{"id":"a37b2a41-7ac0-4d3c-961a-10ed596f0906","arxiv_id":"2505.06886","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper introduces NeuRN, a local-contrast normalization layer, and reports mixed improvements in digit domain generalization and small, unquantified increases in similarity to mouse V1 representations.","lead":"This preprint claims that a normalization layer inspired by mouse visual cortex, called NeuRN, improves domain generalization in deep learning models and makes them more similar to mouse brain representations. The reported results are mixed across tasks, and the method closely resembles existing local contrast normalization.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No normalization baseline for NeuRN; Table 1 is mixed and lacks statistics, so the biological-DG link is untested.","rationale":"I read the paper in good faith. The strongest claim is that NeuRN improves both biological alignment and domain generalization. The reader's weakest_assumption focuses on the RMSE comparison requiring unjustified spatial correspondence between neural responses and feature maps. That is a valid concern, but I find a more load-bearing gap: the causal arrow from 'neural-inspired' to 'better DG' is untested because NeuRN is never compared against the generic normalization methods it cites, LCN and LRN. NeuRN is defined as patch-wise standard-deviation normalization, and the paper's own discussion identifies LCN and LRN as close relatives, yet no baseline experiment with them appears in Table 1 or Figure 2. Without that control, any observed improvement can be attributed to established normalization effects, not to the biological mechanism. Moreover, Table 1 is not uniformly positive: several large drops (e.g., DenseNet121 from 74.3 to 26.4 on M→U) contradict the 'significant improvement' narrative, and no error bars or significance tests support the comparisons. The undisclosed patch size k compounds the problem. These issues leave the central claim unsupported, so the reader's REJECT verdict remains appropriate. My concern differs from the reader's weakest_assumption, which identifies the RMSE spatial-alignment problem, hence agreement_with_reader is disagree.","tokens_in":13078,"tokens_out":5939,"duration_ms":60341,"concrete_test":"Retrain all ten models under the exact Table 1 protocol with NeuRN replaced by Jarrett et al. LCN (local mean subtraction and division by local std over the same k×k patches with k swept), using identical pretraining/fine-tuning, early stopping, and three seeds; report mean±std and paired per-task differences for the 12 M/S/U/MM transfers and the Fig. 2A/B RMSE curves. If LCN matches or beats NeuRN on average, NeuRN's neural-specific mechanism is not load-bearing; if NeuRN clearly beats LCN, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that NeuRN's biologically motivated normalization, rather than generic normalization, drives the reported gains. This is never tested: NeuRN is only compared with unmodified networks. NeuRN's computation, 1/(c·σ_pk) over k×k patches, is a local divisive normalization in the same family as LCN (Jarrett et al. 2009) and LRN (Krizhevsky et al. 2012), yet neither is run as a control. Without that control, any Table 1 differences could be the known robustness effect of local contrast normalization, making the 'neural response' explanation non-load-bearing. The table is also not consistently in NeuRN's favor (e.g., DenseNet121 drops from 74.3 to 26.4 on M→U; several S/U/MM targets show NeuRN below baseline), and no error bars or significance tests accompany the numbers. The layer's patch size k is not reported, so the experiment cannot be reproduced or compared fairly with LCN variants at matched capacity.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a framework for comparing neural representations from the mouse visual cortex (Allen Brain Observatory) with feature representations from deep networks, using RMSE and KDE-based measures. Based on an observed alignment, the authors introduce Neural Response Normalization (NeuRN), a biologically motivated local divisive normalization layer, and integrate it into several DNNs. They report that NeuRN increases biological alignment and improves domain generalization across digit datasets (MNIST, SVHN, USPS, MNIST-M), concluding that NeuRN captures domain-agnostic features. The paper also presents population-level and single-neuron-level comparisons across brain regions and genotypes.","tokens_in":13267,"tokens_out":10468,"duration_ms":90420,"significance":"If the claims were substantiated, the work would offer a simple, biologically inspired layer that improves robustness and strengthens the link between rodent visual cortex and DNN representations. The paper spans multiple architectures and brain regions, and it uses a public neuroscience dataset, which is a strength. However, the current evidence is not sufficient: the domain generalization table is mixed and lacks statistics, no generic normalization baselines are included, and the representational similarity metric relies on an unvalidated spatial correspondence. The absence of code and of the key hyperparameter (patch size k) further limits reproducibility.","major_comments":[{"comment":"The claim that NeuRN yields 'significant improvement' in domain generalization is not supported by the full table. For example, DenseNet121 drops from 74.3 to 26.4 on M→U, ResNet50v2 drops from 78.7 to 31.4 on M→U, and VGG19 drops from 66.2 to 48.4 on U→M, while improvements on other entries are large. No error bars or significance tests are reported for any entry, so the highlighted improvements cannot be distinguished from noise. The text selects four transfers (M→S, M→MM, U→S, MM→S) that improve, but this selection is post hoc. Please report mean and standard deviation over multiple seeds with paired significance tests and discuss the inconsistent pattern.","section":"Table 1"},{"comment":"NeuRN computes 1/(c·σ_pk) over k×k patches, which is a local divisive normalization in the same family as Local Contrast Normalization (Jarrett et al. 2009) and Local Response Normalization (Krizhevsky et al. 2012), both cited but never run as controls. Without a control experiment using LCN or LRN at matched capacity, any improvement in Table 1 could be attributed to generic normalization rather than to the biological inspiration. The biological-alignment results in Figures 2 and 3 need the same control before the claim that NeuRN specifically improves alignment can be accepted.","section":"Methods, Neural Response Normalization (NeuRN)"},{"comment":"The RMSE comparison reshapes a neural representation N = T^T·FS, whose axes are time and pixels, and a DNN feature map, whose axes are spatial positions, into equal-length 1D vectors and computes raw RMSE. This assumes a meaningful spatial correspondence between the two modalities without any justification, and it is not the standard RSA methodology, which compares similarity structures rather than raw vectors. The KDE/IoU analysis (Eqs. 2–5) compares only marginal activation distributions, discarding spatial layout. Since the reported RMSE differences in Figure 2 are small (on the order of 0.01) and no error bars are given, the biological-alignment evidence for NeuRN is not established.","section":"Methods, High-level representational analysis and Single neuron-level analysis"},{"comment":"The manuscript states that the domain generalization results 'validate our previous findings' (Iqbal et al. 2024). Because NeuRN is adapted from that prior work and the biological-alignment improvement is measured with the same RMSE/KDE framework used in the design of NeuRN, the new evidence provided here may be partly circular. Please clarify what independent evidence is added beyond the prior work, and specify whether the evaluation metric was used as an objective during the design of NeuRN.","section":"Results, NeuRN-derived performance boost for domain generalization in DNNs"}],"minor_comments":[{"comment":"The patch size k in the definition of NeuRN is never reported in the main text or the Appendix; without this value the experiments cannot be reproduced. The definition 'c = max(σ)' is ambiguous (max over which set of patches or channels) and no small constant is mentioned to avoid division by zero in flat regions.","section":"Methods, Neural Response Normalization (NeuRN)"},{"comment":"The reference for UMAP is given as 'Hastie and McMahan 2018', but UMAP is by McInnes et al. (2018); please correct the citation.","section":"References"},{"comment":"The text states that differences are 'insignificant' without reporting a statistical test; please add test results or remove the term.","section":"Results, NeuRN improves feature and neural representational alignment in DNNs"},{"comment":"The caption uses 'bold' for improvement and 'underlined' for close-to-benchmark scores, but the criteria for 'close-to-benchmark' are not defined, and the table would be clearer if it reported standard deviations.","section":"Table 1 caption"},{"comment":"The abstract and introduction claim 'significant improvements' from NeuRN while Table 1 contains many degradations; please align the wording with the evidence.","section":"Abstract and Introduction"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the central claim is not established. NeuRN is a local divisive normalization in the same family as LCN (Jarrett et al. 2009) and LRN (Krizhevsky et al. 2012), and the paper never runs either as a control. The DG table is mixed and lacks error bars. So the 'biologically motivated normalization' story is untested.\n\nCredit where due: the authors do a systematic sweep over 10 architectures and multiple cortical areas, they use the Allen Brain Observatory data, and the single-neuron KDE analysis is a nice addition. The paper is clearly written and easy to follow.\n\nThe main problem is the missing baseline. NeuRN's computation, 1/(c·σ) over k×k patches, is arguably LCN with no subtractive step. The paper cites LCN and LRN but only differentiates them verbally. Without running those as controls, any improvement in Table 1 could simply be the known robustness effect of local contrast normalization. The biological interpretation is non-load-bearing.\n\nSecond, Table 1 is not consistently in NeuRN's favor. VGG19 drops from 66.2 to 48.4 on U→M; DenseNet121 drops from 74.3 to 26.4 on M→U. The text highlights four favorable transfers and calls them 'significant,' but there are no error bars or significance tests. That is selective reporting.\n\nThird, the representational comparison via raw RMSE after reshaping time-by-pixels neural responses and conv-channel feature maps to equal-sized vectors assumes a spatial correspondence the paper never justifies. The reported RMSE differences are small and unquantified. If that assumption fails, the biological alignment evidence collapses.\n\nFourth, the patch size k is never reported, so the experiments cannot be reproduced or compared with LCN at matched capacity.\n\nFinally, the circularity concern is real but secondary: NeuRN is adapted from the authors' own prior work, and the alignment metric is the same RMSE/KDE framework used in the design. That does not make the result false, but it raises the burden for independent validation.\n\nWho this is for: someone working on biologically inspired normalization or DG might find the framework interesting, but they should not treat the results as evidence until the controls are run. The paper deserves a serious referee, but on the current evidence I would not accept it; I would ask for LCN/LRN baselines, error bars, the k value, and a justification—or replacement—of the RMSE alignment.","headline":"NeuRN is effectively a local contrast normalization variant, and without that baseline the biological-DG link is untested; Table 1 is mixed and lacks statistics.","tokens_in":13820,"tokens_out":3560,"would_cite":false,"duration_ms":34832,"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":"A brain-inspired normalization layer makes deep networks more robust to image-domain differences.","keywords":["domain generalization","mouse visual cortex","neural response normalization","representational similarity","biologically inspired neural networks","calcium imaging","cross-domain digit classification","contrast normalization"],"falsifier":"A permutation test would settle it: shuffle the spatial locations within the neural representation matrices, or within the feature maps, and recompute RMSE to the other system. If shuffled alignments achieve RMSE as low as the intact ones, the reported biological similarity is an artifact of overall scale rather than spatial structure. A complementary test would compare randomly initialized or random-label-trained networks; if they match mouse visual cortex representations equally well, the alignment is not due to learned features.","tokens_in":12826,"feed_emoji":"🧠","tokens_out":8344,"duration_ms":76317,"temperature":0.7,"pith_summary":"The paper sets out to prove that deep networks become more robust to image-domain shifts when their internal feature representations are pushed toward the response patterns of the mouse visual cortex. It builds a comparison framework that matches feature maps from pretrained networks against calcium-imaging responses of excitatory and inhibitory neurons, and reports that excitatory neurons align more closely with network features. It then adds NeuRN, a normalization layer that divides each pixel's value by the standard deviation of its surrounding patch, and claims that NeuRN increases both biological similarity and cross-domain digit classification accuracy. If right, the paper offers a simple, parameter-free layer that any classifier can adopt.","feed_headline":"Mouse-brain-inspired layer boosts AI's cross-domain accuracy","feed_subtitle":"Feature maps nudged toward mouse visual cortex responses generalize better across digit datasets.","key_machinery":"The load-bearing object is NeuRN, a normalization procedure that replaces each pixel's value by the inverted standard deviation of its local $k \\times k$ patch, scaled by a global maximum; it is inserted into existing deep networks without learned parameters. The supporting comparison machinery is the RMSE between neural representations $N = T^T \\cdot FS$ (trial traces transposed and multiplied by flattened stimuli) and reshaped feature maps, alongside kernel-density overlap measured by intersection-over-union. These comparisons are what the paper uses to argue that NeuRN increases biological alignment and that this alignment is the reason the models generalize.","core_discovery":"The paper's central claim is that the functional mapping of the mouse visual cortex closely tracks the feature representations of high-performing deep networks, especially for excitatory neurons, and that a deliberately inserted normalization layer, NeuRN, pushes artificial representations closer to biological ones while improving classification accuracy under domain shift. NeuRN computes, for each pixel, the standard deviation of its surrounding $k \\times k$ patch and divides by a global maximum standard deviation to produce a contrast-normalized feature map. The authors evaluate this with two measurements: the RMSE between flattened neural and feature representations, and the overlap of kernel-density estimates of activation distributions. On four digit datasets, models with NeuRN often beat their base versions when trained on one style of digits and tested on another.","pith_inferences":["A direct extension would be to test NeuRN on harder natural distribution shifts, such as corrupted versions of standard benchmarks, to see whether local-contrast normalization helps beyond digit fonts and backgrounds.","The success of NeuRN suggests a testable hypothesis for neuroscience: local contrast normalization may be a major reason cortical representations stay stable across illumination and background changes; this could be checked with recordings under controlled contrast manipulations.","Because the paper compares one temporally integrated neural response to static feature maps, a stronger version of its claim would compare layer responses time-locked to stimulus onset; the reported RMSE alignment does not yet control for latency."],"forward_implications":["Models with NeuRN should hold up better when test images differ in background, font, or lighting from training digits; the paper reports gains across twelve source-to-target transfer paths on four digit datasets.","NeuRN can be inserted into any CNN or Vision Transformer before standard training, since it requires no change to loss function or architecture.","The representational comparison pipeline can be applied to other neuron types, brain regions, or species to search for stronger candidate inductive biases.","If biological alignment is the mechanism, networks whose feature statistics match cortical responses more closely should continue to gain robustness as the alignment improves."],"supporting_citations":[{"why":"Supplies the in-vivo calcium-imaging neural responses to natural scenes that define the biological representations.","marker":"Allen Institute for Brain Science 2016"},{"why":"Source of the Winner-Takes-All and neuronal response-normalization idea the paper adapts into NeuRN.","marker":"Iqbal et al. 2024"},{"why":"Justifies using RMSE/Euclidean-style distance as a representational similarity metric.","marker":"Khaligh-Razavi and Kriegeskorte 2014"},{"why":"Provides the large-scale rodent visual cortex comparison against model zoology that the paper extends to single-neuron and NeuRN analyses.","marker":"Conwell et al. 2021"},{"why":"Provides MNIST, one of the digit datasets used in the domain generalization tests.","marker":"LeCun, Cortes, and Burges 2010"},{"why":"Provides SVHN, the real-world digit domain for transfer tests.","marker":"Netzer et al. 2011"},{"why":"Provides USPS, the scanned-mail digit domain for transfer tests.","marker":"Hull 1994"},{"why":"Provides MNIST-M and the domain-adversarial framing for colored-background transfer tests.","marker":"Ganin et al. 2015"},{"why":"Supplies the baseline Local Response Normalization to which NeuRN is compared, along with the reference deep architecture.","marker":"Krizhevsky, Sutskever, and Hinton 2012"}],"fun_headline_variants":["NeuRN: mouse-cortex layer for sharper domain generalization","Mouse cortex inspires robust AI layer","Brain-inspired normalization improves AI's cross-domain skill","From mice to machines: neural layer for robust AI","Cortex-coded layer helps AI adapt to new data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The biological-alignment evidence depends on the assumption that a neuron's response traces, reshaped into the same grid as a network's feature map, have a meaningful spatial correspondence with those feature maps; if the time and pixel axes do not correspond to the spatial layout of features, the RMSE numbers do not establish biological alignment.","fun_headline_variants_meta":{"raw":{"variants":["NeuRN: mouse-cortex layer for sharper domain generalization","Mouse cortex inspires robust AI layer","Brain-inspired normalization improves AI's cross-domain skill","From mice to machines: neural layer for robust AI","Cortex-coded layer helps AI adapt to new data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000861,"raw_usage":{"total_tokens":3744,"prompt_tokens":963,"completion_tokens":2781,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":579,"completion_tokens_details":{"reasoning_tokens":2708}},"tokens_in":579,"tokens_out":2781,"duration_ms":19479,"temperature":1.0,"reasoning_tokens":2708,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:31:03.504592+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A permutation test would settle it: shuffle the spatial locations within the neural representation matrices, or within the feature maps, and recompute RMSE to the other system. If shuffled alignments achieve RMSE as low as the intact ones, the reported biological similarity is an artifact of overall scale rather than spatial structure. A complementary test would compare randomly initialized or random-label-trained networks; if they match mouse visual cortex representations equally well, the alignment is not due to learned features.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the in-vivo calcium-imaging neural responses to natural scenes that define the biological representations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Justifies using RMSE/Euclidean-style distance as a representational similarity metric."}],"review_version":1}