betaFeedback

WorksAgricultural and Biological Sciences

Analysis choices decide whether untrained networks beat trained ones at V1: a multiverse analysis of model-brain RSA

Abstract

Untrained convolutional networks have been reported to rival trained networks as models of visual cortex, and we had reported an untrained-network advantage over backpropagation at V1 in THINGS fMRI, together with a dependence on evaluation resolution. Under a pre-specified ridge readout, the untrained network scored below the trained one at every pooling grid (median r difference -0.043 at g = 4, 0 of 5 seeds positive), and also in MSE skill and per-fold r. In exploratory analyses, across 36 combinations of unit normalisation, distance measure and voxel normalisation (28 distinct analyses), the sign of the V1 contrast depended on the analysis (intervals above 0 in 5 and below 0 in 9 distinct analyses), and in our networks the resolution dependence of the original convention appeared in only one cell. Raw correlation-distance representational similarity analysis (RSA) covaried with the alignment of model and voxel patterns with the unit-mean profile (Spearman 0.97 across our networks, 0.79 across six ImageNet architectures); the advantage was absent from the component orthogonal to the mean profile, and the parallel component alone did not resolve it. On the unweighted angular component of centred patterns, trained networks were ahead at V1 (our network and 5 of 6 ImageNet models) and at LOC (all seven); the norm-weighted angular term was not resolved. A luminance-and-contrast bank scored above every network under ridge (point estimates), and adding network features under a shared penalty did not improve prediction. Intervals are stimulus-bootstrap intervals for three subjects, and 6 pre-stated predictions were not met. We recommend reporting sensitivity to centring over the evaluated stimuli on both sides, reporting components separately, and including a small multiverse and simple baselines.

Introduction

Comparisons between artificial networks and neural data have become a standard way to ask which models capture the computations of visual cortex. Representational similarity analysis (RSA) is one of the most widely used tools for this purpose: it summarises each system by the dissimilarities between its responses to a set of stimuli and compares these summaries directly, without fitting a mapping [Kriegeskorte et al., 2008]. A recurring and surprising result is that networks with random, untrained weights can rival trained networks as models of visual cortex, in some reports exceeding them [9] (Kazemian et al. 2025). If true, such results bear on how much of early visual representation follows from architecture alone and how much from learning.

We reported a result of this kind ourselves. In THINGS fMRI [Hebart et al., 2023], an untrained network with uncalibrated batch-normalisation statistics exceeded a backpropagation-trained network at V1 when images were evaluated at 224 px, but not at 32 px, the resolution at which the networks had been trained [Leutenegger, 2026a]; a related claim appears in Leutenegger [2026b]. Both analyses used a single convention: globally pooled features, correlation distance between units, and voxel responses normalised over all trials of the experiment.

That convention has a property that is easy to overlook. Correlation distance across units is not invariant to constant offsets of individual units over the stimulus set: when units have different mean responses, the dissimilarity between two stimuli depends in part on how strongly each stimulus pattern is aligned with the shared profile of unit means. Euclidean distance is invariant to such offsets but not to the norm of each stimulus’s response vector. Different distance measures and normalisations therefore weight a gain-like quantity and the direction of the response pattern differently. The role of centring is explicit in related similarity measures [Kornblith et al., 2019], and the choice of dissimilarity measure is known to affect the reliability of RSA [10] (Diedrichsen and Kriegeskorte 2017), but its effect on the ranking of candidate models has rarely been examined.

Here we test our earlier result in four ways. First, a pre-specified analysis replaces RSA by a fitted ridge readout on the full feature map. Second, a sequence of exploratory analyses, each with predictions and decision rules committed before it was run, asks where the discrepancy between RSA and ridge arises; among them, we test whether the advantage is carried by the component along the mean profile of the units. Third, a multiverse of 36 combinations of unit normalisation, distance measure and voxel normalisation [8] tests which conclusions hold across defensible analysis choices, and six ImageNet architectures test whether they generalise beyond our networks. Fourth, an exact decomposition of Euclidean distances separates a norm component from an angular component.

We find that whether the untrained network beats the trained one at V1 depends on the analysis. The advantage is not reproduced by the fitted readout, disappears under correlation distance when model units are standardised, and is absent from the angular component and from the component orthogonal to the mean profile, whereas raw correlation-distance RSA closely tracks the mean-profile alignment of model and voxels. The ranking at LOC favours the trained networks wherever it is resolved. A luminance-and-contrast bank scored above every network under the ridge readout at V1 (point estimates). We report all pre-stated predictions, including those that failed, and we revise the corresponding claims of our earlier work.

Methods

Data

We used the THINGS-fMRI data set [Hebart et al., 2023] of 3 subjects. Single-trial betas were taken from the ROI masks of the voxel metadata (V1; LOC as the union of lLOC and rLOC) and z-scored per voxel across all 9,840 trials of the subject before stimulus selection. The 720 evaluation stimuli are one exemplar each of 720 distinct concepts, in the same order for all subjects; all are single-presentation train trials from one session per subject.

Voxel reliability was measured on the 100 test images, shown 12 times each, identical across subjects and disjoint from the 720. Per voxel, the Pearson correlation between the means of two halves of 6 repetitions was averaged over all 462 balanced splits and Spearman–Brown corrected, rSB=2r/(1+r)r_{\mathrm{SB}} = 2r/(1+r). A voxel is reliable if rSB>0.2r_{\mathrm{SB}} > 0.2, a threshold fixed in advance for all conditions; measured on other images, selection cannot leak into fits on the 720. V1 contained 501, 518 and 286 reliable voxels in sub-01, sub-02 and sub-03. Ridge scores use reliable V1 voxels; RSA uses all voxels of the ROI.

RDMs were compared on cross-run pairs only: pairs of distinct stimuli shown in different runs in every subject, 210,205 of the 258,840 pairs of the full set. In every subject the 720 fall into 10 runs of 72; sub-01 and sub-02 saw them in identical run order, so same-run pairs share a run component, and the restriction removes it [Leutenegger, 2026a].

Models

CIFAR-trained CNNs. We evaluated the 30 checkpoints of Leutenegger [2026a]: 6×56 \times5 seeds (42, 123, 456, 789, 1337) of one convolutional architecture, trained on CIFAR-10 [Krizhevsky, 2009] at 32 px, the training resolution. The conditions are Backprop, Feedback Alignment (FA) [Lillicrap et al., 2016], Predictive Coding (PC) [7, 11], STDP [1, 5] and two untrained networks: untrained (uncalibrated BN statistics), with BatchNorm running statistics left at initialisation, and untrained (BN-calibrated), with running statistics set by 50 forward-only passes of CIFAR batches at 32 px. “Untrained” alone means the uncalibrated variant, as in the published headline.

The V1 layer is Conv1, the output of the whole first block (convolution, BatchNorm, ReLU, 2×22 \times2 max pooling; 32 channels, map side half the input size); LOC is mapped to Conv3. Features were extracted in eval mode with the original code, in two input arms. NATIVE: the image is resized and centre-cropped to p∈{32,64,96,128,160,224}p \in\{32, 64, 96, 128, 160, 224\} px. UPSAMPLED: the image is resized and cropped to 32 px and then resized to pp, so its content is limited to 32 px; we used it at p=224p = 224. Both arms apply the CIFAR channel normalisation.

ImageNet models. Six torchvision models [6, 4]: alexnet, vgg11_bn, resnet18, resnet50, convnext_tiny and vit_b_16, in eval mode. Trained models use the IMAGENET1K_V1 weights. Untrained models were built without weights after torch.manual_seed(s), s∈{0,1,2}s \in\{0, 1, 2\}, with BatchNorm running statistics at their defaults (mean 0, variance 1); convnext_tiny and vit_b_16 use LayerNorm. RGB images were resized (shorter side) and centre-cropped to R∈{112,224,448}R \in\{112, 224, 448\} (training resolution 224) and ImageNet-normalised; vit_b_16 accepts only 224. Layers (V1 / LOC), fixed before the run: alexnet features[1] / features[12]; vgg11_bn features[2] / last max pooling of features; resnet18 and resnet50 maxpool / layer4; convnext_tiny features[0] / features[7]; vit_b_16 encoder.layers[0] / encoder.layers[-1]. Outputs were averaged over space, or over tokens without the class token.

RSA convention

Unless stated otherwise, RSA follows Leutenegger [2026a]. The model RDM is the correlation distance (1−1 - Pearson rr across units) between the globally average-pooled (GAP) features of two stimuli; the fMRI RDM is the correlation distance between their beta patterns over all ROI voxels. RSA is the Spearman correlation between the two RDMs over the cross-run pairs, per subject, then averaged. A contrast is the mean over seeds of the per-seed difference between two conditions; for ImageNet models it is untrained (mean over the 3 initialisation seeds) minus trained.

Uncertainty is over stimuli: the 1000 bootstrap resamples of Leutenegger [2026a] (720 stimuli drawn with replacement, seed 20261002). Each draw recomputes the cross-run pair set on the resampled stimuli, dropping pairs of a stimulus with itself, and every difference, slope and ratio is formed within the draw (paired). Intervals are 95% percentile intervals. They describe stimulus variation in 3 subjects, not the population. Unit means and SDs, the feature transformations FF, the voxel normalisations BB, the mean profile and the norms are estimated once on the 720 stimuli and held fixed across resamples; the bootstrap resamples pairs of fixed RDMs. Re-estimating them within 200 resamples left the classification of three cells unchanged (Appendix § 12).

Analysis multiverse

Three factors were crossed. Feature normalisation FF, per unit over the 720 stimuli: raw; center (mean subtracted); scale (divided by the SD, not centered); zz (both). Units with SD below 10−810^{-8} are set to 0. Distance DD: correlation, cosine or euclidean, applied to features and betas alike. fMRI normalisation BB: paper (betas as in § 2.1), center720 (centred per voxel over the 720 stimuli) and z720 (z-scored per voxel over the 720). The cell (raw, correlation, paper) is the convention of Leutenegger [2026a]. All 4×3×3=364 \times3 \times3 = 36 cells were computed for the V1 contrast (Conv1, untrained − Backprop) and the LOC contrast (Conv3, Backprop − untrained). The resolution slope is S=Gap⁡(NATIVE 224)−Gap⁡(NATIVE 32)S = \operatorname{Gap}(\mathrm{NATIVE}\ 224) - \operatorname{Gap}(\mathrm{NATIVE}\ 32), with Gap the V1 contrast. As a check, euclidean RDMs under F=F = raw and center must coincide. For ImageNet models, all 36 cells were computed as point estimates and four were bootstrapped: (raw, correlation, paper), (z, correlation, paper), (raw, euclidean, paper) and (center, correlation, center720); their slope is Sgen=Gap⁡(448)−Gap⁡(224)S_{\mathrm{gen}} = \operatorname{Gap}(448) - \operatorname{Gap}(224).

Ridge readout

Encoding models used dual ridge regression [Hoerl and Kennard, 1970], with the kernel K=XXTK = XX^{\mathsf{T}} eigendecomposed once per split. The Conv1 maps were pooled to a g×gg \times g grid by adaptive average pooling, g∈{1,2,4,8}g \in\{1, 2, 4, 8\}, giving 32g232g^{2} features independent of resolution; g=1g = 1 is GAP, and larger gg retains spatial structure. Features were standardised with statistics of the training rows of each split only.

Cross-validation is nested: 8 outer folds and 5 inner folds within each outer training set, from a fixed partition (seed 20260818); with one exemplar per concept, folds are concept-grouped. Per voxel and outer fold, α\alpha was chosen from 29 log-spaced values between 10−610^{-6} and 10810^{8} by the minimum mean squared error over the inner folds, and the model was refit on the outer training set. The regularisation is selected by mean squared error, not by Pearson rr, the reported score. In simulations with the dimensions of the primary data set, selection by rr placed a substantial share of the selections at reliable voxels on the lowest grid value in every simulated population that contained reliable voxels, irrespective of the strength of the signal, whereas selection by mean squared error placed none there (Phase 2b report in the repository); because rr does not depend on the scale of the predictions, which α\alpha mainly sets, its inner-fold optimum was not localised on the grid. The selected α\alpha therefore need not maximise rr. Out-of-fold predictions were pooled over the outer folds and scored by one Pearson rr per voxel over all 720 stimuli. When the selected regularisation shrinks the predictions of a voxel towards the training mean, this pooled correlation is negative in expectation, so features without signal yield negative rather than zero scores. A model that predicts the mean of the training rows of each fold scored −0.092 (mean over subjects); the 97.5th percentile of a permutation null for the seed-0 networks lay between −0.062 and −0.037, and observed scores were above it in sub-01 and sub-02 (Appendix F). The score is the median rr over reliable V1 voxels, averaged over subjects, with contrasts formed within seed.

A low-edge selection at a reliable voxel is logged, not fatal (amendment of 2026-10-03). With diagnostic values 10−710^{-7}, 10−810^{-8} and 10−910^{-9} added, it is a valid plateau if the inner-fold MSE improves by less than a relative 10−310^{-3} on its value at 10−610^{-6}; otherwise the cell is excluded. The gate never changes a score.

The pre-specified analysis used the 720 single-trial betas at NATIVE 224 px, all checkpoints and subjects, and all four gg; the g=2g = 2 features were extracted on a CPU, subject to a downstream equivalence gate (amendment of 2026-10-13). An exploratory fit at g=1g = 1 covered all NATIVE resolutions and UPSAMPLED 224.

Two controls used the identical readout. The control band consists of low-level banks pooled to the same grid: pixels; the luminance-and-contrast bank, two feature maps (luminance, Rec. 709, Y=0.2126R+0.7152G+0.0722BY = 0.2126R + 0.7152G + 0.0722B, and local RMS contrast in a 9×99 \times9 window) pooled to the g×gg \times g grid, i.e. 2g22g^2 features (2 at g=1g = 1, 128 at g=8g = 8); absolute responses of a Gabor bank (4 orientations ×\times 3 spatial frequencies); and, as negative test, Backprop seed-0 features with permuted stimulus assignment. The spatial-correspondence control permutes the g×gg \times g cells differently for every image (seed 20260820), keeping dimensionality and per-channel distributions; it was run for the seed-0 untrained and Backprop checkpoints at g∈{2,4,8}g \in\{2, 4, 8\}.

Mean-profile components and decomposition

For a stimulus ×\times unit matrix XX (720 rows), let μ\mu be the unit means over the 720 stimuli, μc=μ−μˉ\mu_c = \mu- \bar{\mu} the mean profile with its mean over units removed, u=μc/∥μc∥u = \mu_c/\lVert\mu_c\rVert, and x~i\tilde{x}_i row ii centred across units. The mean-profile alignment of stimulus ii is the cosine between its centred pattern and the mean profile,

ai=⟨x~i,u⟩∥x~i∥.(1)a_i = \frac{\langle\tilde{x}_i, u\rangle}{\lVert\tilde{x}_i\rVert}. \tag*{(1)}

Model alignment was computed on raw features, voxel alignment on all ROI voxels under B=paperB = \text{paper}. calignc_{\mathrm{align}} is the Pearson correlation over the 720 stimuli between model and V1 voxel alignment, averaged over subjects. The alignment RDM is ∣ai−aj∣\lvert a_i - a_j\rvert. Partial RSA regresses the ranks of the model and of the fMRI RDM, per subject and draw, on an intercept and the ranks of their own alignment RDM and correlates the residuals (Pearson). Alignments were described against stimulus mean luminance and RMS contrast, the per-image mean and SD of Rec. 601 luminance (Y=0.299R+0.587G+0.114BY = 0.299R + 0.587G + 0.114B); over the 720 stimuli at 224 px, the image means of the two luminance definitions correlate at Pearson 0.997 (Spearman 0.996).

The projection coefficient pi=⟨x~i,u⟩=ai∥x~i∥p_i = \langle\tilde{x}_i, u\rangle= a_i\lVert\tilde{x}_i\rVert and the remainder ri=x~i−piur_i = \tilde{x}_i - p_i u split each centred pattern into a part along the mean profile and a part orthogonal to it. We compare three variants: “full”, the correlation distance on xix_i (the convention of Leutenegger [2026a]); “orthogonal”, the correlation distance on rir_i; and “parallel”, the RDM ∣pi−pj∣\lvert p_i - p_j\rvert. The beta patterns are decomposed in the same way (voxel means per the 720 stimuli, B=paperB = \text{paper}), and each model component is compared with the same component of the brain.

The norm and angular components start from XcX_c, XX with every unit centred over the 720 stimuli (F=centerF = \text{center}; for F=zF = z, units are also scaled to SD 1). With ni=∥Xc,i∥n_i = \lVert X_{c,i}\rVert and cos⁡θij\cos\theta_{ij} the cosine between rows ii and jj,

∥Xc,i−Xc,j∥2=(ni−nj)2+Wij,Wij=2ninj(1−cos⁡θij).(2)\lVert X_{c,i} - X_{c,j}\rVert^2 = (n_i - n_j)^2 + W_{ij}, \qquad W_{ij} = 2n_i n_j(1 - \cos\theta_{ij}). \tag*{(2)}

We report the norm RDM Nij=∣ni−nj∣N_{ij} = \lvert n_i - n_j\rvert, the angular component Aij=1−cos⁡θijA_{ij} = 1 - \cos\theta_{ij}, the unweighted angular dissimilarity of the patterns centred over stimuli, and the norm-weighted angular term WijW_{ij}. The identity holds for WW, not for AA alone; we therefore report NN, AA and WW separately, and the RSA of the euclidean RDM is not the sum of the RSA of its components. The literature uses “representational geometry” more broadly, for the distance structure including the norms; “angular component” here denotes only AA. The betas are decomposed in the same way, and the RSA of a component compares the model and fMRI RDMs of that component. The identity was checked numerically. These names label quantities defined here; they are not claims about physiological gain.

Analysis plan, amendments and inference

Pre-specified primary analysis. The ridge-readout analysis (§ 2.5) followed a written analysis plan with a frozen configuration (schema version 2, configuration hash 206fa875cca5), dated 20 August 2026. The plan was not publicly registered; its dating rests on in-text date stamps and the configuration hash, and the files were placed under version control on 3 October 2026. We therefore describe its criteria as fixed before the analysis rather than as preregistered. The plan asked whether the untrained-minus-backpropagation difference at V1 reported in Leutenegger [2026a] survives a fitted readout on the full Conv1 feature map. It fixed the layer, the voxel selection (split-half reliability above 0.2), the grid ladder g∈{1,2,4,8}g \in\{1, 2, 4, 8\}, the cross-validation scheme and seed, the regularisation grid and the aggregation (median over voxels, mean over subjects). The unit of inference was the training seed (n=5n = 5), tested with a one-sided exact sign-flip test over all 32 sign assignments. The effect was to count as surviving only if it held at g≥4g \ge4. The minimal detectable effect Δmin⁡\Delta_{\min} was derived from the pilot’s seed-level standard deviation by simulating this test. The reference effect was the value reported in version 1 of Leutenegger [2026a] (+0.044 ± 0.006 at 224 px); the current value is +0.030 [0.018, 0.042], and the verdict does not depend on this choice. The pre-specified rule labels the outcome “The effect disappears” when Δmin⁡\Delta_{\min} is below the reference effect; this is a decision label under the plan, not an equivalence test. Every seed contrast was negative (§ 3.1).

Amendments. Three amendments were adopted before any fit to THINGS data, on the basis of simulations with the dimensions of the primary data set: a hard assertion against selecting the lowest regularisation value for reliable voxels, selection of the regularisation by mean squared error with Pearson rr reported, and a widening of the regularisation grid. Four further decisions were taken after data had been seen, and we list them as such. (i) The regularisation was kept at minimum-MSE selection rather than a tolerance rule, after a tolerance measurement was known; this adds no degree of freedom. (ii) After the pilot aborted on the assertion, its remedy clause was amended to require a diagnosis before any change. (iii) The abort was then replaced by a post-hoc gate that classifies each lower-boundary selection as a plateau or a true optimum below the grid; only the latter would invalidate a cell. This was decided after a diagnosis showed that all observed boundary selections were plateaus, and it leaves all scores unchanged. No cell was excluded in any analysis. (iv) Features for g=2g = 2 were extracted on CPU, whereas all other features came from a GPU extraction. The pre-specified reproduction gate (absolute tolerance 10−610^{-6}) failed because of floating-point differences; we replaced it by a downstream-equivalence gate, which passed (maximum absolute difference in median rr of 1.3×10−81.3 \times10^{-8} at g=4g = 4). Four operational details not stated in the plan (significance level 0.05; “holds at g≥4g \ge4” as p≤0.05p \le0.05 at both g=4g = 4 and g=8g = 8; the decision based on the largest Δmin⁡\Delta_{\min} at g≥4g \ge4; averaging over subjects only for complete cells) were committed before the primary run. An implementation error initially derived Δmin⁡\Delta_{\min} from the primary analysis’s own seed variability instead of the pilot’s; we report the pre-specified value and the original value as supplementary. The verdict is the same under both.

Blinding and deviations. The primary analysis was not blind. Results for subject 1 at g∈{1,4,8}g \in\{1,4,8\} and for subjects 2 and 3 at g=1g = 1 were known from earlier steps; only g=2g = 2 and subjects 2 and 3 at g≥4g \ge4 were new. Several pre-specified instruments were not implemented: a capacity sweep, a PLS readout with 25 components, a ridge fit on the repeated-image test set, and a leave-one-subject-out bound for the ridge scores. The image-wise cell shuffle was run for one seed at 224 px only, and the grid ladder only at 224 px.

Exploratory analyses. All analyses in § 3.2–§ 3.6 lie outside the pre-specified plan, which excluded, among others, other layers and band-pass decompositions. They were developed sequentially. For each step, predictions and decision rules were written down and committed to version control before the step’s scripts were run; the commit timestamps, not the dates in file names, are authoritative and are listed in the repository. We report every pre-stated prediction, including the 6 that were not met (Appendix § 9). Run-grouped cross-validation folds were an additional sensitivity analysis and were not part of the plan.

Inference and its limits. We distinguish three kinds of statements. (a) The pre-specified primary analysis, whose decision rule was fixed before its run. (b) Exploratory analyses with pre-stated rules: each rule was committed before its step was run, and we report it as met or not met. The rules of later steps were written after the results of earlier steps were known, and we chose the thresholds ourselves, so “met” means consistent with a pre-stated rule, not independently confirmed. The analyses of § 3.5 and Appendices § 11 and § 12 (ablation of the mean-profile component, ridge robustness and the bootstrap with re-estimated normalisation) were added in response to criticism of an earlier draft; their rules were committed before their scripts (amendment dated 2026-10-16, committed on 2026-10-18 as 74853dc; the commit timestamp is authoritative). The count of distinct multiverse analyses and the ablation on each side (Appendix § 13) followed further criticism of the revised draft; their rules were committed before their scripts (amendment dated 2026-10-18, committed as ac59694). (c) Descriptive observations without a rule, which we label as such.

The intervals have different sources in different parts of the paper. For RSA, we resampled stimuli (1000 bootstrap resamples, seed 20261002) and computed differences between conditions from the same resamples; intervals are 95% percentile intervals. They describe variation over the 720 stimuli, for these three subjects and these networks; training seeds are averaged within each resample and are not part of the intervals. For ridge scores, the unit of inference is the training seed (n=5n = 5). With 32 sign assignments, the smallest attainable one-sided pp is 1/32 ≈\approx 0.031 and the smallest two-sided pp is 2/32 = 0.0625, so the test cannot support a two-sided claim at the 5% level; we use it, as pre-specified, for the positive direction only. Seed standard deviations do not include variation over stimuli, voxels or subjects. For each ImageNet model we used one set of trained weights, and counts such as “5 of 6 models” are descriptive. The 36 cells of the multiverse share the same stimuli and resamples and are strongly dependent; intervals are not corrected for multiple comparisons, and counts of cells are descriptive. With three subjects, the results describe these subjects and do not generalise to a population of observers (Table 1). Code, configuration, amendments and the full commit history are available in the accompanying repository.1

StatementStatusUncertainty reportedMain caveat
Untrained below Backprop under ridge at V1, also in MSE skill and per-fold rr (Section 3.1)Pre-specified; verdict “effect disappears”; robustness checks exploratory (F-R1 to F-R3 met)5 seeds; one-sided sign-flip test; no interval over stimuli, voxels or subjectsOne layer; sub-03 near zero; low absolute scores (median rr 0.086; MSE skill 0.0111); not blind
V1 ranking depends on the analysis (Section 3.3)Exploratory; GEN2 met; cell counts descriptiveStimulus-bootstrap 95% intervals per cell, uncorrected; cells dependent36 chosen cells, 28 distinct analyses; three subjects
Advantage absent from the component orthogonal to the mean profile (Section 3.5)Exploratory; M1, M2, M4 met; M3 not metStimulus-bootstrap intervalsParallel component alone unresolved; wide interval; paired contrast of the one-sided ablations not tested; paired full-minus-orthogonal contrast resolved (X1)
Raw RSA covaries with mean-profile alignment (Section 3.4)Exploratory; G1, GEN1 met; G2 not metSpearman over 70 and 64 cells, no interval; cells not independentCorrelational; the ablation tests only the orthogonal remainder
Trained ahead on the unweighted angular component at V1 and LOC (Section 3.6)Exploratory; SEP3, SEP4 met; SEP2 not metStimulus-bootstrap intervals; model counts descriptiveWeighted term WW not resolved; one weight set per ImageNet model; convnext_tiny exception
Resolution slope of own CNN vanishes under centring or standardisation (Sections 3.2, 3.3)Exploratory; R3, O2, P_ridge metStimulus-bootstrap intervals; seed SD for ridgeRests on three rules together; ImageNet slope remains (GEN4 not met)
Luminance-and-contrast bank above every network at V1 under ridge (Section 3.1)Descriptive instrument; joint readout F-R4 metPoint estimates per subject, mean over three subjectsNo interval for the bank; neural gain and BOLD not separable

Table 1. Strength of the evidence for the main statements. “Exploratory” denotes analyses with rules committed before their step was run, written after earlier results were known.

Results

Sign flip between RSA and ridge (including the pre-specified primary analysis)

In the pre-specified primary analysis (§ 2.7), the untrained (uncalibrated BN statistics) network scored below Backprop at every rung of the grid ladder (Figure 2). The untrained-minus-Backprop contrast in median rr over reliable V1 voxels, averaged over subjects, was −0.021-0.021, −0.035-0.035, −0.043-0.043 and −0.045-0.045 at g=1,2,4,8g = 1, 2, 4, 8. The number of positive seeds was 0, 0, 0 and 0 of 5, and the one-sided sign-flip pp was 1.000, 1.000, 1.000 and 1.000. The pre-specified minimal detectable effect, derived from the pilot’s seed-level standard deviation (0.0082), was Δmin⁡=0.014\Delta_{\min} = 0.014 (supplementary value from the primary analysis’s own seed variability: 0.016 at g=4g = 4 and 0.016 at g=8g = 8), below the reference effect of +0.044. The criterion for survival at g≥4g \ge4 was not met. Because Δmin⁡\Delta_{\min} was smaller than the reference effect, absence is uninformative under the plan, and the pre-specified rule classifies this outcome as “The effect disappears” rather than “The method cannot resolve it”. In every seed the contrast was negative. With five seeds the smallest attainable one-sided pp is 1/32 and the test addresses only the positive direction, so the negative direction is an observation, not a tested effect. Per-subject values are given in Appendix § 10; the mean over subjects includes sub-03, where ridge scores were close to zero. The four secondary contrasts against Backprop (untrained BN-calibrated, FA, PC, STDP) were also negative at every gg; the largest was −0.018, and at most 0 of 5 seeds were positive in any of them.

Figure 2. Pre-specified primary analysis. Median rr over reliable V1 voxels (Conv1, NATIVE 224 px), averaged over the three subjects, against grid size gg (logarithmic axis). Coloured lines: the six conditions, mean over 5 seeds, error bars SD over seeds. Grey dashed and dotted lines: control banks (luminance and contrast, darker; pixels; Gabor). The permuted-assignment control is not shown (§ 3.1).

In an exploratory check of this result (three subjects, Conv1, NATIVE 224 px; Appendix § 11), the untrained (uncalibrated BN statistics) minus Backprop contrast was negative in two further scores. In the MSE skill (one minus the ratio of the squared error to that of the training-fold mean) it was −0.0041, −0.0055, −0.0069 and −0.0074 at g=1,2,4,8g = 1, 2, 4, 8, and in the per-fold rr it was −0.0166, −0.0220, −0.0281 and −0.0311 (mean over subjects and seeds); no seed was positive in either score at any gg (F-R1 and F-R2: met). A model that predicts the training-fold mean scored a pooled rr of −0.092 (mean over subjects); per-fold rr is undefined for predictions that are constant within each fold. Under 50 permutations of the stimulus assignment of the features (seed-0 untrained and Backprop, g=1g = 1 and g=4g = 4), the 97.5th percentile of the null lay between −0.062 and −0.037; the observed scores in sub-01 and sub-02 (from 0.052 to 0.120) lay above it (F-R3: met). In sub-03, described without a rule, the observed scores (from −0.021 to 0.008) also lay above the 97.5th percentiles of their nulls. A readout of each network together with the luminance-and-contrast bank scored below the bank alone in all six conditions at g=1g = 1 and g=4g = 4 (differences from −0.047 to −0.013): for Backprop the joint readout reached 0.068 and 0.090, against 0.081 and 0.106 for the bank alone, and for the untrained network 0.060 and 0.064 (F-R4, joint minus bank at most 0.010 for Backprop: met). Concatenating network features with the bank under a shared ridge penalty did not improve prediction. With 32 bank and 512 network features at g=4g = 4, the two groups may need different penalties (banded ridge), which we did not fit; the result does not show how much variance the networks and the bank share. Absolute levels were low: the median rr of the best network condition, Backprop at g=4g = 4, was 0.086 (mean over subjects); this correlation score should not be read as variance explained. The absolute MSE skill of Backprop at g=4g = 4 was 0.0111 and that of the untrained network 0.0042 (median over reliable voxels, mean over subjects and seeds, relative to the training-fold mean).

Among the control banks, the luminance and contrast bank reached 0.081, 0.087, 0.106 and 0.111 at g=1,2,4,8g = 1, 2, 4, 8, above the mean of every network condition at every gg, including Backprop (0.064, 0.070, 0.086, 0.088). The pixel and Gabor banks ranged from −0.006 to 0.021. At g=4g = 4 and g=8g = 8 the intact grid exceeded its image-wise cell shuffle for both the untrained seed and the Backprop seed-0 network in 3 of 3 subjects; this instrument covers one seed per condition and is descriptive. The permuted-assignment control, the pipeline’s negative test, for which the plan expected a value near zero, gave −0.078, −0.075, −0.074 and −0.075; negative values are expected for features without signal under the pooled out-of-fold correlation (§ 2.5), and the control is not shown in Figure 2. All 26 lower-boundary selections at reliable voxels (Appendix § 6) were classified as plateaus (26), and no cell was excluded.

In an exploratory analysis of sub-01 (501 reliable V1 voxels, g=1g = 1, NATIVE 224 px), RSA and ridge disagreed in sign on the same voxels. RSA on the reliable voxels gave an untrained-minus-Backprop difference of +0.028; ridge gave −0.029, and −0.051 and −0.058 at g=4g = 4 and g=8g = 8 (Figure 1). Four candidate sources of the disagreement were tested against rules fixed before each run, and none was supported. The voxel set: the RSA difference stays positive on the voxels the ridge readout uses. The pooled out-of-fold score: with per-fold rr the ridge contrast was −0.023 (a positive value would have implicated pooling), and the two networks differed by 0.055 in the share of selections at the largest α\alpha, below the threshold of 0.10 for a collapse asymmetry. The fold convention: with run-grouped folds (Appendix § 7) the contrast was −0.027 (a positive value would have implicated leakage within runs). The score: correlating RDMs built from the ridge predictions with the fMRI RDM gave −0.004 with random folds (2 of 5 seeds positive) and −0.008 with run-grouped folds. This meets the pre-stated rule (≤ 0), but the random-fold value is close to zero: the analysis shows that the RSA advantage of the untrained network does not carry over to RDMs of ridge predictions, not that it reverses there. In the same subject, ridge did not separate the two untrained networks (0.002; rule ≤ 0.010: met), whereas RSA separated them by 0.047. In sub-02 and sub-03 the signs of the predicted pattern held: RSA +0.056 and +0.008, ridge −0.023 and −0.011. Sub-03 has 286 reliable voxels, ridge scores near zero in every condition and RSA and ridge differences close to zero; it contributes little to the pattern.

Figure 1. Untrained (uncalibrated BN statistics) minus Backprop at V1 (Conv1) against evaluation resolution; NATIVE inputs left of the dotted line, UPSAMPLED 224 px (open markers) right. RSA with raw features (circles) and channel-z-scored features (squares): per subject on cross-run pairs, averaged over subjects and 5 seeds; bars are 95% stimulus-bootstrap CIs where computed. Ridge (triangles): median rr over reliable voxels, mean over sub-01 and sub-02, error bars SD over seeds.

Channel standardization removes gap and slope

In an exploratory analysis, each channel of the GAP features was z-scored over the 720 stimuli before RSA. In sub-01 on the reliable voxels (S1), the untrained (uncalibrated BN statistics) minus Backprop difference became −0.015, below the threshold of 0.010 for the scale hypothesis, which was therefore supported; the two untrained networks differed by 0.002 (control criterion ≤ 0.010: met).

With the RSA convention of § 2.3 (all ROI voxels, three subjects, stimulus bootstrap), raw features reproduced the published values (R1, met): +0.030 [0.018, 0.042] at NATIVE 224 px and −0.001 [−0.011, 0.009] at 32 px. Absolute RSA values at V1 were low (ρ\rho between 0.008 and 0.053 in the paper cell at NATIVE 224 px, means of the six conditions). The leave-one-subject-out lower bound of the noise ceiling was 0.051 [0.039, 0.064] on cross-run pairs [Leutenegger, 2026b]; the upper bound is not informative with three subjects, so we cannot say how close the models are to the ceiling. With z-scored channels the V1 gap was −0.011 [−0.017, −0.005] at 224 px (R2, upper bound ≤ 0.010: met) and −0.012 [−0.018, −0.007] at 32 px. The resolution slope fell from Sraw=+0.031S_{\mathrm{raw}} = +0.031 [0.024, 0.038] to Sz=+0.002S_z = +0.002 [−0.004, 0.007] (R3, Sz≤0.5SrawS_z \le0.5 S_{\mathrm{raw}}: met; Figure 1). At LOC the Backprop-minus-untrained gap at 224 px was +0.011 [0.003, 0.021] with raw and +0.007 [0.002, 0.011] with z-scored channels (R4, CI above 0 under z: met). With content limited to 32 px (UPSAMPLED 224), the V1 gap was +0.035 [0.023, 0.046] raw and +0.002 [−0.001, 0.005] z-scored (R5, upper bound ≤0.010\le0.010: met).

The moment ratio m=∥μ∥2/(∥μ∥2+∑kσk2)m = \lVert\mu\rVert^2/(\lVert\mu\rVert^2+\sum_k \sigma_k^2) of the channel means μ\mu and SDs σk\sigma_k tracks the difference dd between raw and z-scored RSA: over 60 checkpoint-by-resolution cells of the untrained network and Backprop, Spearman(m,d)=0.68(m,d) = 0.68 (MR1, called M1 in the step 5 amendment; threshold 0.6: met). For Backprop, mm was ordered UPSAMPLED 224 < NATIVE 224 < NATIVE 32 in 5 of 5 seeds, and the untrained network’s range of mm (0.068) was below half of Backprop’s (P_mech1, met). P_mech2 was not met (Appendix C): Backprop’s mm was higher for band-pass inputs without the coarse band than for the coarse band alone in 5 of 5 seeds, but Spearman(m,d)(m,d) over all 50 cells including the band-pass inputs was 0.27, below 0.6.

Ridge on the same GAP features (g=1g = 1) at every NATIVE resolution (P_ridge; sub-01 and sub-02) gave untrained-minus-Backprop contrasts of −0.026 at 32 px and −0.026 at 224 px, non-positive at all six resolutions, with a slope of −0.000, below half of SrawS_{\mathrm{raw}} (met; Figure 1).

Analysis multiverse

In an exploratory analysis, the V1 contrast between the untrained (uncalibrated BN statistics) network and Backprop was computed in all 36 cells of the multiverse (§ 2.4; Figure 3). The mathematical control O1 was met: euclidean RDMs under raw and centred features, and under paper and center720 betas, agreed to within 4×10−144 \times10^{-14}. O2 was met: the only cell with a strong resolution slope (CI above 0 and S≥0.5SpaperS \ge0.5 S_{\mathrm{paper}}) was the convention of Leutenegger [2026a], raw/correlation/paper, with S=+0.031S = +0.031 [0.024, 0.038]; 0 of 36 cells violated the prediction (at most 2 allowed). At NATIVE 224 px the V1 gap had 21 positive and 15 negative point estimates. Its CI excluded 0 in 13 positive cells, the paper’s cell and the 12 euclidean cells, and in 9 negative cells. Euclidean distances are invariant to centring of units and of voxels, so the 12 euclidean cells comprise 4 distinct analyses and the 36 cells 28; counted over distinct analyses, the CI excluded 0 above in 5, below in 9 and included 0 in 14 analyses. Cell counts are descriptive. Examples are +0.030 [0.018, 0.042] in the paper’s cell, −0.011 [−0.017, −0.005] with z-scored features and correlation distance, and +0.023 [0.004, 0.041] with raw features and euclidean distance. At LOC (Figure 4), Backprop minus untrained was positive in 36 of 36 cells and had a CI above 0 in 20 (O3, thresholds 30 and 18: met); the CI included 0 in all euclidean cells.

Figure 3. Untrained minus Backprop at V1 in the 36 cells of the analysis multiverse. Left: gap at NATIVE 224 px; right: resolution slope SS (224 minus 32 px). Rows: feature normalisation FF; columns: distance DD within fMRI normalisation BB. Colour: value on a diverging scale centred at 0. Dots: 95% stimulus-bootstrap CI excludes 0. Black frame: the convention of Leutenegger [2026a]. Mean over 5 seeds and 3 subjects. Euclidean cells agree in pairs of rows (raw and center; scale and z) and between the paper and center720 columns (centring invariance).

Figure 4. Backprop minus untrained at LOC (Conv3), NATIVE 224 px, in the 36 cells of the analysis multiverse. Layout, colour scale, dots and frame as in Figure 3.

For the six ImageNet models (contrast: untrained minus trained at V1, trained minus untrained at LOC), GEN2 was met: in 4 of 6 models two of the four bootstrap cells gave V1 gaps at 224 px whose CIs excluded 0 in opposite directions. GEN3 was not met: the LOC contrast had a CI above 0 in all four cells for 0 models (at least 5 required). The raw/euclidean cell included 0 for every model, for example +0.009 [−0.024, 0.043] for alexnet and −0.023 [−0.061, 0.013] for convnext_tiny. GEN4 was not met: 5 models had a raw/correlation slope SgenS_{\mathrm{gen}} (448 minus 224 px) with a CI excluding 0, and in none of them did channel z-scoring halve its magnitude. A slope remained under z-scoring, for example +0.008 raw versus +0.005 z-scored for alexnet and +0.008 versus +0.010 for vgg11_bn.

Mean-profile alignment

In an exploratory analysis, the mean-profile alignment of § 2.6 was computed for every model and for the V1 voxels of each subject (Figure 5). Over the 70 cells of the untrained (uncalibrated BN statistics) network and Backprop (5 seeds, six NATIVE resolutions and UPSAMPLED 224 px), calignc_{\mathrm{align}} predicts the raw-feature RSA at V1 with Spearman 0.97 (G1, threshold 0.6: met). Over 64 cells of the ImageNet models the same correlation was 0.79 (GEN1, met). G2 was not met: after the ranks of both RDMs were regressed on the ranks of their own alignment RDMs, the V1 gap at NATIVE 224 px was +0.004 [−0.004, 0.013], whose upper bound exceeds 0.010, and the slope was Spartial=+0.017S_{\mathrm{partial}} = +0.017 [0.011, 0.023], above half of SpaperS_{\mathrm{paper}}.

Figure 5. (a) Own CNN: raw-feature RSA at V1 against calignc_{\mathrm{align}} for the untrained (vermillion) and Backprop (blue) checkpoints; markers code the input condition. (b) ImageNet models: the same for trained (blue) and untrained (vermillion) networks; markers code the model. (c) Pearson correlation over stimuli between mean-profile alignment and stimulus mean luminance (orange) or RMS contrast (light blue), for V1 voxels per subject, the own CNN’s conditions and the ImageNet models.

Descriptively, V1 voxel alignment correlated with stimulus mean luminance at −0.28, −0.27 and −0.23, and with RMS contrast at 0.13, 0.20 and 0.13 (sub-01, sub-02, sub-03). At 224 px the untrained network’s alignment had the same signs (−0.44-0.44 with luminance, 0.450.45 with contrast); Backprop’s alignment correlated positively with luminance (0.400.40), and its correlation with contrast was close to zero (0.000.00).

Removing the mean-profile component

In an exploratory analysis prompted by criticism of an earlier draft (§ 2.7), each raw GAP pattern and each beta pattern was split into its projection on the mean profile and the remainder orthogonal to it (§ 2.6); each model component was compared with the same component of the brain, with the RSA convention and bootstrap of § 2.3. The full variant reproduced the stored paper cell of the own CNN and of the ImageNet models to within 9.9×10−179.9 \times10^{-17} (reproduction gate: met).

M1M1 was met: in the orthogonal remainder, the V1 contrast of the untrained (uncalibrated BN statistics) network minus Backprop at NATIVE 224 px was −0.002-0.002 [−0.007,0.003][-0.007, 0.003], with an upper bound of at most 0.0100.010, against +0.030+0.030 [0.018,0.042][0.018, 0.042] in the full variant. M2M2 was met: the resolution slope of the orthogonal remainder was 0.0000.000 [−0.003,0.003][-0.003, 0.003], below half of SpaperS_{\mathrm{paper}}. M4M4 was met: for the ImageNet models at 224 px, the upper bound of the orthogonal V1 contrast was at most 0.0100.010 in 6 of 6 models. M3M3 was not met: the parallel component gave +0.012+0.012 [−0.011,0.035][-0.011, 0.035] at NATIVE 224 px and a slope of +0.006+0.006 [−0.007,0.020][-0.007, 0.020]. The interval of the parallel gap includes both 0 and the raw gap, so these data do not decide whether the parallel component alone suffices. At LOC, described without a rule, Backprop minus untrained was +0.011 [0.003, 0.021] in the full variant, +0.005 [0.001, 0.010] in the orthogonal remainder and −0.002 [−0.020, 0.017] in the parallel component.

M3 was not met, and sufficiency of the parallel component was not established. The wording of the pre-stated rule (“M3 not met: it is not sufficient”) claimed more than a missed threshold with this interval can show, and we state only what the data resolve: the advantage is absent from the orthogonal remainder, and the parallel component alone did not resolve it. No physiological claim follows.

On the same resamples, the paired contrast of the V1 gap with the full variant on both sides minus the gap with the orthogonal remainder on both sides was +0.032 [0.022, 0.042] (X1: met), which restates that the drop between the full and the orthogonal variant is resolved. With the orthogonal remainder on one side only (Appendix § 13), the V1 contrast was −0.001 [−0.007, 0.006] for the orthogonal model against the full brain RDM and +0.006 [−0.001, 0.012] for the full model against the orthogonal brain RDM; both intervals include 0, so under the pre-stated rule no attribution to a side is claimed. Removing the component from only one side brought the point estimate from +0.030 in the full variant to values close to that of the orthogonal variant, and the intervals of both mixed cells lay below that of the full variant. This is consistent with an advantage that depends on the mean-profile component on both sides, as the alignment correlation calignc_{\mathrm{align}} of § 3.4 suggests; it is a descriptive observation, not covered by a rule, and we did not test a paired contrast.

Norm and angular components

In an exploratory analysis, the euclidean RDM on unit-centred features was decomposed into its norm component and its angular component (§ 2.6; Figure 6). Contrasts are untrained minus trained at V1 and trained minus untrained at LOC; for the own CNN they are untrained (uncalibrated BN statistics) and Backprop. The identity held to a maximum relative error of 3×10−123 \times10^{-12} (SEP1, threshold 10−810^{-8}: met).

Component RSA contrasts with $F =$ center and 95% stimulus-bootstrap confidence intervals for the own CNN and six ImageNet models

Figure 6. Component RSA contrasts with F=F = center and 95% stimulus-bootstrap CIs, for the own CNN (vermillion) and six ImageNet models (blue). Panels from left: V1 norm component, V1 angular component, LOC angular component, and the slope of the V1 angular contrast from 224 to 448 px for the ImageNet models (green; none for vit_b_16). V1: untrained minus trained; LOC: trained minus untrained.

SEP2 was not met: the V1 norm contrast (F=centerF = \text{center}) was +0.017 [−0.004, 0.036] in the own CNN, with a CI including 0, and had a CI above 0 in 4 of 6 ImageNet models. SEP3 was met: the V1 angle contrast was −0.009 [−0.017, −0.002] in the own CNN and had a CI below 0 in 5 of 6 ImageNet models. SEP4 was met: the LOC angle contrast was +0.007 [0.002, 0.012] in the own CNN and had a CI above 0 in 6 of 6 ImageNet models. Because SEP2 was not met, the pre-stated interpretation rule applies: the V1 ranking is not attributed to the norm or the angular component. On the angular component alone, the trained networks were ahead at V1 in the own CNN and in 5 ImageNet models, and at LOC in the own CNN and in 6 ImageNet models. The norm-weighted angular term WW gave a V1 contrast of +0.013 [−0.003, 0.030] in the own CNN, and its interval excluded 0 in 1 of 6 ImageNet models (vgg11_bn, +0.026 [0.004, 0.046]). The angular result thus concerns the unweighted angular dissimilarity AA; the norm-weighted term WW pointed in the direction of the norm component and was not resolved. Over the cross-run pairs, NN and AA were nearly uncorrelated (Spearman 0.09 in the own CNN, mean over checkpoints; 0.00, 0.01 and 0.01 in the fMRI of the three subjects), whereas AA and WW were correlated (0.75; 0.63, 0.68 and 0.67). In a regression of the ranks of the fMRI RDM on the ranks of NN, AA or both (own CNN, V1, NATIVE 224 px), each component accounted for about 0.1% of the rank variance of the fMRI RDM (R2R^2 for NN and AA together: 0.0017 for the untrained network, 0.0012 for Backprop), and no contrast of R2R^2 was resolved.

The resolution slope of the contrasts was a descriptive point fixed in advance without a decision rule. For the ImageNet models, the V1 angle contrast grew from 224 to 448 px with a CI above 0 in 4 of 5 models, for example +0.010 [0.007, 0.014] for resnet50; no norm slope had a CI excluding 0. For the own CNN neither component had a slope from 32 to 224 px under F=centerF = \text{center} (norm −0.004 [−0.013, 0.005], angle +0.001 [−0.004, 0.006]). convnext_tiny was the exception among the ImageNet models: its V1 angle contrast (−0.003 [−0.007, 0.002]) and its angle slope (−0.001 [−0.000, 0.002]) both had CIs including 0. We did not test an explanation for this exception.

Discussion

Strength of the evidence. Table 1 states for each main statement whether it was pre-specified, exploratory with a pre-stated rule, or descriptive, which uncertainty is reported, and the main caveat. Only the first row stems from the pre-specified analysis. All other rows are exploratory, and agreement among them is not independent confirmation, because later rules were written after earlier results were known.

What the V1 advantage measured. The advantage of the untrained network at V1 that we reported earlier [Leutenegger, 2026a] was not robust to the analysis choices examined here. Its sign changed across the multiverse: under correlation distance it disappeared when model units were standardised over the stimulus set, and a pre-specified ridge readout placed the untrained network below the trained one. Under Euclidean distance it persisted, without a resolved resolution slope, but the angular component favoured the trained networks. It was absent from the component orthogonal to the mean profile of the units, and removing the mean-profile component from either side alone brought the contrast close to zero, while the parallel component alone did not resolve it (M3). It was also not shared by the untrained network with calibrated batch-normalisation statistics. Raw correlation-distance RSA, the convention of the earlier work, covaried with the mean-profile alignment of model and voxels, across our networks and across six ImageNet architectures; removing the alignment RDM by rank regression left part of the resolution slope (G2, not met). We therefore read the earlier result as not reflecting angular structure and as sensitive to the mean-profile component under one convention; whether that component alone carries it remains open, and we do not claim that it is a physiological gain.

A luminance-and-contrast bank exceeds every network at V1 in these data. Under the ridge readout, the luminance-and-contrast bank predicted V1 better than every network condition at every pooling grid, including the backpropagation-trained network. These are point estimates (mean over three subjects, five seeds per network condition) without an interval for the bank; in each subject the bank scored above Backprop at every grid (Appendix § 10). A readout of each network together with the bank scored below the bank alone (Section 3.1), so concatenating network features under a shared penalty did not improve prediction. We cannot say how much variance the bank and the networks share. This limits what V1 comparisons in these data can reveal about learning rules. Voxel alignment correlated with luminance and contrast, and the untrained network’s alignment shared these signs. Whether this pattern reflects neural mechanisms such as contrast normalisation [Carandini and Heeger, 2012] or properties of the BOLD measurement cannot be decided with these data; the norm of the voxel response vectors did not depend on luminance or contrast. Data with controlled contrast manipulations or electrophysiological recordings would be needed to separate the two.

Ridge and RSA address different questions. The three analyses ask different questions:

  • raw RSA asks which unaltered distance structure resembles the measured one;

  • centred angular RSA asks which direction-based structure resembles it after centring over the stimuli;

  • ridge asks which features predict the voxel responses better under this regularised linear readout.

A different ranking under ridge therefore does not by itself show that the RSA ranking is wrong.

Sensitivity is not invalidity. That rankings change across defensible analyses shows that the metric matters; it does not show that any single analysis is invalid. Raw, centred and standardised variants and euclidean, correlation and cosine distances may capture different aspects of a representation, and a fitted readout asks another question again. Our ablation shows that the earlier advantage depends on the component along the mean profile of the units; it does not show that this component is a biologically meaningful property of the representation, and we do not claim that it is. Which representational properties a benchmark should preserve is a question about what the comparison is meant to explain, and it cannot be settled by whether a choice yields a stable ranking (on what makes a model explanatory, see Cao and Yamins [2]). Our results support the narrower point that a brain-alignment score is a measurement defined by the evaluation procedure and its assumptions, not an intrinsic measure of a model.

Evaluation resolution. In our CIFAR-trained networks, the dependence of the V1 gap on evaluation resolution vanished under unit centring and standardisation and was absent in the ridge readout. In the ImageNet models, a small dependence on resolution above the training resolution remained under standardisation and resided in the angular component. Evaluation resolution can therefore affect the angular component of model–brain similarity, but in our earlier work the size of the effect depended on uncentred units. The two model families differ in the ratio of evaluation to training resolution (up to sevenfold versus twofold), and we do not claim that the results transfer between these regimes.

Recommendations. Five practices follow from these results. (1) For correlation- or cosine-based ranking claims, report the sensitivity to centring units over the evaluated stimulus set on both the model and the brain side, and justify the chosen reference; normalising voxels over a larger set of trials is not equivalent. (2) Report norm and angular components separately, for example with the norm and angle decomposition used here. (3) For ranking claims in a paper, report a small multiverse of distance measure and normalisation. For benchmark platforms, where a full multiverse for every submission may be too costly, a fixed primary procedure with a stated rationale, plus supplementary diagnostics for the leading models (for example the ranking under centring and the norm and angular components), would make the dependence visible without changing the reference score. (4) Include simple stimulus-statistic baselines alongside network models. (5) State how the normalisation statistics of untrained networks were set, since calibrated and uncalibrated variants behaved differently.

Consequences for our earlier work. The headline of Leutenegger [2026a], that untrained networks exceed backpropagation at V1 at 224 px and that this gap depends on evaluation resolution, and the corresponding claim in Leutenegger [2026b] are not robust to the analysis choices examined here. An audit of our earlier preprints identified the affected statements; revised versions refer to this work.

Limitations. The data comprise three subjects from a single data set, and the intervals reported here describe different sources of variation (§ 2.7): stimulus resampling for RSA and training seeds for ridge. Neither includes variation across subjects, and for each ImageNet model we used one set of trained weights. Ridge scores in one subject were close to zero in every condition, and the primary ridge result is a mean over all three subjects (Appendix § 10). The absolute scale was small: each norm or angular component accounted for about 0.1% of the rank variance of the fMRI RDM, and the median voxelwise correlation of the best ridge readout was 0.086 (a correlation score, not a measure of variance explained; absolute MSE skill 0.0111); no noise ceiling was estimated for ridge, and the pre-specified leave-one-subject-out bound for ridge was not implemented. The primary analysis was not blind, several pre-specified instruments were not implemented, and the exploratory analyses were developed sequentially, with intervals not corrected for multiple comparisons. The re-estimation of the normalisation statistics within the bootstrap covered three cells and not the ablation. Run-grouped folds and the encoding-RSA check were run for one subject. A banded ridge readout, with a separate penalty per feature group, was not fitted. 6 pre-stated predictions failed (Appendix § 9); among them, the LOC ranking was not resolved under Euclidean distance in the ImageNet models (GEN3), and the parallel component alone was not resolved (M3), so the mean-profile account remains a candidate. Whether that component reflects a property of neural representations or of the measurement is not established. The norm-weighted angular term WW behaved differently from the unweighted angular component AA. The repeated-image test set (100 images, 12 repetitions) was used only for voxel reliability, and cross-validated distances were not computed. Counts of lowest-value selections of the regularisation varied between identical runs through near-ties under single-thread BLAS, while the scores agreed (Appendix § 6). One ImageNet model, convnext_tiny, showed no V1 effect in either component; its stem normalises activations across channels at each position [3], which may remove the offset and norm differences studied here, but we did not test this. Our use of “angular component” is a definition, not a claim about neural coding. Layers were fixed in advance and not selected for fit; other layers may behave differently.

Conclusion. Whether untrained networks beat trained ones at V1 in model–brain RSA depends on the analysis. The advantage we reported earlier was absent from the component orthogonal to the mean profile and from centred angular structure; whether the mean-profile component alone carries it was not resolved. Reporting the sensitivity to centring over the evaluated stimuli, reporting components separately and including simple baselines make such comparisons interpretable.

Acknowledgements

We thank Kartik Pradeepan for detailed feedback on a draft of this manuscript, in particular on the distinction between sensitivity and validity and on the recommendations for benchmarking platforms.

Claude (Anthropic), a large language model, was used as an assistant throughout this work: to suggest and refine analyses and decision rules, to write and run analysis code, and to draft most of the manuscript text. The author defined the research question, decided which analyses to run and which claims to make, checked the results with reproduction tests and gates, and takes full responsibility for the content, including any errors.

Lower-edge gate

The ridge fits select the regularisation α\alpha per voxel and outer fold from a grid whose lowest value is 10−610^{-6}. Following the amendment of 2026-10-03 (§ 2.7), the selection of the lowest value at a reliable voxel does not abort the run. It is logged and, after the run, classified on the mean inner-fold MSE curve of that voxel in that outer fold. The curve is extended by the diagnostic values 10−710^{-7}, 10−810^{-8} and 10−910^{-9}, and the relative gain is (MSE(10−6)−MSEmin⁡)/MSE(10−6)(\mathrm{MSE}(10^{-6})-\mathrm{MSE}_{\min})/\mathrm{MSE}(10^{-6}), with MSEmin⁡\mathrm{MSE}_{\min} the minimum over the extended grid. A relative gain below 10−310^{-3} is a plateau and the selection is valid; otherwise the cell would be excluded from every result. Smaller values of α\alpha are not used, because there the solution is determined mainly by eigenvalues of the dual kernel at rounding level (∣λ∣≤1.9×10−12(|\lambda| \le1.9 \times10^{-12} in the diagnosed cases). The gate decides whether a cell is kept and never changes a score.

The first cell with such selections was a condition of the pilot (untrained network with calibrated BN statistics, one seed, g=1g = 1, sub-01). There the standardised training design had full column rank and no dead features (step report in the repository); the pilot diagnosis used a wider diagnostic grid than the gate. Table 2 lists every analysis with such selections. Steps 7 to 9 involve no ridge fits. The number of selections at the lowest value changed between runs of identical fits (4 versus 6 on the model fits of the primary analysis, 3 versus 4 in the pilot cell) because of near-ties under single-thread BLAS; scores agreed to 1.8×10−131.8 \times10^{-13} and the gate classification was unchanged.

AnalysisHitsPlateauExcludedMax. gain
Pilot diagnosis (sub-01, g=1g = 1)4401.6×10−41.6 \times10^{-4}
RSA versus ridge: collapse, random folds (sub-01)3304.4×10−84.4 \times10^{-8}
RSA versus ridge: run-grouped folds (sub-01)4404.8×10−64.8 \times10^{-6}
RSA versus ridge: encoding-RSA (sub-01; repeats the hits above)7704.8×10−64.8 \times10^{-6}
Other subjects, g=1g = 1 (sub-02, sub-03)1103.0×10−53.0 \times10^{-5}
Diagnosis of sub-03 (controls, g∈{1,4,8}g \in\{1, 4, 8\})3301.3×10−51.3 \times10^{-5}
Ridge across resolutions (sub-01, sub-02, g=1g = 1)202004.6×10−54.6 \times10^{-5}
Pre-specified primary analysis (3 subjects)262605.8×10−55.8 \times10^{-5}
Ridge robustness: model fits (3 subjects)4403.0×10−53.0 \times10^{-5}
Ridge robustness: joint readout3303.6×10−53.6 \times10^{-5}
Ridge robustness: permutation null3302.9×10−52.9 \times10^{-5}

Table 2. Selections of the lowest regularisation value at reliable voxels, classified by the plateau gate. In the pilot diagnosis the plateau criterion was applied to the relative gain of the wider diagnostic grid. All hits were plateaus, and no cell was excluded.

Run-grouped folds

Run-grouped cross-validation was an exploratory sensitivity analysis of sub-01 (501 reliable V1 voxels). It was not part of the pre-specified plan, which fixes random folds. The 720 stimuli come from 10 runs of 72. The 10 runs were permuted (seed 20260818) and consecutive pairs of runs formed the 5 outer test folds. Within each outer training set, the 8 remaining runs were permuted (seed 20260818 + 1000 (k+1)(k + 1) for outer fold kk) and consecutive pairs formed the 4 inner folds. Standardisation on training rows, the regularisation grid, the selection rule and the gate were unchanged. All 30 checkpoints were fitted at g∈{1,4,8}g \in\{1, 4, 8\}.

Table 3 gives the untrained (uncalibrated BN statistics) minus Backprop contrast, mean ±\pm SD over the 5 seeds with the number of positive seeds. The contrast was negative in every seed under both fold schemes. Run-grouped folds lowered the pooled rr of all 30 checkpoints by −0.046-0.046, −0.051-0.051 and −0.050-0.050 at g=1,4,8g = 1, 4, 8. Table 4 gives the contrast of encoding-RSA, in which the out-of-fold predictions of the 501 voxels form a model RDM that is compared with the fMRI RDM by the RSA statistic of § 2.3; “centred” denotes predictions centred within each outer fold.

ggPooled rr, random foldsPooled rr, run foldsPer-fold rr, run folds
1−0.029±0.008-0.029 \pm0.008 (0/5)−0.027±0.010-0.027 \pm0.010 (0/5)−0.026±0.008-0.026 \pm0.008 (0/5)
4−0.051±0.011-0.051 \pm0.011 (0/5)−0.048±0.011-0.048 \pm0.011 (0/5)−0.044±0.006-0.044 \pm0.006 (0/5)
8−0.058±0.010-0.058 \pm0.010 (0/5)−0.058±0.008-0.058 \pm0.008 (0/5)−0.048±0.006-0.048 \pm0.006 (0/5)

Table 3. Untrained minus Backprop, median rr over reliable voxels of sub-01 (mean ±\pm SD over 5 seeds; positive seeds of 5).

Folds (g=1g = 1)CentredUncentred
Random folds−0.004±0.009-0.004 \pm0.009 (2/5)−0.001±0.009-0.001 \pm0.009 (3/5)
Run-grouped folds−0.008±0.008-0.008 \pm0.008 (0/5)−0.007±0.009-0.007 \pm0.009 (2/5)

Table 4. Untrained minus Backprop, encoding-RSA of sub-01 (mean ±\pm SD over 5 seeds; positive seeds of 5).

Band-pass inputs

The band-pass inputs were built from the image after resizing and centrecropping to 224 px and before the channel normalisation. Gaussian low-pass filters LrL_r with r∈{112,56,28,14}r \in\{112, 56, 28, 14\} have an amplitude response of 0.5 at r/2r/2 cycles per image; the octave bands are differences of consecutive filters. Band 0–14 is the coarse residual L14L_{14} and contains the image mean. Condition “only 0–14” keeps this band alone, and condition “notch 0–14” is the native image minus this band, shifted by the CIFAR-10 channel-mean grey. Both were evaluated at NATIVE 224 px with all 30 checkpoints.

Table 5 gives the V1 contrast (untrained minus Backprop) and the LOC contrast (Backprop minus untrained) with raw and channel-z-scored features, with 95% stimulus-bootstrap intervals. The V1 contrast with the coarse band alone was positive under raw features, as for native input (+0.030); without the coarse band it was negative. Table 6 gives the moment ratio mm of § 3.2, mean over seeds. The prediction Pmech2P_{\mathrm{mech2}} that relates mm to the difference between raw and z-scored RSA was not met (Appendix § 9).

Inputmm, untrainedmm, Backprop
Native 224 px0.7070.638
Only 0–140.6190.484
Notch 0–140.8320.830

Table 6. Moment ratio mm of the Conv1 features (mean over 5 seeds).

Only 0–14Notch 0–14
V1, raw0.033 [0.021, 0.044]−0.021 [−0.035, −0.007]
V1, z0.004 [0.001, 0.007]−0.010 [−0.018, −0.003]
LOC, raw0.002 [−0.004, 0.009]0.004 [−0.005, 0.013]
LOC, z0.002 [−0.002, 0.006]0.003 [−0.001, 0.006]

Table 5. Contrasts for band-pass inputs (NATIVE 224 px; V1: untrained minus Backprop, LOC: Backprop minus untrained; 95% stimulus-bootstrap intervals).

Failed predictions

Table 7 lists the 6 pre-stated predictions of the exploratory analyses that were not met. Each rule was committed before the scripts of its step were run (§ 2.7).

PredictionRule and thresholdResultSection
P_mech2mBP(notch)>mBP(only)m_{\mathrm{BP}}(\mathrm{notch}) > m_{\mathrm{BP}}(\mathrm{only}) in 5 of 5 seeds, and Spearman(m,d)≥0.6(m,d) \ge0.6 over 50 cells5 of 5 seeds; Spearman 0.273.2
G2partial V1 gap at NATIVE 224 px with upper CI bound ≤0.010\le0.010, and Spartial≤0.5 SpaperS_{\mathrm{partial}} \le0.5\,S_{\mathrm{paper}}gap +0.004+0.004 [−0.004-0.004, 0.0130.013]; Spartial=+0.017S_{\mathrm{partial}} = +0.017 [0.0110.011, 0.0230.023]3.4
GEN3LOC trained −- untrained at 224 px with CI above 0 in all 4 bootstrap cells, for ≥5\ge5 of 6 models0 models; the raw/euclidean cell includes 0 for every model3.3
GEN4for every model whose SgenS_{\mathrm{gen}} (raw, correlation) has a CI excluding 0: ∣Sgen(z)∣≤0.5∣Sgen(raw)∣\lvert S_{\mathrm{gen}}(z)\rvert\le0.5\lvert S_{\mathrm{gen}}(\mathrm{raw})\rvert5 qualifying models, none within the bound3.3
SEP2V1 norm contrast (F=F = center) >0> 0 with CI above 0 in the own CNN and in ≥4\ge4 of 6 ImageNet modelsown CNN +0.017+0.017 [−0.004-0.004, 0.0360.036]; ImageNet 4 of 63.6
M3parallel component: V1 gap at NATIVE 224 px >0> 0 with CI above 0, and resolution slope with CI above 0gap +0.012+0.012 [−0.011-0.011, 0.0350.035]; slope +0.006+0.006 [−0.007-0.007, 0.0200.020]3.5

Table 7. Pre-stated predictions that were not met. Intervals are 95% stimulus-bootstrap percentile intervals.

Per-subject results of the primary analysis

Table 8 gives the untrained (uncalibrated BN statistics) minus Backprop contrast of the pre-specified primary analysis for each subject (median rr over reliable V1 voxels; mean ±\pm SD over the 5 seeds; positive seeds of 5). The main text reports the mean over the three subjects. Table 9 gives the median rr of the luminance and contrast bank and of the Backprop condition (mean over seeds) for each subject. In each subject and at each gg the bank scored above the Backprop condition. Sub-03 has 286 reliable voxels. All values are point estimates per subject, without an interval over stimuli or voxels.

ggsub-01sub-02sub-03
1−0.029±0.008-0.029 \pm0.008 (0/5)−0.023±0.004-0.023 \pm0.004 (0/5)−0.011±0.006-0.011 \pm0.006 (0/5)
2−0.049±0.011-0.049 \pm0.011 (0/5)−0.034±0.006-0.034 \pm0.006 (0/5)−0.021±0.006-0.021 \pm0.006 (0/5)
4−0.051±0.011-0.051 \pm0.011 (0/5)−0.045±0.011-0.045 \pm0.011 (0/5)−0.032±0.009-0.032 \pm0.009 (0/5)
8−0.058±0.010-0.058 \pm0.010 (0/5)−0.047±0.011-0.047 \pm0.011 (0/5)−0.030±0.009-0.030 \pm0.009 (0/5)

Table 8. Untrained minus Backprop per subject.

ggsub-01sub-02sub-03
10.118 / 0.0940.102 / 0.0960.023 / 0.002
20.129 / 0.1040.106 / 0.1040.027 / 0.002
40.155 / 0.1230.132 / 0.1220.030 / 0.013
80.161 / 0.1260.135 / 0.1220.035 / 0.014

Table 9. Median rr of the luminance and contrast bank / of Backprop, per subject.

Ridge robustness

These checks of the pre-specified ridge result are exploratory (§ 2.7). They used the pre-specified configuration (three subjects, Conv1, NATIVE 224 px, reliable V1 voxels, the same nested cross-validation and regularisation grid) with the fitting code unchanged; the pooled median rr of every model fit reproduced the primary analysis to 1.8×10−131.8 \times10^{-13}. Table 10 gives the untrained (uncalibrated BN statistics) minus Backprop contrast in two further scores, the MSE skill (one minus the squared error relative to that of the mean of the training rows of each outer fold, median over reliable voxels) and the per-fold rr (Pearson rr within each outer fold, averaged over folds, median over reliable voxels), mean over subjects.

ggMSE skillper-fold rr
1−0.0041±0.0008-0.0041 \pm0.0008 (0/5)−0.0166±0.0051-0.0166 \pm0.0051 (0/5)
2−0.0055±0.0010-0.0055 \pm0.0010 (0/5)−0.0220±0.0053-0.0220 \pm0.0053 (0/5)
4−0.0069±0.0013-0.0069 \pm0.0013 (0/5)−0.0281±0.0068-0.0281 \pm0.0068 (0/5)
8−0.0074±0.0013-0.0074 \pm0.0013 (0/5)−0.0311±0.0067-0.0311 \pm0.0067 (0/5)

Table 10. Untrained minus Backprop, mean ±\pm SD over 5 seeds (positive seeds of 5), mean over subjects.

A model that predicts the mean of the training rows of each fold scored a pooled rr of −0.091-0.091, −0.088-0.088, −0.097-0.097 in sub-01, sub-02 and sub-03; its per-fold rr is undefined, because the prediction is constant within each fold. Table 11 gives the null distribution of the pooled median rr under 50 permutations of the stimulus assignment of the feature rows (the same permutation for all voxels; seed-0 checkpoints) next to the observed value.

SubjectggConditionNull mean[2.5%, 97.5%]Observed
sub-011untrained−0.070[−0.083, −0.040]0.052
sub-011Backprop−0.072[−0.083, −0.046]0.088
sub-014untrained−0.073[−0.084, −0.041]0.060
sub-014Backprop−0.072[−0.086, −0.046]0.118
sub-021untrained−0.068[−0.078, −0.056]0.073
sub-021Backprop−0.069[−0.079, −0.062]0.096
sub-024untrained−0.071[−0.081, −0.060]0.071
sub-024Backprop−0.069[−0.080, −0.055]0.120
sub-031untrained−0.074[−0.091, −0.038]−0.014
sub-031Backprop−0.077[−0.093, −0.048]−0.004
sub-034untrained−0.075[−0.094, −0.037]−0.021
sub-034Backprop−0.076[−0.093, −0.048]0.008

Table 11. Permutation null of the pooled median rr (seed-0 checkpoints).

For the joint readout, the network features and the luminance-and-contrast bank of the same gg were concatenated into one feature matrix. Each feature was standardised with the statistics of the training rows of each split, as in every other fit, and the cross-validation scheme and regularisation grid were unchanged. Table 12 compares the bank alone, the network alone and the joint readout.

ggNetworkBank aloneNetwork aloneNetwork and bank
1untrained0.0810.0430.060
1Backprop0.0810.0640.068
4untrained0.1060.0430.064
4Backprop0.1060.0860.090

Table 12. Median rr over reliable V1 voxels, mean over seeds and subjects.

Bootstrap with re-estimated normalisation

In the main analyses, unit means and SDs, the feature transformations FF, the voxel normalisations BB, the mean profile and the norms are estimated once on the 720 stimuli, and the bootstrap resamples pairs of the resulting fixed RDMs (§ 2.3). In this exploratory check, the statistics were re-estimated from the resampled rows of each of the first 200 resamples (duplicates kept, self-pairs dropped), and the RDMs were rebuilt with them. Table 13 compares the V1 contrast of the untrained (uncalibrated BN statistics) network minus Backprop at NATIVE 224 px in three cells. The classification of the interval relative to 0 was the same in all three cells (BN1: met). With euclidean distance, centring does not change the distances, so the two intervals agree by construction over the same resamples.

Cell (FF, DD, BB)ValueFixed (1000)Re-estimated (200)Interval
zz, correlation, paper−0.011-0.011[−0.017,−0.005][-0.017, -0.005][−0.017,−0.005][-0.017, -0.005]below 0
center, euclidean, paper0.0230.023[0.004,0.041][0.004, 0.041][0.002,0.039][0.002, 0.039]above 0
center, correlation, center7200.0010.001[−0.005,0.007][-0.005, 0.007][−0.005,0.007][-0.005, 0.007]includes 0

Table 13. V1 contrast with 95% stimulus-bootstrap intervals, normalisation fixed (1000 resamples) or re-estimated within each resample (first 200 resamples).

Ablation on each side

This exploratory analysis was added after further criticism of a revised draft; its rule was committed before its script (§ 2.7). The mean-profile decomposition of § 2.6 was applied to the model side, the brain side, both or neither. Table 14 gives the V1 contrast at NATIVE 224 px (raw GAP features, correlation distance, BB = paper) for the untrained (uncalibrated BN statistics) network minus Backprop in all four combinations, and for the six ImageNet models (untrained, mean of three initialisations, minus trained, 224 px) in the two mixed combinations. All values use the same 1000 stimulus-bootstrap resamples. The full/full and orthogonal/orthogonal cells reproduce the full and orthogonal variants of § 3.5. In the ImageNet models, the CI lay above 0 in 0 and below 0 in 3 models with the orthogonal model against the full brain RDM, and above 0 in 0 and below 0 in 1 models with the full model against the orthogonal brain RDM (descriptive, no rule).

ModelModel RDMBrain RDMV1 contrast [95% CI]
own CNNfullfull0.0300.030 [0.018,0.042][0.018, 0.042]
own CNNfullorthogonal0.0060.006 [−0.001,0.012][-0.001, 0.012]
own CNNorthogonalfull−0.001-0.001 [−0.007,0.006][-0.007, 0.006]
own CNNorthogonalorthogonal−0.002-0.002 [−0.007,0.003][-0.007, 0.003]
alexnetorthogonalfull−0.002-0.002 [−0.012,0.008][-0.012, 0.008]
vgg11_bnorthogonalfull−0.003-0.003 [−0.015,0.008][-0.015, 0.008]
resnet18orthogonalfull−0.011-0.011 [−0.018,−0.004][-0.018, -0.004]
resnet50orthogonalfull−0.012-0.012 [−0.020,−0.005][-0.020, -0.005]
convnext_tinyorthogonalfull0.0050.005 [−0.002,0.012][-0.002, 0.012]
vit_b_16orthogonalfull−0.024-0.024 [−0.041,−0.008][-0.041, -0.008]
alexnetfullorthogonal−0.001-0.001 [−0.007,0.005][-0.007, 0.005]
vgg11_bnfullorthogonal0.0050.005 [−0.001,0.013][-0.001, 0.013]
resnet18fullorthogonal−0.003-0.003 [−0.008,0.003][-0.008, 0.003]
resnet50fullorthogonal−0.003-0.003 [−0.008,0.003][-0.008, 0.003]
convnext_tinyfullorthogonal0.0030.003 [−0.002,0.008][-0.002, 0.008]
vit_b_16fullorthogonal−0.018-0.018 [−0.029,−0.008][-0.029, -0.008]
X1, own CNN: gap[full, full] − gap[orthogonal, orthogonal]+0.032+0.032 [0.022,0.042][0.022, 0.042] (met)

Table 14. V1 contrast with 95% stimulus-bootstrap CIs (own CNN: untrained minus Backprop, mean over 5 seeds; ImageNet: untrained minus trained); mean over three subjects.

References

  1. [1]G.-q. Bi and M.-m. Poo. Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type. Journal of Neuroscience, 18:10464–10472, 1998.pubmed.ncbi.nlm.nih.gov/9852584
  2. [2]R. Cao and D. Yamins. Explanatory models in neuroscience, part 2: functional intelligibility and the contravariance principle. Cognitive Systems Research, 85:101200, 2024. doi: 10.1016/j.cogsys.2023.101200.
  3. [3]Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976–11986, 2022. doi: 10.1109/CVPR52688.2022.01167.DOI
  4. [4]Sébastien Marcel and Yann Rodriguez. Torchvision the machine-vision package of torch. In Proceedings of the 18th ACM International Conference on Multimedia, pages 1485–1488, 2010. doi: 10.1145/1873951.1874254.DOI
  5. [5]T. Masquelier and S. J. Thorpe. Unsupervised learning of visual features through spike timing dependent plasticity. PLOS Computational Biology, 3(2):e31, 2007.
  6. [6]Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zachary DeVito, Martin Rai son, Aly Khan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024–8035, 2019.arxiv.org/abs/1912.01703
  7. [7]R. P. N. Rao and D. H. Ballard. Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2:79–87, 1999.pubmed.ncbi.nlm.nih.gov/10195184
  8. [8]Sara Steegen, Francis Tuerlinckx, Andrew Gelman, and Wolf Vanpaemel. Increasing transparency through a multiverse analysis. Perspectives on Psychological Science, 11(5):702–712, 2016. doi: 10.1177/1745691616658637.DOI
  9. [9]Anna Truzzi and Rhodri Cusack. Can visual object representations in the human brain be modelled by untrained convolutional neural networks with random weights? In Proceedings of the 42nd Annual Meeting of the Cognitive Science Society, 2020. URL https://cognitivesciencesociety.org/cogsci20/papers/0691/index.html.
  10. [10]Alexander Walther, Hamed Nili, Naveed Ejaz, Arjen Alink, Nikolaus Kriegeskorte, and Jörn Diedrichsen. Reliability of dissimilarity measures for multi-voxel pattern analysis. NeuroImage, 137:188–200, 2016. doi: 10.1016/j.neuroimage.2015.12.012.DOI
  11. [11]J. C. R. Whittington and R. Bogacz. An approximation of the error backpropagation algorithm in a predictive coding network with local Hebbian synaptic plasticity. Neural Computation, 29:1229–1262, 2017.DOI

Paper details

vlnc:W_k6m9ydwq
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Agricultural and Biological Sciences

Claude (Anthropic)

View statement
Looking for related papers…
Contents