the two-branch projection design extends.","marker":"30"},{"why":"Defines the prior set-to-set domain adaptation method (GRL) that the proposed point-to-point alignment is compared against.","marker":"35"},{"why":"Supports the premise that ANN representations resemble cortical responses to audio, motivating the reversed prediction direction.","marker":"9"},{"why":"Provides evidence that cortical encoding of musical features occurs around 200 ms, grounding the delay tuning experiment.","marker":"8"}],"fun_headline_variants":["Reverse ANN to EEG supervision lifts music ID to 78%","EEG trained on ANN features hits 78% accuracy for song ID","Training EEG to predict ANN representations improves music ID to 78%","ANN supervision flips direction: EEG predicts features, hits 78%","Invert the mapping: ANN-guided EEG reaches 78% on music ID"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The supervisory 'ANN representation' comes from a music encoder that is randomly initialized and trained jointly with the EEG model on the same ten songs, so it is not an independently pre-trained ANN representation.","fun_headline_variants_meta":{"raw":{"variants":["Reverse ANN to EEG supervision lifts music ID to 78%","EEG trained on ANN features hits 78% accuracy for song ID","Training EEG to predict ANN representations improves music ID to 78%","ANN supervision flips direction: EEG predicts features, hits 78%","Invert the mapping: ANN-guided EEG reaches 78% on music ID"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000556,"raw_usage":{"total_tokens":2649,"prompt_tokens":951,"completion_tokens":1698,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":1604}},"tokens_in":567,"tokens_out":1698,"duration_ms":10493,"temperature":1.0,"reasoning_tokens":1604,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:18:36.401508+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same model but replace the jointly-trained music encoder with a fixed, pretrained music model that is not updated by the PredANN loss, then compare accuracy; if accuracy does not exceed the classification-only baseline, the reported benefit is not due to predicting ANN representations.","supporting_citations":[{"cited_title":"T., Dmochowski, J","cited_arxiv_id":null,"evidence_quote":"Supplies the NMED-T EEG dataset of 20 subjects listening to 10 songs that all experiments are run on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the multi-modal contrastive learning approach (CLIP) that the two-branch projection design extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the premise that ANN representations resemble cortical responses to audio, motivating the reversed prediction direction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides evidence that cortical encoding of musical features occurs around 200 ms, grounding the delay tuning experiment."}],"review_version":1}