acceptodds
Under review as a conference paper at ICLR 2027

When Good Brain Predictors Make Bad Brain Controllers: A Regression Theory of Prediction–Control Dissociation

Abstract

Deep neural networks are increasingly used both as encoding models of visual cortex and as controllers that synthesize images to drive neural responses toward desired states. Recent in vivo experiments reveal a striking dissociation: models with nearly identical held-out accuracy can generate qualitatively different stimuli and differ sharply in control, and these differences track input-gradient structure better than predictive accuracy. We develop a geometric and analytical account of this phenomenon in a linear teacher–student model. The key is that control creates a self-generated distribution shift. Natural-image generalization, self-accentuation, and peer review evaluate the same fitted predictor under three different input distributions; correspondingly, their equal-error sets in weight space are ellipsoids, spheres, and hyperplanes. Using deterministic equivalents for ridge regression, we show that prediction-optimal and control-optimal regularization satisfy different spectral conditions, so cross-validation can leave large control error when responses are noisy and inputs are anisotropic. Extending the analysis to fixed linear feature maps, we construct a prediction-preserving gauge family with identical generalization and cross-validated regularization but arbitrarily different control. Control is most fragile when feature directions that survive regularization backpropagate into low-variance, high-spatial-frequency image directions. The resulting gradient-geometry factor extends to nonlinear networks as a Jacobian-based diagnostic and predicts across-model variation in in vivo control better than held-out prediction error. Thus, predictive accuracy alone cannot certify control; the model’s gradient geometry or generated stimuli must also be evaluated.

Primary area:
applications to neuroscience & cognitive science
Keywords:
regression theory, ridge, random matrix theory, OOD generalization, visual neuroscience, foundation model, digital twin, causal intervention

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_fGdAOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper starts from a puzzling experimental observation concerning in-vivo neural control via presented (generated) visual stimuli: models that scored favourable on held-out neural predictivity were not automatically good model for neural control (i.e. via model-based control obtained by gradient ascent on the model's prediction), and viceversa; moreover even poor-controller models were generally good at estimating the efficacy of synthesised control-images (i.e. good "peer-review" predictivity). In this work the author propose a linear theory to mechanistically explain this puzzling behaviour: in a simplified teacher-student framework they show how this asymmetry fundamentally stems from the fact that control evaluates models on a different distribution than what predictivity measures do. Fundamentally, prediction is weighted by the data covariance matrix, while control puts equal weight on all weight-direction, so control fails when control optimisation steps into low-variability data directions. The author offer a precise geometrical picture for this fact and validate their predictions via Monte Carlo estimates. Finally, the authors apply their linear theory to non-linear network by considering the Jacobian-linearization and showing that their geometrically-derived factor yields a better predictor for the model performance as controller then the corresponding held-out neural predictivity.

Critical strengths

In this work the authors have tackled an initially puzzling experimental finding and proceeded to build a methodical mechanistic explanation for the phenomenon in a simplified linear theory which yielded an intuitive and satisfying explanation: prediction and control are evaluated on different distributions, the latter being generated by the controller itself. The geometrical intuition behind this asymmetry (conveyed in Fig.2) is particularly helpful. The author then support their claim via extensive Monte Carlo simulations and an explicit construction that recreates the same phenomenon in the studied teacher-student setting. The final extension to the non-linear network case is also appreciated, where the theory-inform geometrical factor provides a better (although not perfect) predictor of downstream controller performance than the standard held-out neural predictivity. This last piece might provide especially useful to the experimental practitioners for designing the next neural control-based experiments and informing the model-selection process.

Critical weaknesses

Minor points:

  • Figure 1 is hard to read due to the small fonts and high visual complexity. Perhaps the author can consider simplifying the design to focus and convey the most important aspects (panel B,C) and moving the rest to the appendix.
Questions for authors

Some curiosity-based questions for the authors:

  1. What are the author recommendations for obtaining reliable neural controller? Can those theory-derived insights be applied to training a model specifically for control?
  2. Adversarially trained models tend to be the best neural controller. Do the authors think this is because adversarial training tends to probe-and-exploit low-variance directions, thus implicitly regularising the data covariance to be "more spherical", with in turns reduces the risk of mis-control? Or is there some other mechanism involved?
3 comments
AuthorsOpenReview ↗

Response to `fGdA`

We appreciate reviewer fGdA for their detailed reading, insightful comment, and positive evaluation of our work!

W1:

Thanks for the helpful comment! We are remaking Figure 1 to break it down and make it easier to read.

Q1: Recommendations for obtaining reliable neural controller

  1. Training side: Our current insight is to directly regularize or supervise the model’s gradients so that they have a more naturalistic structure. Empirically, adversarial training is an effective way to encourage natural-image-like gradients. In practice, adversarial fine-tuning offers a good recipe: start from a strong foundation model (CNN or transformer) and then make its gradients more perceptually aligned. Among our ten models, CLIPAG is the result of this kind of adversarial fine-tuning, and it performed well on neural control, as did ResNet50-robust (which was adversarially trained from scratch).

    One interesting future direction is to find alternative training method other than adversarial training: is directly regularizing frequency spectrum of input gradient a competitive alternative?

  2. Generation side: According to the theory, the failure of control is a form of out-of-distribution (OOD) generalization failure. If a stronger prior is supplied during the image generation process (e.g., diffusion models, GANs, etc. Pierzchlewicz et al. 2023; Wang & Ponce 2026), we expect the predictive model’s judgements to be better calibrated and the control to be more successful. We think this will be especially helpful for models that were not robustly trained.


  1. Pierzchlewicz, P. A., et al. (2023). Energy Guided Diffusion for Generating Neurally Exciting Images. Advances in Neural Information Processing Systems, 36, 32574–32601.
  2. Wang, B., & Ponce, C. R. (2026). Neuronal tuning aligns dynamically with object and texture manifolds across the visual hierarchy. Nature Neuroscience, 29, 864–875.

Q2: Cause of Adv. Robust models performing better

For adversarially trained models, they tend to have gradients more aligned with low freq / high variance dimension of data — as we can see the gradients from adv. trained models generally have better spatial frequency profiles (decay like natural images), and visually they looks more like shapes instead of meaningless noise.

So related to the feature regression theory, when we do accentuation and derive control stimuli from adv robust models, the image sweeps are more aligned with natural data manifold, so we think the models have less out-of-distribution prediction error for their images. Thus on these robust model generated images the robust models have better prediction of neural out come thus control score. Actually, as a matter of fact, all other models have higher neural prediction scores on these robust model generated images — suggesting they may be closer to training distribution.

Dear Reviewer fGdA,

Can you please take a look at the authors' response and provide your further input? Thank you.

AC

I thank the authors for their response, especially for the recommendations on how to improve future controller, both at the training and generation stage. I confirm my score and recommend acceptance for this work.

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

Previous studies used neural response models to design images that drive neuronal activity. This paper asks why this approach can fail despite accurate prediction on natural images.Using simplified linear models, the authors show how image optimization can exploit errors hidden by natural image tests. They find that the best regularization can differ between prediction and stimulus design. They also construct equally predictive models with different design performance. Finally, they propose a gradient geometry diagnostic and report correlations with stimulus design performance in existing neural recordings.

Critical strengths
  1. The feature map construction provides a concrete illustration about optimized input performance that ordinary prediction statistics do not determine.

  2. The numerical experiments and appendices contain substantial validation and sensitivity analyses.

Critical weaknesses
  1. Limited scientific contribution. The paper does not clearly explain what its linear regression analysis adds to prior work on input optimization and response geometry. The biological evidence is descriptive, and the diagnostic’s advantage over existing gradient-based measures remains unclear.

  2. Unclear exposition and substantive mathematical errors. Dense notation obscures the argument. Corollary 19 claims unbounded growth under a fixed finite dimensional positive definite covariance, which imposes a finite bound. An exact equality in the minimum norm interpolation discussion also requires an unstated noiseless assumption. These errors undermine the paper’s central theoretical contribution.

Questions for authors

What additional predictive value does the diagnostic provide over existing gradient spectrum measures?

What evidence beyond the reported correlations supports the claim that “A single mechanism accounts for the three neuronal phenomena”?

6 comments
AuthorsOpenReview ↗

Response to reviewer `SBun` (1/n)

We appreciate SBun's careful examination of our paper, we address the questions and weakness below!

W1: Limited contribution.

We’d like to use this space to better contextualize our scientific contribution with respect to the literature.

Regression theory for neural prediction. There are ample prior works using kernel and ridge regression theory to explain neural prediction and brain-machine alignment (Canatar et al., 2021a; 2023). These works have made contact with the practise of encoding benchmarks such as BrainScore (Yamins et al., 2014; Schrimpf et al., 2018). However, these theory and practise evaluate models on held out samples from the same distribution.

Stimuli optimization and neural control. Recently, there is a growing line of experimental work on stimuli optimization for neural control (Bashivan et al., 2019; Walker et al., 2019; Ponce et al., 2019; Prince et al., 2026), adapting techniques from feature visualization (Olah et al., 2017). These works offered many new phenomena, but they are surprisingly much less theorized. Specifically, it has not been treated within the regression theory framework, and model backbones and visualization methods are chosen largely empirically. In particular, it’s not clear why some networks can predict the neurons well (in classic BrainScore sense) but their accentuated stimuli does not control the neurons well (Prince et al., 2026).

Existing OOD regression theory (Canatar et al., 2021b; Patil et al., 2024) does not cover this case either: the test distribution was fixed in advance, while in “neural control” type experiment, the test is generated by the fitted model itself.

Our Contribution. Our work fill exactly this gap: we adapt Ridge regression theory to study input optimization and neural control. On the theory / technical side, we derive a few new deterministic-equivalent (random-matrix) results for the control error, i.e. the error on the model’s own optimized inputs. We show that it has different geometric and statistical properties comparing to the classic prediction error. On the neuroscience / practical side, the theory explains why some good brain predictors are poor controllers and why held-out prediction alone is not enough to constrain this control error. It also gives theory-grounded rationale to the diagnostic (see Q1 also.).

More broadly, the paper connects two literatures that have developed separately: the regression and random-matrix theory of neural prediction, and feature visualization and neural control. We believe this bridge is valuable to both communities, and we will revise the introduction and related work to make it explicit.

References

  • Bashivan, P., Kar, K., & DiCarlo, J. J. (2019). Neural population control via deep image synthesis. Science
  • Canatar, A., Bordelon, B., & Pehlevan, C. (2021a). Spectral bias and task-model alignment explain generalization in kernel regression and infinitely wide neural networks. Nature Communications
  • Canatar, A., Bordelon, B., & Pehlevan, C. (2021b). Out-of-distribution generalization in kernel regression. NeurIPS
  • Canatar, A., Feather, J., Wakhloo, A., & Chung, S. (2023). A spectral theory of neural prediction and alignment. NeurIPS
  • Olah, C., Mordvintsev, A., & Schubert, L. (2017). Feature visualization. Distill.
  • Patil, P., Du, J.-H., & Tibshirani, R. J. (2024). Optimal ridge regularization for out-of-distribution prediction. ICML 2024
  • Ponce, C. R., et al. (2019). Evolving images for visual neurons using a deep generative network reveals coding principles and neuronal preferences. Cell
  • Prince, J. S., et al. (2026). Parametric neural control differentiates top neural network models of primate visual cortex. bioRxiv.
  • Schrimpf, M., et al. (2018). Brain-Score: Which artificial neural network for object recognition is most brain-like? bioRxiv.
  • Walker, E. Y., et al. (2019). Inception loops discover what excites neurons most using deep predictive models. Nature Neuroscience
  • Yamins, D. L. K., et al. (2014). Performance-optimized hierarchical models predict neural responses in higher visual cortex. PNAS
AuthorsOpenReview ↗

Response to reviewer `SBun` (2/n)

W2: Math typos

We thank the reviewer for checking our theoretical statements so carefully!!

(1) Corollary 19. Indeed, the reviewer is right that “grow without bound” in the Corollary 19 was informal, and it’s not mathematically correct. Given a fixed data covariance Σ≻0\Sigma\succ0, the quantities have bounds and and bound depends on lowest eigenvalue of Σ\Sigma.

To truly “grow without bound”, we need an additional assumption on the data covariance Σ\Sigma. For example, let ϵ\epsilon be the lowest eigenvalue of Σ\Sigma. If we take a sequence of covariances with ϵ→0\epsilon\to0, then TFT_F can be made arbitrarily large while prediction stays unchanged. Similarly, for a power-law spectrum, letting the data dimension d→∞d\to\infty produces arbitrarily small eigenvalues, and hence unbounded growth.

Natural-image covariances have many eigenvalues close to zero, so this regime holds in practice. The missing assumption therefore does not weaken Proposition 18, or our main point that feature gradients loading onto the low-variance eigenspace are dangerous for control. Still, we agree that the assumption on Σ\Sigma should be stated explicitly, and we will do so in the revision.

To make it more rigorous, we clarified the set up for Σ\Sigma and say “without bound” happens only as a limit of λmin⁡→0\lambda_{\min} \to 0 in a remark.

(2) Minimum-norm interpolation. Indeed, the exact equality refers to the noiseless minimum-norm interpolator, not the one through noisy data, and we will state this explicitly. This was a side remark; it is not used in any derivation. We have removed the remark where it is not needed, and we state the noiseless assumption where it is used.

(3) Exposition. We agree the notation is dense. We would be grateful if the reviewer could point to the sections or appendices that were hardest to follow, and we will revise those to explain the argument more clearly.

AuthorsOpenReview ↗

Response to reviewer `SBun` (3/n)

Q1: Additional predictive value

In terms of predicting a model’s control capability, TϕT_\phi (and VϕV_\phi) is comparable to the predictors previously reported in Prince, Wang et al. (e.g., statistics of the spatial frequency of the gradient), and it does not confer additional predictive value:

Predictor10 models9 trained7 conventional
Held-out error (control session)0.060.00−0.01
Adversarial sensitivity (Prince et al.)0.51—0.05
TϕT_\phi, exact-local0.510.620.17
TϕT_\phi, neighborhood (16/255)0.580.610.19
Gradient PR (Prince et al.)0.640.670.26

We reproduced the gradient participation ratio (PR) results using the protocol in Prince, Wang et al’s public code base. Further TϕT_\phi and gradient PR are strongly collinear (r=0.81∼0.89r=0.81\sim0.89), and the variance they explained are the same chunk (To test for incremental value, we regressed PR out of both log⁡10Tϕ\log_{10}T_\phi and the control slope. The partial correlation is negligible: ∣r∣≤0.17|r| \le 0.17 across variants and subsets). So from a purely predictive angle, we agree that TϕT_\phi does not add predictive value on its own.

One possible reason why TϕT_\phi is less predictive is that TϕT_\phi uses less information. TϕT_\phi depends only on the feature-space PCs (sk,vks_k, v_k) of the representation on the training images and on the feature Jacobian JϕJ_\phi. Apart from the ridge level κ\kappa, it ignores the fitted readout weights w^\hat w on top of those PCs. Gradient PR, by contrast, is computed from the input gradient of the specific readout head, Jϕ(x)⊤w^J_\phi(x)^\top \hat w. It therefore weights each feature direction by how much that neuron’s readout actually uses it. This extra, neuron-specific information may account for part of PR’s advantage.

However, we insist there is value in our theory: our intention is that it adds justification and a theoretical framework for thinking about this control error. In this sense, a predictor based on the spatial-frequency structure of the gradient is not ad hoc; it follows from the linear theory we solved. In the feature regression theory, the gradient geometry term is TF=∑k(sksk+κ)2∣gk∣2gk⊤ΣgkT_F=\sum_k\big(\tfrac{s_k}{s_k+\kappa}\big)^2\tfrac{|g_k|^2}{g_k^\top\Sigma g_k}, the ratio of each feature gradient’s Euclidean energy to its energy under the natural-image covariance. Natural images have a power spectrum that falls off steeply with frequency. This ratio is therefore large exactly when the gradient’s power sits at high spatial frequencies, which natural data barely constrain.

The theory also tells us not to expect TFT_F (or TϕT_\phi) to fully explain the control slope. The control error is zacc=D/N−1z_{acc}=D/N-1, with N=β⋆⊤βˉF,κN={\beta^\star}^\top\bar{\beta}_{F,\kappa} and D=∣βˉF,κ∣2+1n(Egen+σ2)TFD=|\bar\beta_{F,\kappa}|^2+\tfrac{1}{n}(E_{gen}+\sigma^2)T_F. So beyond gradient geometry, it also depends on teacher alignment (how well the neuron’s tuning lies in the model’s learnable subspace), and on the noise level (neuron specific) and sample size (fixed). TϕT_\phi is blind to the alignment term by construction. Gradient PR captures it only indirectly, through the fitted readout weights. This may explain why gradient-based predictors remain imperfect, especially among the conventional models. It also suggests where to improve: estimate the alignment term from neural data, rather than further refining just gradient statistics.

Reviewer_hi6ZOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies whether models that are good at predicting visual-cortex activity are equally good when used as controllers, that is, to synthesize images that drive a neuron to target responses and generate testable hypotheses on real neurons. Empirically this often fails: models with nearly identical predictive accuracy differ widely in how well their images control the neuron. To explain this, the authors use a linear teacher–student framework. The key insight is that control evaluates the model on a distribution it generates itself: accentuated images move along the model’s own gradient. Prediction on natural images barely penalizes weight errors in low-variance, high-spatial-frequency directions, but accentuation exposes those errors at full Euclidean strength. They show that response noise, cross-validated regularization (which is not control-optimal), and the feature map’s gradient geometry all amplify this gap, and that equally predictive models can differ arbitrarily in control. A Jacobian-based version of their gradient-geometry measure predicts in-vivo control across ten models where held-out accuracy does not. The conclusion is that predictive accuracy alone cannot certify a good controller; the model’s gradient geometry or its generated stimuli must be validated directly.

Critical strengths

The paper gives an elegant and unifying account of why predictive accuracy does not determine a model’s usefulness as a neural controller, and why gradient geometry does. Particularly, I find very encouraging:

  • Matched prediction with divergent control, recognition without generation, and the link between control and input-gradient spectra all follow from the fact that control evaluates the model on its own gradient path.
  • Rigorous and interpretable theory. Deterministic-equivalent analysis yields closed-form, interpretable results: the norm-inflation term, an explicit residual control error at the cross-validated penalty, and a prediction-preserving gauge family that isolates feature geometry as the cause.
  • A practical, falsifiable diagnostic. T_φ can be computed from the model and image statistics before any experiment. On real macaque data, it ranks models by in-vivo control (r ≈ −0.6) where held-out accuracy does not (r ≈ 0).
  • Thorough validation. Theory is checked against natural-image Monte Carlo, the predicted sign pattern of peer- versus self-review is confirmed in the neural data, and the empirical analyses include robustness checks across model subsets, rank correlations, and alternative endpoints.
Critical weaknesses

The related-work discussion would benefit from citing Gaziv, Lee & DiCarlo (NeurIPS 2023), who showed that robustified ANNs, unlike standard ones, generate low-norm perturbations that strongly and precisely modulate human category percepts. This is behavioral evidence consistent with the paper’s finding that adversarially trained models have the smallest T_φ and are the best neural controllers. Discussing whether T_φ would also predict their human-perception results would strengthen the generality of the claim.

Questions for authors
  1. Can this mechanism predict an intervention? Is there a tradeoff between control and accuracy?
4 comments
AuthorsOpenReview ↗

Response to reviewer `hi6Z` (1/2)

We appreciate reviewer hi6Z for their detailed reading and insightful comment on our work!

Weaknesses:

Thank you for mentioning the important study of Gaziv, Lee & DiCarlo (1). We are big fans of this line of work, and we will be sure to incorporate it into our discussion of related work.

We are also grateful for the reviewer’s framing suggestion: if we consider perception / perceptual judgement p(a∣s)p(a|s) as a function to be predicted or controlled, then the linear theory we studied can similarly be adapted to consider the problem of creating perceptually modulating stimuli.

As a side note, when reviewing the generated stimuli from Prince, Wang et al. (2), we noticed that the stimuli that successfully controlled neuronal responses also strongly modulated perception (i.e., the perception of the image changed from that of the seed image). This may be another clue that modulating neurons and modulating perception could involve some similar mechanisms.

At the same time, it’s important to note that several recent studies complicate the assumption that the factors underlying optimal control of ventral stream neurons (e.g., adversarial robustness training) transfer straightforwardly to yield strong forms of downstream behavioral alignment. For example, Doshi et al. (3) developed a clever paradigm for studying whether models show human-like configural shape processing using “visual anagram” stimuli, and found that while robustness training offered slight boosts in behavioral alignment to humans compared to baseline models, these models were substantially below the human threshold, and other inductive biases (e.g. language alignment) appeared to offer substantially greater boosts. Similarly, Reznik et al. (4) have shown that adversarial robustness training does not enable models to match human categorization judgments when stimuli are synthesized to elicit disagreement between robust models.

So, overall, the links between brain and behavioral alignment in DNN models are not as straightforward as one might expect; we are grateful that the reviewer suggested strengthening this facet of our discussion, and there are many exciting and testable questions still to explore to better understand these links.

Q1: Intervention

  1. Training side: Our current insight is to directly regularize or supervise the model’s gradients so that they have a more naturalistic structure. Empirically, adversarial training is an effective way to encourage natural-image-like gradients. In practice, adversarial fine-tuning offers a good recipe: start from a strong foundation model (CNN or transformer) and then make its gradients more perceptually aligned. Among our ten models, CLIPAG is the result of this kind of adversarial fine-tuning, and it performed well on neural control, as did ResNet50-robust (which was adversarially trained from scratch).
  2. Generation side: According to the theory, the failure of control is a form of out-of-distribution (OOD) generalization failure. If a stronger prior is supplied during the image generation process (e.g., diffusion models, GANs, etc. Pierzchlewicz et al. 2023; Wang & Ponce 2026), we expect the predictive model’s judgements to be better calibrated and the control to be more successful. We think this will be especially helpful for models that were not robustly trained.

  1. Gaziv, G.**,* Lee, M.J. *, & DiCarlo, J. J. (2023). Strong and precise modulation of human percepts via robustified ANNs. Advances in Neural Information Processing Systems 36 (NeurIPS 2023). arXiv:2308.06887.
  2. Prince, J. S., Wang, B., Fel, T., Jagadeesh, A. V., Vaziri, P. A., Alvarez, G. A., Livingstone, M. S., & Konkle, T. (2026). Parametric neural control differentiates top neural network models of primate visual cortex. bioRxiv.
  3. Doshi, F. R., Fel, T., Konkle, T., & Alvarez, G. (2025). Visual anagrams reveal hidden differences in holistic shape processing across vision models. arXiv:2507.00493.
  4. Reznik, I., Lapid, R., Sipper, M., & Golan, T. (2026). Adversarial robustness does not entail human-aligned categorization. arXiv:2611.00715.
  5. Pierzchlewicz, P. A., et al. (2023). Energy Guided Diffusion for Generating Neurally Exciting Images. Advances in Neural Information Processing Systems, 36, 32574–32601.
  6. Wang, B., & Ponce, C. R. (2026). Neuronal tuning aligns dynamically with object and texture manifolds across the visual hierarchy. Nature Neuroscience, 29, 864–875.
AuthorsOpenReview ↗

Response to reviewer `hi6Z` (2/2)

Q1.5: Tradeoffs

This is a very interesting question! It is well known that there is a tradeoff between adversarial robustness and object recognition accuracy.

In our case, however, the theory suggests that there should be no tradeoff between prediction and control: the optimal brain predictor is one that correctly predicts brain responses to every image, whether in or out of distribution. Such a predictor should do well on both control and prediction, since control is a special case of OOD prediction. Empirically, we also didn’t see evidence that top neural control models do worse on BrainScore type held out prediction — they are in line with each other. Centered within each neuronal site, the prediction score (Pearson r) and control score (slope) are positively correlated; same is true for average model score across all neuronal sites. So in this sense, improving BrainScore / held-out prediction is also helpful, it’s just not the sole factor determining control.

Table: Held-out prediction Pearson r in the control session, versus neural control slope.

ComparisonnPearson rp-valueSpearman ρp-value
Both variables centered within site250+0.3471.71 × 10⁻⁸ ***+0.1980.00162 **
Both variables centered within site (excluding untrained AlexNet)225+0.1080.106 NS+0.0470.480 NS
Model means across sites10+0.7090.0218 *+0.5390.108 NS

Dear Reviewer hi6Z,

Can you please take a look at the authors' response and provide your further input? Thank you.

AC.

I thank the authors for such a great discussion. They have addressed all my questions, and I keep my score it should be a clear acceptance.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We thank all reviewers for their careful reading and generally positive assessment. In addition to the point-by-point replies, we have polished the whole paper and especially the theory appendix:

  • Exposition: we added intermediate steps to the derivations, improved the layout, and reorganized several remarks and proofs so the arguments are easier to follow.
  • Notation: we removed notation collisions, along with propositions and symbols that were not needed for the argument.
  • Rigor: following the reviewers’ comments, we corrected statements that were imprecise. These include the growth claim in Corollary 19, which is bounded at fixed Σ\Sigma and diverges only as smin⁡→0s_{\min}\to0; the noiseless assumption behind the minimum-norm remark; and the definition of deterministic equivalence.
  • Related work: we added a connection to recent evidence that robustified networks steer human perception (Gaziv et al., 2023).
  • Intervention: thanks to the thought provoking discussion with the reviewers, we also added a discussion section on potential intervension methods during training and image generation to make better controllers.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.