{"id":"41e9fd13-2f7b-4f73-b9fa-8ae3271303e0","arxiv_id":"2412.03192","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A two-stage pipeline first learns features with Hebbian updates (no backprop), including new transpose-convolution rules, then fine-tunes on few labels, improving Dice on several medical segmentation benchmarks.","lead":"A segmentation network's first training stage can run on the brain-inspired rule 'fire together, wire together' instead of backpropagation. The paper adds new Hebbian rules for upsampling layers and shows the pretrained weights improve label-efficient medical image segmentation on most tested benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline superiority claim is not established by the reported statistics: most DC gaps are within overlapping 90% CIs, the LA results contradict 'across all label levels,' and the per-dataset SWTA temperature in Table 5 has no documented validation protocol.","rationale":"The reader's verdict is CONDITIONAL, and my stress-test reinforces that conditional status rather than moving it. I agree with the reader's observation that the abstract overclaims and that gains often sit inside overlapping CIs, and I also notice the temperature-selection protocol is undocumented. I do not select the TSA convergence gap as the single most load-bearing issue: although Appendix A proves only the forward HPCA setting and not the TSA variant, the update rule is a direct transpose of the standard Hebbian/Sanger rule and the empirical SWTA-TSA ablation (Table 6) shows it outperforms the naive SWTA-S variant, so the theoretical gap is plausible but not the main threat to the central claim. The main threat is that the empirical evidence, as reported, does not currently establish 'substantially outperforms SOTA across different levels of label availability' because the comparisons lack paired significance testing, the LA 3D results are inconsistent with the claim, and the per-dataset temperature may have been selected on test. A single reproducible re-analysis with a fixed or validation-selected temperature and paired tests would settle whether the concern lands. If it does land, the paper remains conditionally acceptable after softening the claim and adding the missing statistical/protocol details.","tokens_in":24492,"tokens_out":12341,"duration_ms":117219,"concrete_test":"Run the released code on GlaS/PH2/HMEPS/LA with a single fixed temperature (e.g., t=20) chosen a priori or on validation, and compute paired bootstrap or Wilcoxon signed-rank tests over the 10 folds for Ours versus the best SOTA baseline at every regime. If the majority of wins are not significant at p<0.05, or if LA 10%/20% remain worse by more than 1 DC point, the paper should be revised to claim 'competitive or better on 2D, not consistently better on 3D' rather than 'outperforms SOTA across different levels of label availability.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the two-stage Hebbian pretraining 'outperforms SOTA approaches across different levels of label availability' (Abstract) and 'substantially improves' results. For this to hold, the reported gains must be real and the comparisons fair. Three facts make this condition insecure. First, in Tables 2-4 the DC advantage over the best baseline is typically 0.2-1.3 points and the 90% confidence intervals overlap in nearly every regime; with 10-fold splits this does not establish a significant difference unless paired tests are reported. Second, the 3D LA results in Table 8 contradict the across-the-board claim: at 2%, 5%, 10%, and 20% labels, Ours is not the best method, and at 10%/20% UAMT beats it by 1.84 and 1.74 DC points. Third, Table 5 selects a different SWTA temperature per dataset (100 for GlaS, 20 for PH2/HMEPS), but the paper never states that this choice was made on a validation split; if temperatures were selected on test, the comparison with default-tuned baselines is invalid. The consistent direction on 2D datasets and the linear-probing ablation (Table 7) are supportive, but they do not overcome these protocol issues. The absence of a convergence proof for the TSA transpose-convolution rules is a secondary concern: even if the rules are heuristic, the empirical claim still needs the statistical and protocol checks above.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage semi-supervised semantic segmentation pipeline for biomedical imaging. In the first stage, an encoder-decoder (UNet-like) architecture is pretrained without labels using Hebbian learning rules (SWTA and HPCA), including newly formulated rules for transpose-convolutional layers (SWTA-TSA and HPCA-TSA). In the second stage, the pretrained weights are fine-tuned with standard backpropagation on a small labeled subset. The method is evaluated on three 2D datasets (GlaS, PH2, HMEPS) and one 3D dataset (LA), across label regimes from 1% to 20%. The authors report Dice/Jaccard/boundary metrics, compare with several single- and two-stage semi-supervised baselines, and show that initializing those baselines with their Hebbian pretraining often improves performance. The central claim is that the proposed method outperforms SOTA across different levels of label availability.","tokens_in":24800,"tokens_out":4884,"duration_ms":43632,"significance":"If the claims are substantiated, the paper makes a useful contribution by showing that biologically plausible Hebbian updates can pretrain upsampling layers without labels, complementing prior work that focused on convolutional or fully connected layers. The release of code and the inclusion of a linear-probing ablation are strengths, as is the evaluation across multiple imaging modalities and a volumetric dataset. However, the manuscript's headline superiority claim is not yet supported by the reported statistics or by its own 3D results, and several protocol details (temperature selection, baseline tuning, significance testing) need to be clarified before the empirical contribution can be fully assessed.","major_comments":[{"comment":"The claim that the method outperforms SOTA is not supported by the reported statistics in the 2D experiments. For example, on GlaS at 1% labeled data, Ours achieves DC 69.95 ± 1.09 versus CPS at 69.32 ± 0.59; on PH2 at 10%, Ours achieves 88.26 ± 0.51 versus URPC at 88.06 ± 0.40. In nearly all regimes the 90% confidence intervals overlap substantially, and no paired significance tests over the ten runs are reported. Without paired tests (e.g., Wilcoxon signed-rank or paired t-test), the numerical advantages of 0.2–1.3 DC points do not establish a reliable improvement. Please add appropriate statistical tests or temper the abstract and conclusions to 'numerically higher in most 2D regimes' where significance is not demonstrated.","section":"§5.2, Tables 2–4"},{"comment":"The abstract's claim of outperforming SOTA 'across different levels of label availability' is directly contradicted by the paper's own LA results. At 10% and 20% labels, UAMT exceeds Ours by 1.84 and 1.74 DC points (88.95 vs. 87.11 and 90.91 vs. 89.17, respectively), and at 2% and 5% DTC or UAMT are best. Similarly, on HMEPS at 1% (Table 4), CCT achieves 91.09 vs. Ours 90.75. The paper itself acknowledges the HMEPS 1% case but the abstract remains overbroad. Please restrict the claims to the regimes and datasets where the method is actually best, or provide additional experiments/analysis to support a global claim.","section":"§5.5, Table 8, and Abstract"},{"comment":"The per-dataset SWTA temperature (100 for GlaS, 20 for PH2 and HMEPS) appears to be selected on the test set: no validation protocol is described for this choice. Since the main results in Tables 2–4 use these selected temperatures, any selection based on test performance would invalidate the comparison with baselines that used default or differently tuned hyperparameters. Please document explicitly how the temperature (and the learning rates for competitor methods mentioned in Appendix B) were chosen on a validation split, and report the sensitivity of the final results to these choices.","section":"§5.4, Table 5, and Appendix B"},{"comment":"The newly proposed TSA Hebbian rules for transpose-convolutional layers are the main theoretical novelty, but the manuscript does not provide a convergence or pattern-discovery analysis for them. Appendix A proves the HPCA equilibrium only for the forward convolutional case; the TSA variant reverses the roles of input and output and introduces a new reconstruction block ρ*, so the claim that SWTA-TSA/HPCA-TSA inherit the clustering/PCA behavior of the forward rules is an unproven inductive assumption. Please either provide an equilibrium/convergence analysis for the TSA rules or explicitly frame them as heuristics whose usefulness is established empirically (e.g., via the linear probing and initialization results).","section":"§4.2 and Appendix A"},{"comment":"The description of baseline hyperparameter tuning is not sufficiently specified. The statement that for some competitor techniques the authors 'opted for a lower initial learning rate, as we observed better results empirically' does not constitute a reproducible protocol. It is unclear whether the same validation-based selection procedure was applied to the proposed method and to all baselines, or whether the baselines were tuned with equal effort. Please specify the exact selection procedure for all hyperparameters of all methods, or the fairness of the comparison cannot be assessed.","section":"§5.2 and Appendix B"}],"minor_comments":[{"comment":"There is a typo: 'Straightfarward' should be 'Straightforward' in the description of the first T-Conv strategy.","section":"§4.2"},{"comment":"For GlaS, the best temperature is the maximum tested (100), so the statement that performance 'reaches a plateau at that point' is not warranted; higher values should be tested or the wording should be softened.","section":"§5.4, Table 5"},{"comment":"The linear probing results are reported without confidence intervals or a description of the probing protocol (which features are used, classifier details, number of runs). Please add these details for consistency with the rest of the evaluation.","section":"§5.4, Table 7"},{"comment":"The y-axis ranges are described as 'embedded in the most convenient range for best readability,' which can visually exaggerate small differences. Please use fixed axis ranges or add error bars so that the improvements/worsenings are not misleading.","section":"Figure 4"},{"comment":"The phrase '10-fold cross-validation protocol (5-fold for the LA dataset), varying the training and test splits' is ambiguous. Please clarify whether this is standard k-fold cross-validation or k independent random splits, and how the validation split for model selection was obtained.","section":"Appendix B"},{"comment":"The wording 'substantially improves performance' is stronger than the quantitative evidence in the tables. Please align the abstract and introduction with the actual effect sizes and statistical support.","section":"Abstract and §1"},{"comment":"Reference [54] is listed as 'to appear'; if the ECCV 2024 workshop paper has been published, please update the citation.","section":"References"},{"comment":"The text contains typographical artifacts such as 'di fferent' and 'e fficient' that should be corrected in the final version.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript extends the authors' own ECCV 2024 workshop paper [54]. The novelty over that paper appears incremental (new TSA rules, additional datasets, initialization experiments), but the editor may wish to verify the overlap with the workshop paper. The journal scope is appropriate for this topic. The main concern is that the empirical claims are not yet supported by the reported statistics and protocol details; however, the issues are addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work on semi-supervised medical segmentation or Hebbian learning. The genuinely new piece is the TSA (transposed-structure-aware) Hebbian rule for transpose-convolutional layers; the earlier workshop paper had only a preliminary version. The authors also do a fair thing: they reimplement the single-stage baselines under a common UNet and report 10-fold means with 90% CIs, plus a linear-probing ablation that isolates the value of the pretraining stage. Code is public. On the 2D benchmarks (GlaS, PH2, HMEPS) the method consistently tops or ties the best baseline on Dice, usually by 1-3 points, and initializing the SOTA methods with the Hebbian weights helps more often than it hurts. That is a real, reproducible empirical contribution.\n\nThe soft spots are not fatal but they are real. First, the abstract claims superiority \"across different levels of label availability,\" and Table 8 on LA does not support that: at 2%, 5%, 10%, and 20% the method is not best, and UAMT beats it by about 1.8 points at 10% and 20%. The text hedges (\"ranks best or second best\"), but the abstract was not updated. Second, most Dice gaps on the 2D tables sit inside overlapping 90% CIs; with ten runs, overlapping intervals do not establish a significant difference unless paired tests are reported. The consistent direction across datasets helps, but it is not a substitute for a paired test. Third, the temperature sweep in Table 5 picks different values per dataset (100 for GlaS, 20 for PH2/HMEPS) with no stated validation protocol. If those values were chosen on test, the comparison with default-tuned baselines is invalid; the paper needs to say how t was selected. The missing convergence proof for the TSA rules is a secondary issue: the empirical claim does not depend on a proof, but the paper should be honest that TSA is a heuristic extension. Even the linear-probing table (Table 7) reports single numbers without CIs; the gap over VAE is small.\n\nWho this is for: people building label-efficient pipelines for biomedical segmentation, and the Hebbian learning community. It deserves a serious referee: the method is clearly described, reproducible, and mostly honestly discussed. I would send it out but ask for (1) a corrected abstract, (2) paired significance tests or at least error bars on the ablation and initialization figures, and (3) a description of how the temperature was chosen. After that, it could be a solid contribution.","headline":"Worth a look if you work on semi-supervised medical segmentation or Hebbian learning; the new TSA rule for transpose-conv layers is a genuine contribution, but the abstract overclaims what the paper's own tables show.","tokens_in":25360,"tokens_out":1733,"would_cite":false,"duration_ms":15026,"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 claims that a two-stage pipeline—unsupervised Hebbian pretraining of both the downsampling and upsampling paths, followed by a short backpropagation fine-tuning on a few labels—beats current semi-supervised segmentation methods…","keywords":["Hebbian learning","semi-supervised learning","semantic segmentation","biomedical imaging","transpose-convolutional layers","SWTA","HPCA","encoder-decoder pretraining"],"falsifier":"Fine-tune the full model from the SWTA-TSA initialization and from the same initialization with only the transpose-convolutional layers reset to random weights; if the Dice scores are statistically indistinguishable across several seeds, the new T-Conv rule is not carrying the claimed benefit.","tokens_in":24240,"feed_emoji":"🧠","tokens_out":11095,"duration_ms":94408,"temperature":0.7,"pith_summary":"The paper tries to show that an encoder-decoder segmentation network can be pretrained entirely without labels using Hebbian “fire together, wire together” updates, and that this pretraining makes the subsequent supervised fine-tuning data-efficient enough for biomedical imaging. The first stage applies local Hebbian rules to the ordinary convolutional layers and to the transpose-convolutional (upsampling) layers, for which the paper derives new SWTA-TSA and HPCA-TSA update rules. The second stage fine-tunes the pretrained weights with standard backpropagation on a small labeled subset. On four biomedical datasets across 1% to 20% label regimes, the method reports higher Dice overlap scores than state-of-the-art pseudo-labeling, consistency-training, variational-autoencoder, and self-supervised baselines, and it improves those baselines when used as their initialization. The practical upshot, if the claim holds, is that accurate pixel-level segmentation needs far fewer annotated images.","feed_headline":"Label-free first stage lifts medical segmentation accuracy","feed_subtitle":"A backprop-free first stage learns from unlabeled images, so fine-tuning needs only 1-20% of the labels.","key_machinery":"The load-bearing machinery is the generalized Hebbian update $\\Delta w_{i,j} = \\eta\\, g_j\\,(s_i - s^*_{i,j})$, built from three blocks: a gating block $\\gamma(y)$ that sets the update size, a reconstruction block $\\rho(y,w)$ that reproduces the layer input from its outputs, and the plasticity function that moves the weight along the difference between a target signal $s_i$ and a reconstruction $s^*_{i,j}$. For transpose-convolutional layers the paper adds the named Transposed-Structure-Aware reconstruction block $\\rho^*(y,w)$: it takes patches from the upsampled feature map at different offsets and combines them with the weight matrix through an ordinary convolution, fixing the shape mismatch while preserving the SWTA/HPCA pattern-discovery reading. This block is what allows the upsampling path to be pretrained without backpropagation or labels.","core_discovery":"The central discovery the authors aim to establish is that a bio-inspired local learning rule, applied layer-by-layer and including the upsampling path, produces a useful unsupervised initialization for semi-supervised semantic segmentation. The main technical step is the definition of Transposed-Structure-Aware (TSA) Hebbian rules for transpose-convolutional layers, where a new reconstruction block $\\rho^*$ treats the downsampled feature map as the target and derives gate and reconstruction signals from the upsampled feature map, so that the clustering (SWTA) or principal-component (HPCA) interpretation survives the reversed input-output geometry. In their comparisons, SWTA-TSA is the best of the four derived rules, and supplying its pretrained weights as initialization improves most single-stage semi-supervised competitors on most datasets and label regimes. The authors also report that the same two-stage pipeline carries over to 3D volumetric MRI segmentation, where it finishes best or second best in most settings.","pith_inferences":["Beyond the paper's experiments, the TSA shape-mismatch argument applies to any upsampling layer, so SWTA-TSA could be tested as a label-free pretraining step for natural-image super-resolution and image synthesis.","Because the first stage uses only local updates and no backpropagation, it may run on neuromorphic or energy-constrained hardware, but the paper does not measure energy consumption, so that benefit remains an open empirical question.","The paper's ablation shows SWTA-TSA beats HPCA-TSA on GlaS, which suggests cluster discovery is a better inductive bias than PCA-style decorrelation for pixel-wise classification; a cross-task study would show whether that ranking is general.","The reported small performance drops when the Hebbian initialization is applied to some consistency methods indicate that the pretraining is not universally beneficial, so mapping which methods and regimes tolerate it would sharpen the recipe."],"forward_implications":["If the central claim holds, an encoder-decoder segmentation model can be initialized from unlabeled images alone, so the number of pixel-level annotations needed for a usable biomedical segmentation system drops substantially.","The new TSA Hebbian rules mean the upsampling path is no longer a leftover that needs backpropagation; the entire architecture can be pretrained with local, biologically plausible updates.","Since the Hebbian initialization improves most pseudo-labeling and consistency-training baselines, the benefit compounds with existing semi-supervised losses rather than replacing them.","The same pipeline transfers from 2D histology, dermoscopy, and pupil images to 3D volumetric MRI, where it reaches best or second-best results at most label fractions.","The method's gains are concentrated in low-label regimes; the authors observe a smaller unsupervised-stage benefit on HMEPS at 1% labels, where the labeled set is already relatively large."],"supporting_citations":[{"why":"supplies the Hebbian SWTA/HPCA learning rules and the semi-supervised sample-efficiency setting that the first stage extends to segmentation.","marker":"[24]"},{"why":"the authors' earlier workshop paper that introduced preliminary Hebbian rules for transpose-convolutional layers and is extended here with the TSA rules and a broader evaluation.","marker":"[54]"},{"why":"defines the encoder-decoder biomedical segmentation architecture used as the shared backbone for all methods in the comparison.","marker":"[1]"},{"why":"provides the competing-hidden-units formulation from which the soft winner-take-all clustering rule is drawn.","marker":"[19]"},{"why":"derives the Hebbian/anti-Hebbian subspace learning result that underlies the HPCA principal-component rule.","marker":"[22]"},{"why":"supplies the SoftHebb Bayesian interpretation of soft winner-take-all updates used in the SWTA gate.","marker":"[20]"},{"why":"the cross-consistency training baseline the method compares against in the semi-supervised experiments.","marker":"[31]"},{"why":"the uncertainty-rectified pyramid-consistency baseline used in both 2D and 3D comparisons.","marker":"[33]"},{"why":"the variational-autoencoder two-stage baseline compared against in the low-label regimes.","marker":"[35]"},{"why":"the superpixel self-supervised two-stage baseline compared against in the low-label regimes.","marker":"[36]"}],"fun_headline_variants":["Hebbian pre-training slashes labels for medical segmentation","Bio-inspired unsupervised pretrain boosts segmentation","No-backprop pretraining for few-label medical imaging","Local learning rules teach segmenters with 1% labels","Transposed-Hebbian rules lift semi-supervised segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the new TSA rule for upsampling layers still discovers useful patterns even though the usual input and output roles are reversed, a property the paper proves only for the forward convolutional HPCA case and not for the TSA variant.","fun_headline_variants_meta":{"raw":{"variants":["Hebbian pre-training slashes labels for medical segmentation","Bio-inspired unsupervised pretrain boosts segmentation","No-backprop pretraining for few-label medical imaging","Local learning rules teach segmenters with 1% labels","Transposed-Hebbian rules lift semi-supervised segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000474,"raw_usage":{"total_tokens":2333,"prompt_tokens":907,"completion_tokens":1426,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":1361}},"tokens_in":523,"tokens_out":1426,"duration_ms":8371,"temperature":1.0,"reasoning_tokens":1361,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:40:37.257828+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fine-tune the full model from the SWTA-TSA initialization and from the same initialization with only the transpose-convolutional layers reset to random weights; if the Dice scores are statistically indistinguishable across several seeds, the new T-Conv rule is not carrying the claimed benefit.","supporting_citations":[{"cited_title":"Lagani, F","cited_arxiv_id":null,"evidence_quote":"supplies the Hebbian SWTA/HPCA learning rules and the semi-supervised sample-efficiency setting that the first stage extends to segmentation."},{"cited_title":"Ciampi, G","cited_arxiv_id":null,"evidence_quote":"the authors' earlier workshop paper that introduced preliminary Hebbian rules for transpose-convolutional layers and is extended here with the TSA rules and a broader evaluation."},{"cited_title":"Ronneberger, P","cited_arxiv_id":null,"evidence_quote":"defines the encoder-decoder biomedical segmentation architecture used as the shared backbone for all methods in the comparison."},{"cited_title":"Moraitis, D","cited_arxiv_id":null,"evidence_quote":"supplies the SoftHebb Bayesian interpretation of soft winner-take-all updates used in the SWTA gate."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"the variational-autoencoder two-stage baseline compared against in the low-label regimes."}],"review_version":1}