{"id":"5dc0e20f-ff08-4e55-9ebe-35884f415771","arxiv_id":"2501.14854","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Pruned encoding models for visual cortex fMRI use as few as 1% of neural features per layer without losing accuracy, and different pretrained backbones predict the BOLD signal via distinct visual features.","lead":"Researchers trained fMRI encoding models that predict brain activity from images using pre-trained neural networks, then used pruning and AI explainability tools to find that a surprisingly small fraction of neural features is enough to match full-model accuracy. The study provides an open-source toolkit, BOLDreams, for building, pruning, and interpreting such encoding models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The pruning-invariance claim is established only for the standard-deviation filter ranking; without a random-pruning control, the 'maximally minimal' conclusion is not isolated from the selection heuristic.","rationale":"The reader's weakest_assumption identifies the pruning rule, and I agree that this is the most load-bearing point. The central claim has two components: numerical invariance to readout pruning and the existence of a maximally minimal model M0. Both depend on how filters are selected when the readout is narrowed. If the standard-deviation ranking is not compared against alternative selection rules, then the '1–10%' figure is an artifact of the heuristic rather than a property of the encoding model or the BOLD signal. This is not a fatal flaw—the paper's qualitative observation may survive a control—but it is precisely the kind of assumption that should be explicitly tested before the abstract's 'maximally minimal' language is accepted. The concrete test I propose would settle the issue directly by re-running a small subset of the experiments with different selection rules. The other concerns (subject 7 exclusion, test-set model selection, RN50x4 finetuning failure) are real but secondary; they affect generalizability and incremental claims, whereas the pruning rule is load-bearing for the main quantitative sweep. I therefore keep the reader's CONDITIONAL verdict unchanged.","tokens_in":14378,"tokens_out":4367,"duration_ms":41472,"concrete_test":"Run the AlexNet (no-finetune and finetune) pipelines at p = 1, 5, and 10% with three alternative filter-selection rules: (i) random selection without replacement, (ii) selection by the L2 norm of readout weights learned with the full 100% readout, and (iii) selection by mutual information between filter activations and the mean voxel response across training images. Compare the resulting mean test correlations (averaged over subjects 1, 2, and 5) against the standard-deviation-ranked curves in Figure 2. If all three procedures fall within the existing standard error bars at each p, then the invariance is not an artifact of the ranking rule. If random selection degrades by more than the standard error, the paper must restate the claim as 'pruning to 1–10% of high-variance filters' and cannot support M0 as a global minimal model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2 defines the pruning procedure: filters are ranked by the standard deviation of their activations over the training set, and the top p% are kept. The central quantitative claim—that readout width can be reduced to 1–10% without loss of accuracy—is established only for this specific ranking. The paper provides no control condition, such as random filter selection, selection by learned readout-weight magnitude, or selection by voxel-response mutual information. If high-variance filters are a particularly informative subset, then the flat accuracy curves could reflect the quality of that subset rather than genuine redundancy in the neural code; if random subsets at the same p% perform equally well, the claim is robust but the 'maximally minimal' M0 is still minimal only under a heuristic that was never varied. The post-hoc exclusion of subject 7 and the test-set selection of the best model are additional caveats, but the pruning rule is the load-bearing assumption because it defines what 'pruning' means in every figure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper trains feature-weighted receptive field encoding models on the Natural Scenes Dataset using three pretrained backbones (AlexNet, VGG11, CLIP RN50x4), with and without finetuning, and with readouts pruned by retaining the top p% of filters per layer ranked by activation standard deviation. Voxel-wise correlations are reported as a function of this pruning percentage. The main empirical claim is that reducing the readout to 1–10% of filters per layer has little effect on prediction accuracy, and that a finetuned AlexNet with 10% of filters per layer is the best-performing model, motivating a proposed 'maximally minimal' model M0. The paper also presents 'dreams' (maximally exciting images) and integrated-gradient attention maps as interpretability analyses.","tokens_in":14598,"tokens_out":3554,"duration_ms":35372,"significance":"If the pruning-invariance result is robust, it is an interesting and practically useful finding: it suggests that large pretrained feature spaces are substantially redundant for linear readout to BOLD responses and that much smaller readouts suffice. The released BOLDreams toolkit and the systematic comparison across backbones, finetuning, and readout width are useful contributions. The interpretability analyses (dreams, attention maps, word clouds) provide qualitative illustrations that may generate hypotheses. However, the central pruning claim is currently supported only under a single filter-ranking heuristic, without a control condition, and the statistical and data-handling caveats (exclusion of subject 7, failed RN50x4 finetuning, test-set model selection) need to be addressed before the strength of the claim is established.","major_comments":[{"comment":"The pruning-invariance claim is established only for the standard-deviation filter ranking defined in §2.2; there is no control condition using random filter selection, readout-weight magnitude, or mutual-information-based selection at matched pruning levels. Without such controls, the flat accuracy curves could reflect the informativeness of high-variance filters rather than genuine redundancy in the neural code, and the 'maximally minimal' model M0 is defined relative to a heuristic that is never varied. Please add a random-pruning control and ideally at least one alternative ranking, reporting the variance of performance across randomly sampled filter subsets.","section":"§2.2 and Figure 2"},{"comment":"Subject 7 is excluded from the main accuracy and histogram analyses because it 'appears to be an outlier,' but no quantitative criterion or preregistered rule for this exclusion is given. Since all central claims are averaged over subjects, post-hoc exclusion can change the conclusions. Please report results with and without subject 7, justify the exclusion statistically, or treat the exclusion explicitly as a sensitivity analysis.","section":"§3.1, Figures 2–4"},{"comment":"The fine-tuned RN50x4 condition is reported as producing zero or NaN correlations and is omitted from all subsequent plots, but no explanation or diagnostic is provided. This leaves the backbone-by-finetuning comparison incomplete: the claim that the finetuned AlexNet is the best-performing model is made without a successful finetuned CLIP comparison. Please document the failure mode (for example, loss divergence or learning-rate sensitivity) or provide a corrected finetuning protocol.","section":"§3.1, Figure 2"},{"comment":"The 'best percentage parameter' for each model appears to be selected using the same test set on which the final comparisons are made. This is test-set model selection and can optimistically bias the reported differences between models. Please clarify whether a separate validation split or nested cross-validation was used for selecting the percentage parameter, or report the selection results on a validation subset.","section":"§3.1, Figure 4"},{"comment":"The paper repeatedly states that pruning 'does not significantly reduce' correlation and that differences are 'marginal,' but no statistical test is reported that compares ρ_i with ρ_100 across voxels or subjects. Error bars showing standard error are not a substitute. Please provide paired tests or confidence intervals for the pruning effect, and state whether the invariance claim holds statistically for each backbone and finetuning condition.","section":"§3.1 and Discussion"}],"minor_comments":[{"comment":"There are numerous typographical and naming inconsistencies, for example 'LeCunn' in the introduction, 'Vgg' instead of 'VGG' in several places, and inconsistent hyphenation of 'finetuned' versus 'fine tuned.' These should be corrected.","section":"Throughout"},{"comment":"The notation for layers is inconsistent: the text uses P_l for pixel spaces and then uses script ℓ both for the readout layers and for the pixel-space dimension in Eq. (1). Please unify the notation to avoid ambiguity.","section":"Equation (1)"},{"comment":"The interpretation of the dreams and word clouds is entirely qualitative and subjective. I do not object to qualitative XAI analysis, but the claims that specific backbones produce 'biologically plausible' or 'relevant' features would be strengthened by a quantitative evaluation, such as a rating study or a similarity metric between dream features and known category-selective responses.","section":"§3.2, Figures 5–7"},{"comment":"The caption states that intensity to the left of the red line means the 100% model performs better. Since the y-axis is the maximum of ρ_i and ρ_100, the relationship between a point's location and the sign of ρ_i − ρ_100 is not immediately clear; a more direct scatter plot of ρ_i versus ρ_100 with a diagonal reference line would be easier to interpret.","section":"Figure 3"},{"comment":"The decomposition M = M0 ⊕ M1 is introduced in prose without a formal definition of the fusion operation or a criterion for when a connection belongs to M1. As a conceptual proposal this is acceptable, but the paper should state explicitly that this is a qualitative framework rather than a construction with algorithmic content.","section":"§4, Discussion of M0"}],"recommendation":"major_revision","confidential_remarks":"The core empirical observation—flat accuracy across readout width—is potentially interesting and the toolkit is useful, but the manuscript currently lacks a control for the pruning heuristic and contains several unresolved data-handling issues (subject 7 exclusion, NaN finetuning, test-set model selection). These issues are fixable within the scope of the paper, so I recommend major revision rather than rejection. Given the qualitative nature of the interpretability sections, the strongest path to acceptance would be to make the pruning-control experiment a central, quantitative part of the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read BOLDreams. Short version: this is a solid, incremental extension of the feature-weighted receptive field program. The genuinely new parts are the systematic pruning comparison across AlexNet, VGG11, and CLIP-RN50x4 on NSD, and the XAI visualizations (dreams, integrated gradients, word clouds) applied to pruned models. The central quantitative claim—readout width can be cut to 1–10% of filters per layer with little change in voxel-wise correlation—holds up in the figures, and the accompanying open-source toolkit is a real asset.\n\nWhat the paper does well: clean use of a rigorous public dataset, voxel-level distribution plots rather than only means, a fair comparison of text vs. image encoders, and an honest discussion that treats the 'maximally minimal' model as a framing rather than a discovery.\n\nSoft spots, in order of weight:\n\n1. The pruning result is only demonstrated for the standard-deviation filter ranking. There is no control with random pruning or an alternative selection criterion. Without that, the flat accuracy curves could reflect the quality of the heuristic, not redundancy in the neural code. This is the main weakness. It does not sink the claim, but it limits the 'maximally minimal' interpretation.\n\n2. Subject 7 is excluded post hoc as an outlier, with no quantitative criterion or sensitivity analysis. That matters because the paper's headline numbers are all subject-7-free.\n\n3. The best percentage per model is chosen on the test set. It is standard model selection, but with nine discrete settings and a single test set, there is some selection bias. Not fatal, but worth noting.\n\n4. The RN50x4 finetuning failed (NaN correlations) and is silently dropped. That leaves the 'finetuning' column incomplete for the largest backbone.\n\n5. The dream and word-cloud interpretation is explicitly subjective. That is acceptable for a qualitative section, but the abstract's 'biologically plausible features' is stronger than the evidence supports.\n\nThe math is simple and standard; no errors spotted. The citation pattern is fair, with appropriate references to prior encoding and XAI work. The self-citation is minor and justified.\n\nWho is this for? Researchers working on fMRI encoding models and anyone who wants a ready-made toolkit for pruning and XAI in this setting. It is not a breakthrough, but it is a useful data point.\n\nRecommendation: worth a serious referee round. I would ask for a random-pruning control, a proper handling of subject 7, and an explanation of the RN50x4 finetuning failure. With those, the pruning invariance result would be solid.","headline":"Pruning invariance is real but under-specified: the result needs a random-pruning control and better handling of subject exclusion, but it is a useful incremental contribution.","tokens_in":15063,"tokens_out":4364,"would_cite":true,"duration_ms":36788,"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":"Across three CNN backbones, cutting the readout to 1-10% of filters per layer leaves fMRI prediction accuracy essentially unchanged.","keywords":["fMRI encoding models","visual cortex","BOLD signal","readout pruning","feature-weighted receptive field","maximally minimal model","dreaming","Natural Scenes Dataset"],"falsifier":"Train the same readout at 1% and 10% widths with filters chosen at random instead of by highest activation variance, and compare per-voxel correlations with the variance-ranked versions; if random selection reproduces the full-model accuracy, the invariance is a property of readout redundancy rather than of the BOLD signal's feature content.","tokens_in":14182,"feed_emoji":"🧠","tokens_out":8570,"duration_ms":71408,"temperature":0.7,"pith_summary":"Using the Natural Scenes Dataset, the authors train feature-weighted receptive field encoding models that map CNN features from three backbones to visual-cortex voxel responses. They report that pruning the readout to as little as 1-10% of the filters per layer leaves per-voxel prediction accuracy essentially unchanged, with a fine-tuned AlexNet at 10% giving the highest correlations. They interpret this as evidence that the BOLD signal encodes a coarse scaffold of visual features, and they use the finding to propose a maximally minimal model, M0, as the smallest readout that still encodes the signal. The accompanying dream and attribution analyses show that different backbones reach similar accuracy through different predictive mechanisms.","feed_headline":"Pruning 90% of CNN filters barely hurts brain-response predictions","feed_subtitle":"A fine-tuned AlexNet with 10% of filters per layer matches full-width models on voxel-wise correlation.","key_machinery":"The central object is the feature-weighted receptive field readout of Equation 1: for each voxel, trainable receptive fields pool feature maps at selected layers, and a weighted sum maps the pooled values to a predicted BOLD response. The pruning mechanism ranks the filters in each layer by the standard deviation of their activations over the training set and keeps only the top p% for the readout. This pairing of a linear readout over a fixed CNN feature hierarchy with variance-based filter selection is what carries the claim that readout width has little effect on accuracy.","core_discovery":"Across all three backbone architectures, voxel-wise correlation between predicted and measured BOLD responses stays essentially flat as the readout is pruned from 100% of the filters per layer down to 1%, and the best overall accuracy is reached by a fine-tuned AlexNet using only 10% of filters per layer. Because the pruning rule keeps the highest-variance filters, the authors conclude that the BOLD signal is well captured by a small set of broadly informative features and does not require the detailed feature repertoire of a large pretrained network. They cast this as a maximally minimal model, M0, with the full model decomposed as M = M0 ⊕ M1, where M1 comprises connections that are compatible with the signal but need external evidence to be justified.","pith_inferences":["Inference: If the variance-ranking result generalizes, a similar pruning sweep on other fMRI encoding datasets and cortical areas should reproduce the flat accuracy curve; that would make aggressive readout pruning a cheap preprocessing step for large-scale encoding studies.","Inference: The paper's dream images suggest a direct behavioral test: present maximally exciting images from pruned and full models to a new sample of participants; if pruned-model dreams drive BOLD responses as strongly as full-model dreams, the omitted filters are confirmed irrelevant to that signal.","Inference: The M0 ⊕ M1 decomposition predicts representational degeneracy at the voxel level; two models with disjoint sets of kept filters could still predict the same voxel responses, which could be probed by comparing prediction consistency across random filter subsets.","Inference: The pruning-by-variance recipe could be carried over to decoding pipelines, where shrinking the feature space before image reconstruction from fMRI might lower computational cost without hurting reconstruction quality."],"forward_implications":["Pruned models with 1-10% of filters per layer can be used in place of full-width readouts for fMRI prediction with nearly identical voxel-wise accuracy, reducing storage and computation substantially.","The existence of M0 implies that, at least for the NSD paradigm, the BOLD signal in visual cortex does not require the fine-grained features of large pretrained networks; broad scaffolding features suffice.","Because different backbones reach similar accuracy through different learned features, model accuracy alone cannot identify which features the brain actually uses; mechanism-level analyses such as dreams and attention maps are necessary.","The M0 ⊕ M1 decomposition suggests that additional data modalities, such as electrophysiology or behavior, are needed to constrain the features that are compatible with but not required by the BOLD signal."],"supporting_citations":[{"why":"provides the NSD fMRI dataset and image stimulus set used to train and test all models.","marker":"[2]"},{"why":"introduces the feature-weighted receptive field readout architecture that maps CNN features to voxels.","marker":"[35]"},{"why":"AlexNet is one of the three pretrained backbones and produces the best-performing pruned model when fine-tuned.","marker":"[19]"},{"why":"VGG11 is the second pretrained backbone used for comparison.","marker":"[27]"},{"why":"CLIP RN50x4 is the third backbone and supplies the text encoder for the language-based model.","marker":"[23]"}],"fun_headline_variants":["Pruning 90% of CNN filters barely hurts brain predictions","Tiny AlexNet with 10% filters matches full model on brain activity","Filter pruning to 10% keeps visual cortex predictions accurate","Minimal model achieves best BOLD prediction with 10% filters","Heavy pruning of CNN filters has minimal impact on BOLD forecasts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pruning invariance rests on the assumption that the filters with the largest activation variance across training images are the ones carrying the BOLD-relevant information; if that ranking rule is not the right one, the flat accuracy curve could be an artifact of the pruning procedure rather than a sign that the BOLD signal encodes only a coarse feature scaffold.","fun_headline_variants_meta":{"raw":{"variants":["Pruning 90% of CNN filters barely hurts brain predictions","Tiny AlexNet with 10% filters matches full model on brain activity","Filter pruning to 10% keeps visual cortex predictions accurate","Minimal model achieves best BOLD prediction with 10% filters","Heavy pruning of CNN filters has minimal impact on BOLD forecasts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000433,"raw_usage":{"total_tokens":2169,"prompt_tokens":870,"completion_tokens":1299,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":1209}},"tokens_in":486,"tokens_out":1299,"duration_ms":24317,"temperature":1.0,"reasoning_tokens":1209,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:56:43.138395+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same readout at 1% and 10% widths with filters chosen at random instead of by highest activation variance, and compare per-voxel correlations with the variance-ranked versions; if random selection reproduces the full-model accuracy, the invariance is a property of readout redundancy rather than of the BOLD signal's feature content.","supporting_citations":[{"cited_title":"The feature-weighted receptive field: an interpretable encoding model for complex feature spaces","cited_arxiv_id":null,"evidence_quote":"introduces the feature-weighted receptive field readout architecture that maps CNN features to voxels."},{"cited_title":"ImageNet Classification with Deep Convolutional Neural Networks","cited_arxiv_id":null,"evidence_quote":"AlexNet is one of the three pretrained backbones and produces the best-performing pruned model when fine-tuned."}],"review_version":1}