acceptodds
Under review as a conference paper at ICLR 2027

BrainPFN: Amortized Brain-Behaviour Prediction via Prior-Data Fitted Networks

12 views

Abstract

The standard approach to predicting behavioral phenotypes from functional MRI is represented by Kernel Ridge Regression (KRR) applied on functional connectivity (FC) matrices. However, this approach is limited: separate models must be fit and tuned for every score and cohort, they do not transfer across datasets, and accuracy degrades sharply when only a handful of subjects are available, as is typical of clinical or cognitive studies. Despite an increasing interest in developing brain foundation models, none has consistently outperformed KRR on behaviour prediction; most reported gains come from easier targets such as age, sex, or diagnosis rather than continuous behavioural scores. We introduce BrainPFN, a Prior Data Fitted Network for brain behaviour regression. BrainPFN is meta-trained on a domain-specific prior built from  7k real FC matrices, from which we sample one billion synthetic training samples. At inference, a labelled context is simply passed through the network in a single forward pass, with no per-score or per-cohort refitting required. Across four independent cohorts, BrainPFN beats KRR on small sample prediction, and leads recent foundation models pre-trained at four to nine times larger scale. Our meta-training pool, aggregated from open repositories, is a fraction of the size of the corpora behind competing models, most of which sit behind restrictive data use agreements such as UK Biobank. We release this pool publicly alongside code and checkpoints, showing that a modest, openly available pool can match or beat models trained at Biobank scale.

Primary area:
applications to neuroscience & cognitive science
Keywords:
neuroscience, neuroimaging, machine learning, brain-behavior prediction, foundation model

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_cTnkOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

BrainPFN trains a prior-data fitted network for brain–behaviour prediction using synthetic regression tasks defined over real, unlabeled functional-connectivity data. Connectomes are transformed into a feature space, kernel functions generate synthetic targets, and signal calibration controls task difficulty. A transformer then predicts new phenotypes from labeled context examples without task-specific fine-tuning. The paper evaluates transfer across cohorts and context sizes, compares with regression and pretrained models, and studies task-generation choices. Its main practical benefit is in small-context prediction; larger-context results also show the continued competitiveness of ridge regression.

Critical strengths
  1. The task generator addresses an important obstacle to pretraining across neuroimaging datasets. Using real connectomes with synthetic targets avoids requiring a common set of behavioral labels while retaining empirical input structure. The kernel family, coefficient distribution, and task difficulty define a concrete reusable learning scheme. This is a substantive methodological contribution independent of whether the prior has a direct biological interpretation.

  2. The evaluation addresses the intended small-context use case. Testing across independent cohorts and a range of context sizes assesses whether a frozen model transfers to new prediction problems. The ablations of kernel choice, coefficient sampling, and task difficulty examine consequential parts of the proposed generator. The concentration of benefits at small context sizes gives the contribution a useful, identifiable scope.

Critical weaknesses
  1. The spectral explanation does not clearly distinguish the proposed prior from the stated alternative. For a fixed positive-definite kernel matrix KK, Equation 6 gives the raw context signal f=Kαf=K\alpha covariance proportional to KK. If the no-RKHS alternative samples coefficients isotropically, its raw signal covariance is instead proportional to K2K^2, which concentrates relative variance even more strongly in large-eigenvalue directions. Thus the argument about favoring smooth directions does not establish that this alternative produces rougher targets. This calculation concerns the signal before task calibration and does not invalidate Equation 6. However, the actual alternative sampling procedure and subsequent calibration need to be specified clearly enough to explain what the ablation demonstrates about the chosen prior.

  2. The claimed exact GP counterpart is not established for the calibrated generator. Section 3.2 standardizes each sampled function using a separate calibration block. Although that block is disjoint, its scale still depends on the sampled function, so this transformation generally changes the original Gaussian function distribution. It defines a valid implicit prior, and the log-loss principle in Equations 1–4 remains applicable. The issue is the stronger statement in Appendix C.2 that the standard GP is an exact analytical counterpart of the training tasks. That correspondence needs justification under the actual calibration, rather than only for an uncalibrated Gaussian construction. The description of a linear regression head in Section 4.2 also leaves unclear which predictive density or posterior summary is represented. These distinctions affect the interpretation of the Bayesian-computation claim and GP comparison.

Questions for authors
  1. What property of the coefficient prior explains the no-RKHS ablation? Section 3.2 motivates Equation 6 by favoring high-eigenvalue, smooth directions. For a fixed positive-definite kernel matrix KK and the raw context signal f=Kαf=K\alpha, Equation 6 gives Cov⁡(f)=σα2K\operatorname{Cov}(f)=\sigma_\alpha^2 K. If the isotropic alternative instead samples α∼N(0,τ2I)\alpha\sim\mathcal{N}(0,\tau^2 I), its signal covariance is τ2K2\tau^2 K^2, placing relatively more variance in the large-eigenvalue directions. Thus this spectral argument alone does not establish that the alternative produces rougher targets. What distribution is actually used in the no-RKHS ablation, and do both samplers undergo the same function-scale and task-difficulty calibration? Please explain which difference between the resulting function priors the ablation isolates. In particular, distinguish evidence that Equation 6 improves prediction from evidence for the proposed smoothness explanation; the covariance calculation above applies before function-dependent calibration.

  2. What posterior quantity does BrainPFN approximate, and is the GP baseline exact for the calibrated generator? Section 3.2 divides each sampled function by a scale estimated on a third subject block. The subjects are disjoint, but the scale still depends on the sampled coefficients. As described, this generally changes the Gaussian function distribution, even at fixed kernel hyperparameters. Appendix C.2 nevertheless identifies the task posterior with a standard GP posterior and its hyperparameter weights with GP marginal likelihoods. Please show how the function-dependent scale is accounted for in this correspondence, or qualify the GP as a related analytical baseline. If the intended equivalence concerns only the posterior mean of the uncalibrated finite kernel expansion, rather than the full predictive distribution, please state that scope and explain whether it survives calibration. Finally, does the linear regression head in Section 4.2 parameterize the predictive density in Equation 4 or output only a point estimate? Please specify the output, implemented training loss, and posterior quantity being approximated so that the Bayesian interpretation can be assessed against the actual generator.

4 comments

We thank the reviewer for the careful and rigorous reading. To answer the 2 concerns in short: the reviewer's calculation is correct, so we withdrew the smoothness explanation and revised Sec. 3.2 and 5.4 (the no-RKHS ablation shows that Eq. 6 improves prediction, not why); and the GP is not the exact counterpart of the calibrated generator, so Appendix C.2 now presents it as a matched analytical baseline, which BrainPFN still outperforms.

1. Spectral explanation and RKHS prior

We thank the reviewer for this careful reading. The calculation is correct, and the argument we gave in Section 3.2 was a mistake on our side. The no-RKHS ablation shows that Eq. 6 improves prediction, and it does not support the smoothness explanation, which we fixed. We revised two sections (in green): Sec. 3.2 now states both priors and the covariance each one gives, and Sec. 5.4 now specifies the isotropic sampler, which shares the kernels and the calibration of Eq. 6.

What was wrong in our argument. With f=Kαf = K\alpha, Eq. 6 gives Cov(f)=σα2K\mathrm{Cov}(f) = \sigma_\alpha^2 K, and isotropic coefficients α∼N(0,τ2I)\alpha \sim \mathcal{N}(0, \tau^2 I) give Cov(f)=τ2K2\mathrm{Cov}(f) = \tau^2 K^2. As the reviewer points out, the isotropic alternative places relatively more variance in the large-eigenvalue directions, so it does not produce rougher targets. The same holds for the RKHS norm: at equal signal variance, the expected norm of the isotropic draws is smaller than or equal to that of Eq. 6, since (∑jλj)2≤rank(K)∑jλj2(\sum_j \lambda_j)^2 \le \mathrm{rank}(K) \sum_j \lambda_j^2. We removed the statement that Eq. 6 avoids "arbitrarily rough targets".

What we meant. Eq. 6 samples the coefficients according to the RKHS norm. It is the Gaussian prior p(α)∝exp⁡(−α⊤Kα/2σα2)p(\alpha) \propto \exp\left(-\alpha^\top K \alpha / 2\sigma_\alpha^2\right) , where α⊤Kα=∥f∥H2\alpha^\top K \alpha = \Vert f\Vert_{\mathcal{H}}^2 is the penalty of kernel ridge regression. Writing K=∑jλjvjvj⊤K = \sum_j \lambda_j v_j v_j^\top and f^j=vj⊤f\hat f_j = v_j^\top f , this norm is ∥f∥H2=∑jf^j2/λj\Vert f\Vert_{\mathcal{H}}^2 = \sum_j \hat f_j^2 / \lambda_j . It penalises the amplitude of ff more along directions with smaller eigenvalues, and accordingly the sampled functions have variance σα2λj\sigma_\alpha^2 \lambda_j along vjv_j. Our intent was to state that the coefficients are sampled under the same penalty that kernel ridge applies to them. Our explanation then mixed up two statements. The norm penalises the amplitude of ff along the directions with smaller eigenvalues, usually read as the rougher ones. We took this to mean that coefficients drawn without the norm give rough targets. This does not follow: the norm acts on ff, while the isotropic prior is flat on α\alpha.

What is used in the no-RKHS ablation. The ablation draws α∼N(0,I)\alpha \sim \mathcal{N}(0, I) and changes nothing else. Both samplers use the same real FC matrices, the same kernel family and bandwidth distribution, and the same calibration: the signal ff is standardised with the mean and standard deviation measured on a separate block of subjects, and noise is then added to reach the target correlation rr. Because of this standardisation the scales σα\sigma_\alpha and τ\tau drop out. The covariance calculation applies before this calibration, which is the same for both priors. The two priors therefore differ only in how the signal variance is distributed over the eigen-directions of KK before calibration: proportional to λj\lambda_j under Eq. 6 and to λj2\lambda_j^2 under the isotropic prior.

What the ablation isolates. It compares training on functions with covariance KK (before calibration) against training on functions with covariance K2K^2, for the same kernel. Before calibration and for a fixed kernel, the Bayes-optimal in-context predictor under Eq. 6 is kernel ridge with that same kernel: for noise of variance σ2\sigma^2, it keeps a fraction λj/(λj+σ2/σα2)\lambda_j / (\lambda_j + \sigma^2/\sigma_\alpha^2) of the label component along vjv_j. Under the isotropic prior this fraction becomes λj2/(λj2+σ2/τ2)\lambda_j^2 / (\lambda_j^2 + \sigma^2/\tau^2), which corresponds to a kernel ridge regressor with K2K^2 as the kernel. After calibration the tasks are no longer exactly Gaussian (App. C.2), so these two filters describe the priors up to the calibration shift and scale. The ablation therefore removes the agreement between the prior over targets and the RKHS norm of the kernel that generates them.

What the ablation shows. It shows evidence that Eq. 6 improves prediction: removing it gives the largest drop of all ablations (Sec. 5.4). It does not show evidence for the smoothness explanation, which was incorrect. The revised Sec. 5.4 reports this result and presents the following explanation as a hypothesis: isotropic targets concentrate on a few leading directions of KK, so a model trained on them learns to treat the weaker directions as noise, while real behavioural scores may also depend on those directions.

2. GP counterpart for the calibrated generator:

We agree. The rescaling factor depends on the sampled function, so the calibrated tasks are not Gaussian and the standard GP is not their exact posterior. The calibrated generator still defines a valid implicit prior, and the log-loss principle of Equations 1 to 4 applies to it, so the Bayesian interpretation of BrainPFN holds with respect to that prior. The revised Appendix C.2 presents the GP as a matched analytical baseline, without the claim of exactness.

Output, loss and target. The linear head returns one point estimate per query subject and is trained with the squared error, which is Equation 4 with a fixed-variance Gaussian. BrainPFN is therefore trained to approximate the posterior predictive mean under the calibrated prior. Section 4.2 and Appendix B.1 will state this.

Does the comparison survive calibration? The exact posterior mean of the calibrated tasks has no closed form and would need Monte Carlo integration for every prediction. We therefore compare with the closest tractable model, a GP that also models the calibration effect: the signal is stronger on the context than on new subjects, by a factor computed from unlabelled features ("GP, training signal model" below). Only the task-to-task variation of this factor is left out. It acts like a spread in the signal-to-noise ratio. GP-BMA, which averages over a range of signal-to-noise ratios, scores at the level of the GP with a fixed ratio ("GP, training hyperprior"), so this approximation is unlikely to change the comparison. BrainPFN leads this GP on every target set (p < 0.05).

(Log-AUC of Pearson r over N = 5 to 800, mean over 5 folds)

ModelEffect size rCalibration effectHCP-YA, all targetsHCP-YA, cognitionHCP-Aging, all targets
BrainPFN0.5 in trainingIn the training tasks0.1370.3250.333
GP-BMA (Appendix C.2)Averaged over 0.2 to 0.8Not modelled0.1270.3050.319
GP, training hyperprior0.5Not modelled0.1270.3040.319
GP, training signal model0.5Modelled0.1270.3070.320

Thank you for clarifying both the prior comparison and the GP interpretation. Your response addresses both of my concerns. I do not request any additional experiments and I will maintain my score.

Dear reviewer cTnk,

Thank you for acknowledging our response. As all of your concerns have been addressed, would you consider raising your score to a full acceptance?

Reviewer_RXY1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces BrainPFN, a brain-behavior prediction network based on prior-data fitted networks (PFN). The authors generate synthetic tasks by sampling from real resting-state FC data and from the kernel ridge regression (KRR) function space. The authors use these tasks to train a Transformer based on PFN for achieving one-shot forward predictions in small-sample settings (N<100N < 100). The scientific question is good. However, in my understanding, this paper is not a ture "brain foundation models", but an an amortized KRR solver. The following are my concerns.

Critical strengths
  1. Small sample size has always been a challenging issue in clinical practice, and this paper is very good from this perspective.
  2. The idea of amortized Inference is interesting.
Critical weaknesses
  1. I noticed that the experimental results (as shown in Figure 2 and Figure 4) show that when the sample size $N>150-200 $, the traditional Ridge Regression gradually equalizes or even surpasses BrainPFN. This indicates that PFN has encountered bottlenecks in processing slightly larger-scale data (such as HCP-YA and HCP-A) due to limitations in capacity saturation or in context learning. Can you please explain?
  2. The prior task of BrainPFN is to generate samples from the KRR function space (linear, RBF, Laplace kernel) (Eq. 5). Essentially, the network is trained as a fast kernel ridge regression approximator capable of performing Bayesian model averaging (BMA) (as mentioned in Appendix C.2, where BrainPFN performs slightly better than precise Gaussian processes). So, has BrainPFN only learned how to perfectly fit KRR in one second, without learning deeper and more complex nonlinear brain behavior mapping mechanisms than traditional kernel methods?
  3. All evaluations are only based on Schaefer-400 spectra, didn't test on different atlases. This is very important validation for the generalization of your method.
  4. BrainPFN trained with resting-state data can also predict well on task-state FC (HCP-YA Working Memory), this phenomenon is interesting but lacks a neuroscience explanation.
Questions for authors
  1. If the input is the original FC matrix (didn't conduct dimensionality reduction operation), or FC based on other atlases, can the method achieve the same performance?
  2. Since the covariance structure between the resting state and the task state is significantly different? Why can the prior sampled from the resting-state pool generalize to the distribution shift of the task state?
5 comments

We thank the reviewer for the thoughtful feedback. To answer the 4 concerns in short: the gap with Ridge at large N comes from the pretraining context size (a model pretrained on a context of 2,000 subjects is above Ridge at every N on HCP-YA); we provide evidence that BrainPFN regressors are not kernel ridge functions from the families used to train it; BrainPFN gains hold across atlases; finally, regarding the foundation model remark, we consider BrainPFN a foundation model because it predicts any out-of-scanner measure without refitting, with amortization as the underlying mechanism.

1. BrainPFN scaling to large contexts

We thank the reviewer for this question. Our main focus was the small-sample regime, so we did not put much effort into large sample sizes, where Ridge is competitive. Still, we investigated the reason for this: the gap at larger sample sizes actually comes from the pretraining setup, not from a capacity or in-context learning limit. The model in the paper was pretrained on synthetic tasks with 200 subject contexts, which limited how well it could use larger samples. We therefore pretrained BrainPFN on tasks with 2000 subject contexts. The resulting model is above Ridge at every context size. We added these results to the paper (Sec. C.6). Below we report the results for HCP-YA.

HCP-YA (grand average)

ModelN=5N=10N=20N=50N=100N=200N=400N=600N=800log-AULC
BrainPFN paper0.036*0.056*0.077*0.118*0.146*0.1740.2130.2310.2430.130 ± 0.008
BrainPFN N=20000.036*0.057*0.079*0.123*0.154*0.192*0.225*0.2430.2550.138 ± 0.009
Ridge0.0260.0410.0570.0950.1280.1750.2160.2360.2500.119 ± 0.009

* denotes p<0.05p<0.05 . For this N=2000 model, the synthetic tasks use a linear kernel because it is much cheaper to run. Drawing GP targets with the Laplace kernel requires factorizing the full N x N kernel matrix, which scales cubically with N and is about 25x slower on our machine. The linear kernel has rank at most d (the number of features), so its targets can be drawn at O(Nd2)O(Nd^2) . The N=2000 model pretrained with the Laplace kernel will be included in the camera-ready version.

2. Is BrainPFN only a fast KRR approximator?

Thank you for this very interesting question. We find that BrainPFN does not behave as the kernel ridge of its prior. We tested this directly in a new section, Appendix C.3 (in green in the revised PDF). We asked whether BrainPFN behaves like a kernel ridge predictor on HCP contexts of 100 subjects, using two properties of a kernel ridge. Given a context of FC matrices XX with labels yy and a kernel kk, let f(x)f(x) be the kernel ridge prediction for a new FC matrix xx. The two properties are:

  • Linear in the labels. Kernel Ridge is linear in yy, so its weights on the labels J=∂f/∂yJ=\partial f/\partial y depend only on the FC matrices: ∂J/∂y=0\partial J/\partial y = 0. We tested whether this is also 0 for BrainPFN, by measuring how much JJ changes when the target moves from one behavioral score to another (details in Appendix C.3). A value different from 0 would indicate that BrainPFN does not behave like kernel ridge.

  • Shape set by the kernel. the regressor is a sum of kernel bumps k(x,xi)k(x,x_i) at the context subjects, then the kernel determines its shape in FC space. Along the line x(s)x(s) from a context subject xix_i (s=0s=0) to its nearest neighbour (s=1s=1), we measured how sharply the prediction bends at the context subject, i.e. its change of slope at s=0s=0 (the kink). For linear ridge the shape is a straight line along x(s)x(s), for Laplace ridge it has a sharp corner at s=0s=0. For BrainPFN the shape is different from two previous ones.

We tested two BrainPFN models: the one with Laplace-kernel prior (from the main text), and the large-context model of Appendix C.6 (linear-kernel prior). The label weights JJ of the Laplace-prior network change by 25 to 41% between pairs of scores, and those of the linear-prior network by 16 to 21%. The kink of the Laplace-prior network is 0.25 times that of a Laplace ridge. We also computed the difference in shape between each network and the closest kernel ridge of its prior's family (Laplace ridges over 7 bandwidths and 6 penalties, or the linear ridge). It is 0.40 for the Laplace-prior network and 0.25 for the linear-prior network, whereas each network differs from itself by only 0.046 and 0.025 on another context. These networks are therefore not kernel ridges with a fixed kernel and penalty. We found that this departure comes from the attention layers, which can change the label weights: when we keep the attention constant, 92 to 94% of the departure in shape is removed.

3. Generalization to other atlases

We re-trained the BrainPFN on three additional atlases: Yan with 400 parcels (Yan et al., NeuroImage 2023) which is a homotopic atlas based on Schaefer, and BASC with 444 and 325 parcels (Bellec et al., NeuroImage 2010) which is a different family. BrainPFN generalizes across all atlases. We report below a summary in terms of log-AULC, and we revised the text by adding a new section in appendix C.7 with these results.

AtlasHCP-YA BrainPFNHCP-YA RidgeHCP-Aging BrainPFNHCP-Aging Ridge
Schaefer-400 17N (paper)0.130 ± 0.0080.119 ± 0.0090.333 ± 0.0120.310 ± 0.010
BASC 4440.122 ± 0.0070.115 ± 0.0040.320 ± 0.0120.306 ± 0.009
BASC 3250.125 ± 0.0090.115 ± 0.0060.334 ± 0.0100.308 ± 0.005
Yan-400 7N0.115 ± 0.0050.111 ± 0.0060.318 ± 0.0140.307 ± 0.008

4. Generalization to task-fMRI

We thank the reviewer for this observation, which also relates to their second question on the distribution shift between resting-state and task data.

We hypothesize that resting-state connectivity provides a good representation space for task-related brain activity and behavior. Resting-state FC captures the intrinsic functional organization of the brain, which is largely preserved during task performance. This hypothesis is consistent with a growing body of empirical evidence. Early work showed that task-free MRI predicts individual differences in task-evoked activity (Tavor et al., Science, 2016). Subsequently, representations built on resting-state networks were shown to transfer across task studies and improve decoding (Mensch et al., PLOS Computational Biology, 2021). More recently, Brain-Semantoks (Gijsen et al., ICLR 2026), a foundation model pretrained on rs-fMRI, was also shown to generalize to task-based fMRI (decoding block type in the Hariri emotion task). Together, this evidence suggests that transfer from rest to task is a general property of resting-state representations rather than a peculiarity of BrainPFN. Under this view, BrainPFN does not need to model the task-state covariance exactly, because the resting-state feature space remains informative about task-state phenotypes. We leave a more detailed neuroimaging investigation of this mechanism for future works. We have added a discussion of this hypothesis, with these references, to the revised manuscript.

Thank you to the authors for the detailed rebuttal and additional experiments. While these revisions strengthen the empirical results, my fundamental conceptual concern remains, i.e., it is still debatable whether this framework represents a true "brain foundation model". Therefore, I will maintain my current score.

Reviewer_JSUKOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The manuscript proposed BrainPFN, a method inspired by TabPFN, pre-trained on over 7K FC matrices. The method shows improvements on datasets for multiple tasks on small sample sizes. In large sample size, the method is comparable or better in some tasks.

Critical strengths
  • The experiments performed on multiple datasets and tasks, and compared with multiple baselines and ablations.
  • Overall, I am excited about this direction and found it novel and interesting.
Critical weaknesses
  • The proposed approach is outperformed by a simple baseline at higher sample size as Ridge Regression, which seems a downside. Hence, the method does not scale well, and benefit early just due to pre-training and large set of seen samples. This issue have to be addressed.
  • There is no interpretability analysis which is important part of brain analysis. It is not clear whether the performance improvement actually benefit the brain data analysis, and it is not clear whether the model actually can get robust insights from the brain data at low sample size.
  • No statistical analysis on the performance. In addition, at low sample size some results are within the spread, so can not be treated as better.
Questions for authors

See weaknesses

3 comments

We thank the reviewer for the positive and constructive feedback. To answer the 3 concerns in short: the gap with Ridge at large N comes from the pretraining context size (a model pretrained on a context of 2000 subjects is above Ridge at every N on HCP-YA); the statistical analysis was already in the paper but not clearly explained (asterisks mark p < 0.05 against Ridge) and is now stated explicitly in Sec. 4.2; finally, the interpretability analysis is still ongoing, and we will post the results as soon as they are ready.

1. Scaling of BrainPFN

We thank the reviewer for this observation. Our main focus was the small-sample regime, so we did not put much effort initially into large sample sizes, where Ridge is competitive. Still, the submitted model was pre-trained on synthetic tasks with 200 context subjects, which we discovered that ultimately limited how well it could use larger samples. Following the reviewer's comment, we pretrained BrainPFN on tasks with 2000 context subjects. The resulting model is above Ridge at every context size. Pretraining on larger contexts unlocked BrainPFN scaling with sample size. We added these results to the paper (Sec. C.6). Below we report the results for HCP-YA.

HCP-YA (grand average)

ModelN=5N=10N=20N=50N=100N=200N=400N=600N=800log-AULC
BrainPFN paper0.036*0.056*0.077*0.118*0.146*0.1740.2130.2310.2430.130 ± 0.008
BrainPFN N=20000.036*0.057*0.079*0.123*0.154*0.192*0.225*0.2430.2550.138 ± 0.009
Ridge0.0260.0410.0570.0950.1280.1750.2160.2360.2500.119 ± 0.009

* denotes p<0.05p<0.05 . For this N=2000 model, the synthetic tasks use a linear kernel because it is much cheaper to run. Drawing GP targets with the Laplace kernel requires factorizing the full N x N kernel matrix, which scales cubically with N and is about 25x slower on our machine. The linear kernel has rank at most d (the number of features), so its targets can be drawn at O(Nd2)O(Nd^2) . The N=2000 model pretrained with the Laplace kernel will be included in the camera-ready version.

3. Statistical analysis of the performance

We thank the reviewer for this comment, and we apologize: the statistical analysis is in fact in the paper, but it was not stated clearly enough. The asterisks in the figures mark p<0.05p < 0.05 against Ridge throughout (paired test over the 5 folds, Holm-corrected across context sizes), and the log-AULC summaries use the same test and threshold. Where BrainPFN's advantage falls within the spread, the test is not significant and no asterisk is shown, so we do not claim an improvement there. The significant results are at small N on all four cohorts, and in log-AULC on the cognition target of all four cohorts. We chose Ridge as the reference because the other foundation models fall below it most of the time. We stated the test and the p<0.05p < 0.05 threshold explicitly in the revised text in “Evaluation protocol” in Sec. 4.2.

2. Interpretability of BrainPFN

Thank you for raising this. We added two interpretability analyses (new Appendix C.9, in green). They ask whether BrainPFN's accuracy gain over Ridge comes with different features of the connectome. In both, the two models make different predictions, both close to the original target, and recover the same connections. For each model we compute a feature map: which functional connections covary with its predictions (Haufe et al., 2014). We compare BrainPFN with Ridge on the same subjects, against maps fitted on shuffled labels.

  • Both models recover the same connections on synthetic and real scores. To validate the maps on fully controlled examples where we know which features are actually relevant we built synthetic scores from the real connectivity of 953 HCP subjects. Each depends on a few known groups of between-network connections, additively or non-linearly, with noise at real effect sizes. With 10 to 100 labelled subjects the predictions of the two models correlate at 0.82 to 0.90 (Fig. 19a). Both recover the known connections (Fig. 19b); the pre-specified test finds no significant difference in any condition (Holm p ≥ 0.29). The same holds on two real HCP measures: for Flanker, recovery is 0.24 for BrainPFN and 0.19 for Ridge at 25 subjects, and 0.39 and 0.36 at 100.

  • Projecting the maps on the cortex Here we focus on the picture vocabulary score (picture naming) as an illustrative example. With 500-subject context, BrainPFN and Ridge reach similar accuracy (r = 0.32 and 0.33), their maps correlate at 0.96, and 7 of their 10 strongest connections are shared (Fig. 20). We compared the maps with NeuroQuery meta-analytic maps (Dockès et al., 2020) of 440 Cognitive Atlas terms (Poldrack et al., 2011). The five terms that best match the map are the same for BrainPFN and Ridge: cognitive control, executive function, rule, problem solving and goal maintenance (Fig. 20, right). The strongest connections link the dorsal and ventral attention networks with each other or with the control network, in line with reports on crystallised ability (Lombardo & Kaufmann, 2023; Dhamala et al., 2021), which is also among the functions probed by picture naming (Akshoomoff et al., 2014).

At low sample size, the benefit of BrainPFN is therefore in predictive accuracy, and it comes at no cost in interpretability: the connections it identifies are those that Ridge identifies. Whether higher accuracy will translate into new insights on brain data remains an open question. At the effect sizes of current brain-behaviour prediction, feature maps are likely limited by the signal in the data, for any model. We see more accurate few-shot predictors such as BrainPFN as a necessary step towards that goal.

Reviewer_JSUKOpenReview ↗

Rebuttal Response

Thank you for addressing my concerns on performance and interpretability. I have increased the score to clear accept.

Comments on the paper

1 comment

We thank all reviewers for their appreciation of the work and for their interesting questions, which made the paper stronger. We are glad that the direction was found "novel and interesting" (JSUK), the small-sample problem relevant for clinical practice (RXY1), and the task generator "a substantive methodological contribution" (cTnk). All changes are in green in the revised PDF. The key highlights:

  • BrainPFN now scales with sample size (JSUK, RXY1). The gap with Ridge at large N came from the pretraining context size, not from a capacity limit. Pretrained on contexts of 2,000 subjects, BrainPFN is above Ridge at every N from 5 to 800 on HCP-YA (App. C.6).

  • The gains hold across atlases (RXY1). Retrained on three additional atlases (Yan-400, BASC-444, BASC-325), BrainPFN is above Ridge in log-AULC on HCP-YA and HCP-Aging (App. C.7).

  • BrainPFN goes beyond kernel ridge (RXY1). Its label weights change by 16 to 41% between behavioural scores, whereas those of a kernel ridge regressor do not change at all, and its shape in FC space matches no kernel ridge of its prior's family (App. C.3).

  • Higher accuracy at no cost in interpretability (JSUK). BrainPFN and Ridge recover the same connections, on synthetic scores with known ground truth and on real HCP measures (App. C.9).

  • Corrected theoretical claims (cTnk). We fixed the smoothness explanation of the RKHS prior and now present the GP as a matched analytical baseline, not an exact counterpart. BrainPFN still outperforms a GP that models the calibration, with p < 0.05 on every target set (Sec. 3.2, Sec. 5.4, App. C.2).

  • Statistical testing stated explicitly (JSUK). Asterisks mark p < 0.05 against Ridge (paired test over the 5 folds, Holm-corrected); we claim an improvement only where the test is significant (Sec. 4.2).

We believe the revision addresses the main concerns raised, and we are happy to answer any further questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.