Visualizing and Controlling Cortical Responses Using Voxel-Weighted Activation Maximization

Mark D. Lescroart; Matthew W. Shinkle

REVIEW 4 major objections 5 minor 24 references

Visualizing and Controlling Cortical Responses Using Voxel-Weighted Activation Maximization

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that activation maximization applied to a whole-network encoding model built from a pretrained Inception V3 network can synthesize images that characterize and reliably drive fMRI responses in targeted regions across the…

desk verdict A legitimate first application of whole-network activation maximization to human fMRI encoding models, with real external validation, but the abstract oversells the reliability of the effects. read the letter →

arxiv 2506.04379 v1 pith:RUXH6PGU submitted 2025-06-04 cs.CV cs.AIq-bio.NC

Matthew W. Shinkle , Mark D. Lescroart This is my paper

classification cs.CVcs.AIq-bio.NC

keywords activationmaximizationvoxelwiseencodingmodelsfMRIvisualcortexfeaturevisualizationsyntheticimagesInceptionV3deepneuralnetworks

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that activation maximization, a technique originally built for interpreting deep neural networks, can be applied directly to encoding models of the human brain. The authors fit a voxelwise linear mapping from multi-layer, downsampled activations of a pretrained Inception V3 network to fMRI responses to naturalistic movies, then run gradient ascent on that mapping to synthesize images predicted to drive individual cortical voxels or whole regions of interest. They report that single-voxel images carry features consistent with known retinotopic and category selectivity, and that whole-region images reliably activate their target regions in new fMRI sessions, including sessions from subjects whose data were not used to generate the images. If correct, the method provides a flexible, image-computable way to both visualize and control cortical responses without requiring a generative model.

What carries the argument

The central machinery is an augmented encoding model: a pretrained Inception V3 network whose activations from 21 layers are adaptively downsampled to roughly 5,000 features per layer, flattened and concatenated into roughly 78,000 features, and mapped to each voxel's z-scored fMRI response by ridge regression. The regression weights are treated as a new fully connected layer appended to the network, so gradients of the predicted response with respect to the input image can be computed by backpropagation. Activation maximization then performs gradient ascent in Fourier space with random augmentations and color decorrelation, synthesizing images that maximize the predicted response of a target voxel or a contrast-weighted average over an ROI. The contrast step z-scores the weights, subtracts a shared mean selectivity, and normalizes the result, so that optimization emphasizes features unique to the target rather than common visual features.

What would settle it

A decisive falsifier would be a new-subject fMRI experiment comparing images optimized for voxels with opposite known selectivity, such as upper- versus lower-field retinotopy: if each image fails to evoke the predicted spatial pattern of activation, or if images optimized on an encoding model fit to randomly shuffled fMRI responses look equally structured, the central claim collapses.

Watch

Extended reading notes

Core claim

The central discovery claimed is that an encoding model assembled from layers spanning an entire pretrained network can serve as the target of activation maximization, and the resulting synthetic images preserve the selectivity of the voxels and regions for which they were optimized. Evidence comes from both qualitative and quantitative fronts: images optimized for early visual voxels contain high-contrast patches in the expected retinotopic locations, images for face-selective regions show eye- and face-like elements, and images for scene-selective regions show building- and street-like structure. In fMRI validation, ROI-optimized images evoked the highest response in the targeted region in four of five ROIs, with statistically significant differences in EBA, FFA, and RSC for one subject and FFA for another, and group-averaged weights produced images that activated the same regions in a new set of subjects. The paper also reports exploratory findings from voxels outside established ROIs, including a posterior parietal scene-selective patch and body-selective voxels posterior to PPA.

Load-bearing premise

The approach rests on the assumption that the fitted equation from brain-image features to measured brain activity is a faithful causal model of what each brain location responds to, so that images found by climbing that equation truly excite the intended locations instead of exploiting quirks of the model.

Editorial extensions

If this is right

Single-voxel and ROI-optimized images can be used to survey selectivity across the entire visual cortex, including voxels that fall outside established functional regions.
ROI-optimized images evoke stronger fMRI responses in their target regions than in comparison regions in most tested areas, indicating that the method can be used to control cortical responses, not just visualize them.
Images generated from group-averaged weights drive responses in subjects whose data were not used for synthesis, suggesting the captured tuning is shared across individuals rather than idiosyncratic.
Because the approach requires only a pretrained, non-generative network, it can be adapted to other architectures and other differentiable optimization objectives, including optimization against a property or for stimulus formats like video.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

An ablation study could determine how much of the observed specificity comes from the contrast-weighting step: optimizing raw, uncontrasted weights would show whether the pipeline depends on subtracting shared selectivity to produce interpretable images.
Systematically varying which of the 21 layers feed the regression could map which representational levels drive each region, and would test whether the full-network design, rather than any single layer, is what enables control of high-level areas.
A control experiment training the same encoding model on shuffled fMRI responses, then running activation maximization, would show whether the synthesis pipeline generates structured images even without meaningful brain data; if it does, image interpretability alone cannot attest to cortical selectivity.
The cross-subject success suggests a closed-loop protocol in which synthetic images are refined against real-time fMRI could be used to probe selectivity in newly identified voxels without waiting for group-level ROIs.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two things to know about arXiv:2506.04379. It's a legitimate first: whole-network, multi-layer activation maximization applied to human fMRI encoding models, with real external validation. And the abstract oversells the reliability — the significant effects are confined to a subset of ROIs and the comparisons lack a neutral baseline.

What's actually new: Bashivan et al. did single-layer AM for macaque V4; this extends to 21 layers of Inception V3 and human fMRI, with voxelwise ridge regression. The cross-subject experiment (generate on four subjects, test on four different subjects) is the right design and reduces circularity. Optimizing contrastive weights, with the average ROI pattern subtracted, is a sensible way to generate discriminative images. The encoding model accuracy in the supplement looks solid — high noise-ceiling-corrected R2 across early and category-selective areas.

Where it's soft: the 'reliably drive' claim is not supported by the pattern. In Experiment 2, only EBA, FFA, and RSC for subject 1 and FFA for subject 2 reach significance; the rest are null. In Experiment 3, V3, EBA, FFA survive, RSC doesn't, and LO is not reported. That's roughly half the comparisons. Also, because the optimization target is contrastive, the validation is inherently relative — there is no phase-scrambled or otherwise matched control, so you cannot tell whether the images actually drive activity above a neutral baseline, or just look different from each other. The single-voxel visualizations in Experiment 1 are assessed qualitatively, as the authors admit in their limitations. That is acceptable for an exploratory tool, but not for the strong claim that these images characterize selectivity. Code is on GitHub but with no commit hash, and data are not deposited, so reproducibility is partial.

Who it's for: vision neuroscientists who build encoding models and want a flexible synthesis tool that works with any differentiable network. The method is straightforward and should be useful. The paper deserves a serious referee: the core idea is sound and the external validation is real, but the authors need to temper the abstract, add a low-level control, and report all ROI comparisons including LO and RSC in Experiment 3.

Referee Report

4 major / 5 minor

Summary. The paper proposes a method that applies activation maximization to a whole-network DNN-based encoding model of human fMRI responses. The authors extract and adaptively downsample activations from multiple layers of a pretrained Inception V3 model, train a ridge regression mapping to predict voxelwise BOLD responses to natural movies, and then use gradient ascent in Fourier space to synthesize images that maximize predicted responses for individual voxels or ROIs. They report qualitative correspondence between single-voxel images and known selectivity, and present two fMRI experiments in which images optimized for whole ROIs are presented to participants to test whether they drive targeted responses. The abstract claims this is the first demonstration that activation maximization can be successfully applied to whole-network encoding models to characterize and modulate fMRI responses across the visual system.

Significance. If the central claims hold, this work would provide a useful, image-computable tool for visualizing and controlling cortical responses without requiring a natively generative model, extending earlier single-layer work to multi-layer, whole-network encoding models. The strongest evidence is the real fMRI validation with externally presented synthetic images, including a cross-subject generalization experiment and the use of multiple layers of a pretrained network. The authors provide code and an interactive visualization, which are valuable for reproducibility. However, the validation results are only partially significant, and several methodological ambiguities (notably the handling of temporal lags and the absence of low-level controls) currently undermine the strength of the central claim.

major comments (4)

[Sec. 6.3, Eq. (2); Sec. 6.4, Eq. (3)] There is a mismatch between the encoding model and the activation-maximization loss. The encoding model is fit with a design matrix that combines temporal lags of 2, 4, and 6 seconds, so the regression weight vector beta presumably contains separate weights for each lag. In contrast, the activation-maximization loss in Eq. (3) is defined as the dot product between the instantaneous feature vector f(x) and beta. If beta includes all lags, the dimensions do not match; if the weights across lags were collapsed into a single vector, that step is not described anywhere. This is a load-bearing technical detail because it determines what the optimization target actually represents, and it needs to be clarified or corrected.
[Sec. 3.2, Sec. 3.3, Abstract] The statistical evidence is weaker than the abstract and conclusions claim. In Experiment 2, significant same-ROI preference was observed in only EBA, FFA, and RSC for subject 1 and only FFA for subject 2; in Experiment 3, only V3, EBA, and FFA reached significance, with RSC non-significant and LO not reported. This pattern does not support the statement that generated images 'reliably drive activity in targeted regions across both low- and high-level visual areas and across subjects.' The claim should be moderated to reflect the observed partial and variable significance, or additional analyses/experiments should be provided to strengthen the evidence.
[Sec. 3.2, Sec. 3.3, Sec. 6.5] The experiments lack a control condition that matches the low-level image statistics of the synthetic stimuli. Since the optimized images differ substantially in low-level properties (for example, V3 images are high-contrast patterns while FFA images contain facial-like elements), the observed ROI-specific responses could be driven by differences in luminance, contrast, or spatial frequency rather than by the DNN-derived feature selectivity that the method aims to exploit. A phase-scrambled or spectrally matched control would help rule out this alternative explanation and is essential for interpreting the causal validity of the generated images.
[Sec. 3.3, Supplementary Sec. 5.1] The reporting of Experiments 2 and 3 is incomplete in ways that affect the assessment of the central claim. For Experiment 3, LO is listed as one of the five target ROIs, but no significance result is reported for it, and the text only notes that RSC was non-significant. Additionally, the main text states that the encoding model predicts responses accurately 'including in all areas examined in our experiments', yet the actual prediction accuracy values are only shown in a supplementary figure without numerical details. The authors should report full results for all ROIs and provide quantitative encoding accuracy values, so that readers can evaluate the model's reliability in the regions where validation failed.

minor comments (5)

[Title and header] The title contains a typographical spacing error: 'V oxel-Weighted' should be 'Voxel-Weighted'.
[Fig. 2 and Sec. 3.1] The qualitative evaluation of single-voxel images against known selectivity is inherently circular because the voxels are selected from the same ROIs whose selectivity is already characterized. Consider adding a more quantitative or blinded assessment, or at least clearly label these results as descriptive rather than confirmatory.
[Sec. 6.4] The description of the Fourier-domain optimization ('taking the magnitude to obtain real-valued pixel intensities') does not explain how gradients flow through the magnitude operation. Please clarify this step, as it is important for reproducibility.
[Sec. 3.2 and Sec. 3.3] The text reports only the t-test threshold (t(18) > 2.101, p < .05) rather than actual t-statistics or effect sizes. Reporting the full statistics, including confidence intervals, would allow readers to assess the magnitude of the effects.
[Sec. 6.3] The choice of the ridge regularization parameter alpha over a range from 100 to 10^10 is extremely wide; it would be helpful to report the distribution of selected alpha values to ensure the model is not over- or under-regularized.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central validation is genuinely external (new fMRI responses to synthesized images, with Experiment 3 using separate subjects), and the only self-citation (Pycortex) is a visualization tool, not load-bearing.

full rationale

The derivation chain is self-contained and does not reduce to its own inputs. The encoding model is fit with ridge regression to naturalistic movie fMRI responses; the same model is then used to synthesize images that maximize predicted responses. The validation consists of presenting those synthesized images to human participants and measuring BOLD responses that were not used to fit the model. This is an out-of-stimulus, and in Experiment 3 out-of-subject, test. The held-out movie prediction accuracy is standard cross-validation, not a circular prediction. The qualitative comparison in Experiment 1 uses known selectivity from the literature, and the ROI validation in Experiments 2 and 3 uses measured responses, so even though generation and testing share subjects in Experiment 2, the measured responses are independent of the fitting procedure. The only self-citation is reference [4] (Pycortex), which is used for surface visualization and is not load-bearing for any scientific claim. No uniqueness theorem is imported, no ansatz is smuggled via self-citation, and no fitted parameter is relabeled as a prediction. The paper's own limitation (qualitative characterization of image contents) and the partially non-significant ROI results are concerns about evidence strength and statistical power, not about circularity. Therefore no circular step can be exhibited by quotation and reduction.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new particles, forces, dimensions, or conserved quantities. The posterior parietal scene patch and PPA-adjacent body patch are data-driven observations, not postulates.

free parameters (6)

Fmax (per-layer feature budget) = 5000
Adaptive downsampling target in Eq. 1; chosen by hand to balance feature counts across layers.
Ridge regularization alpha = Log-spaced 100 to 1e10, selected per voxel via 10-fold CV
Hyperparameter selected on training data; affects the encoding model weights used for image optimization.
Number of optimization iterations = 2500
Activation maximization stopping point chosen by hand; affects image quality and interpretability.
Adam learning rate = 1e-2
Optimizer hyperparameter for gradient ascent; chosen to balance convergence and stability.
Augmentation ranges = rotation -5 to 5 deg, scale 0.95 to 1.05, padding 3 to 5 px
Regularization choices that shape the statistics of the synthesized images.
Contrast objective components = z-scored weights minus mean across voxels or ROIs, L2 normalized
The contrastive reweighting in Eqs. 4-6 determines what the optimization targets; chosen by the authors, not derived from neural data.

assumptions (4)

domain assumption ImageNet-pretrained Inception V3 features resemble hierarchical human visual cortex representations.
Invoked in the Introduction and Methods to justify using fixed DNN features without training a deep network on fMRI data.
domain assumption Voxelwise fMRI responses can be approximated as a linear (ridge) function of DNN feature activations.
Core encoding model assumption; standard in the encoding-model literature the paper cites.
domain assumption Voxel selection by noise ceiling (p<0.05, uncorrected) yields reliable voxels without biasing later ROI analyses.
Supplementary Methods 6.1; the uncorrected threshold may affect which voxels are included in the model and the fMRI validation.
domain assumption BOLD responses estimated with GLMsingle and temporal downsampling to TR 2s provide unbiased estimates of image-evoked responses.
Supplementary Methods 6.1; assumes the response estimation pipeline is appropriate for rapid event-related image presentation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Visualizing and Controlling Cortical Responses Using Voxel-Weighted Activation Maximization." pith.science (2026). https://pith.science/paper/RUXH6PGU

@misc{pith2026250604379,
  author       = {Pith},
  title        = {Pith review of: Visualizing and Controlling Cortical Responses Using Voxel-Weighted Activation Maximization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RUXH6PGU}},
  note         = {Machine review of arXiv:2506.04379}
}

read the original abstract

Deep neural networks (DNNs) trained on visual tasks develop feature representations that resemble those in the human visual system. Although DNN-based encoding models can accurately predict brain responses to visual stimuli, they offer limited insight into the specific features driving these responses. Here, we demonstrate that activation maximization -- a technique designed to interpret vision DNNs -- can be applied to DNN-based encoding models of the human brain. We extract and adaptively downsample activations from multiple layers of a pretrained Inception V3 network, then use linear regression to predict fMRI responses. This yields a full image-computable model of brain responses. Next, we apply activation maximization to generate images optimized for predicted responses in individual cortical voxels. We find that these images contain visual characteristics that qualitatively correspond with known selectivity and enable exploration of selectivity across the visual cortex. We further extend our method to whole regions of interest (ROIs) of the brain and validate its efficacy by presenting these images to human participants in an fMRI study. We find that the generated images reliably drive activity in targeted regions across both low- and high-level visual areas and across subjects. These results demonstrate that activation maximization can be successfully applied to DNN-based encoding models. By addressing key limitations of alternative approaches that require natively generative models, our approach enables flexible characterization and modulation of responses across the human visual system.

Figures

Figures reproduced from arXiv: 2506.04379 by the authors.

**Figure 1.** Model Overview. Schematic of our augmented DNN framework. Movie frames are processed by a pretrained Inception V3 model. Activations are extracted from multiple layers and adaptively downsampled, flattened, and concatenated to form a feature vector that is passed through a new fully connected layer. The weights of this layer are trained via ridge regression to predict voxelwise fMRI responses. Core Encoding Model T… view at source ↗

**Figure 2.** Synthetic images optimized for individual cortical voxels. Images synthesized for individual voxels in V2 show highcontrast patches in expected retinotopic locations; those for voxels drawn from face-selective regions (OFA/FFA) exhibit facial features, including eye- and face-like components; and those from scene-selective regions (OPA) display scene-like elements. In addition, images synthesized for voxels outsi… view at source ↗

**Figure 3.** Real fMRI responses to images synthesized for whole ROIs. (A) Synthetic images optimized via activation maximization for five target regions (V3, LO, FFA, EBA, and RSC) from Subject 1. These images contain characteristics which qualitatively correspond to known region selectivity. For example, V3 images exhibit high-contrast patterns spanning the visual field (consistent with the relatively low-level visual feature … view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Prediction Accuracy of the Inception V3-Based Encoding Model. (A) Flatmaps of voxelwise prediction accuracy for subjects 1 and 2 in natural movie data, illustrating the encoding model prediction accuracy across the visual cortex. (B) Bar chart displaying ROIwise predi…

**Figure 5.** Figure 5: Cross-Subject Generalization of ROI-Level Synthetic Images. (A) Example synthetic images for each ROI, generated using predicted responses from a group of four subjects and then presented to a separate set of four subjects. These images capture regionspecific features…

**Figure 6.** Figure 6: Generation Process. Diagram illustrating the activation maximization procedure. Starting from an initial neutral grayscale image in the Fourier domain, the image is transformed into the spatial domain using an inverse FFT. Random invariance transforms (e.g., cropping,…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages

[1]

Pouya Bashivan, Kohitij Kar, and James J. DiCarlo. Neural population control via deep image synthesis. Science, 364 (6439):eaav9436, 2019. 1

work page 2019
[2]

Nunez- Elizalde, and Jack L

Tom Dupr ´e La Tour, Michael Eickenberg, Anwar O. Nunez- Elizalde, and Jack L. Gallant. Feature-space selection with banded ridge regression. NeuroImage, 264:119728, 2022. 3, 5

work page 2022
[3]

Visualizing higher-layer features of a deep network

Dumitru Erhan, Yoshua Bengio, Aaron Courville, and Pascal Vincent. Visualizing higher-layer features of a deep network. University of Montreal, 1341(3):1, 2009. 2

work page 2009
[4]

Gao, Alexander G

James S. Gao, Alexander G. Huth, Mark D. Lescroart, and Jack L. Gallant. Pycortex: an interactive surface visualizer for fmri. Frontiers in Neuroinformatics, 9, 2015. 5

work page 2015
[5]

Allen, Yihan Wu, Ghislain St-Yves, Thomas Nase- laris, Kendrick Kay, Mert R

Zijin Gu, Keith Wakefield Jamison, Meenakshi Khosla, Emily J. Allen, Yihan Wu, Ghislain St-Yves, Thomas Nase- laris, Kendrick Kay, Mert R. Sabuncu, and Amy Kuceyeski. NeuroGen: Activation optimized image synthesis for discov- ery neuroscience. NeuroImage, 247:118812, 2022. 1, 4

work page 2022
[6]

Guclu and M

U. Guclu and M. A. J. Van Gerven. Deep neural networks reveal a gradient in the complexity of neural representations across the ventral stream. Journal of Neuroscience, 35(27): 10005–10014, 2015. 1

work page 2015
[7]

Hansen, Kendrick N

Kathleen A. Hansen, Kendrick N. Kay, and Jack L. Gallant. Topographic organization in and near human visual area v4. The Journal of Neuroscience, 27(44):11896–11911, 2007. 3

work page 2007
[8]

Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition , pages 346–361

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition , pages 346–361. Springer Interna- tional Publishing, 2014. 3

work page 2014

Show all 24 references

[9]

Huth, Shinji Nishimoto, An T

Alexander G. Huth, Shinji Nishimoto, An T. Vu, and Jack L. Gallant. A continuous semantic space describes the repre- sentation of thousands of object and action categories across the human brain. Neuron, 76(6):1210–1224, 2012. 2, 1

2012
[10]

Kennedy, Sarala N

B. Kennedy, Sarala N. Malladi, R. Tootell, and S. Nasr. A previously undescribed scene-selective site is the key to en- coding ego-motion in naturalistic environments. eLife, 2024. 3, 4

2024
[11]

Deep neural networks: A new frame- work for modeling biological vision and brain information processing

Nikolaus Kriegeskorte. Deep neural networks: A new frame- work for modeling biological vision and brain information processing. Annual Review of Vision Science, 1(1):417–446,
[12]

Deep learning

Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436–444, 2015. 1, 2

2015
[13]

Luo, Margaret M

Andrew F. Luo, Margaret M. Henderson, Leila Wehbe, and Michael J. Tarr. Brain diffusion for visual exploration: Cor- tical discovery using large scale generative models, 2023. 1, 4

2023
[14]

Kay, Shinji Nishimoto, and Jack L

Thomas Naselaris, Kendrick N. Kay, Shinji Nishimoto, and Jack L. Gallant. Encoding and decoding in fmri. NeuroIm- age, 56(2):400–410, 2011. 2, 3

2011
[15]

Multifaceted feature visualization: Uncovering the different types of fea- tures learned by each neuron in deep neural networks, 2016

Anh Nguyen, Jason Yosinski, and Jeff Clune. Multifaceted feature visualization: Uncovering the different types of fea- tures learned by each neuron in deep neural networks, 2016. 2

2016
[16]

Feature visualization

Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization. Distill, 2(11), 2017. 1, 2, 3, 4

2017
[17]

Ponce, Will Xiao, Peter F

Carlos R. Ponce, Will Xiao, Peter F. Schade, Till S. Hartmann, Gabriel Kreiman, and Margaret S. Livingstone. Evolving images for visual neurons using a deep generative network reveals coding principles and neuronal preferences. Cell, 177(4):999–1009.e10, 2019. 1

2019
[18]

Visual and linguistic semantic representa- tions are aligned at the border of human visual cortex.Nature Neuroscience, 24(11):1628–1636, 2021

Sara F Popham, Alexander G Huth, Natalia Y Bilenko, Fatma Deniz, James S Gao, Anwar O Nunez-Elizalde, and Jack L Gallant. Visual and linguistic semantic representa- tions are aligned at the border of human visual cortex.Nature Neuroscience, 24(11):1628–1636, 2021. 1

2021
[19]

Improving the accuracy of single-trial fmri response estimates using glmsingle.eLife, 11, 2022

Jacob S Prince, Ian Charest, Jan W Kurzawski, John A Pyles, Michael J Tarr, and Kendrick N Kay. Improving the accuracy of single-trial fmri response estimates using glmsingle.eLife, 11, 2022. 2, 5

2022
[20]

Rethinking the in- ception architecture for computer vision, 2015

Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the in- ception architecture for computer vision, 2015. 2, 3

2015
[21]

Daniel L. K. Yamins, Ha Hong, Charles F. Cadieu, Ethan A. Solomon, Darren Seibert, and James J. DiCarlo. Performance-optimized hierarchical models predict neural responses in higher visual cortex. Proceedings of the Na- tional Academy of Sciences, 111(23):8619–8624, 2014. 1 Vi...

2014
[22]

Prediction accuracy was noise-ceiling cor- rected and averaged across voxels and subjects for each re- gion of interest (ROI)

Supplementary Results To evaluate the overall quality of the DNN-to-brain map- ping, we computed the prediction accuracy of our Inception V3-based encoding model on the testing set of naturalis- tic movie clips. Prediction accuracy was noise-ceiling cor- rected and averaged ac...
[23]

fMRI Data Acquisition and Preprocessing For fitting our DNN-based encoding models, we used BOLD fMRI responses to a large set of naturalistic movie clips from Huth et al

Supplementary Methods 6.1. fMRI Data Acquisition and Preprocessing For fitting our DNN-based encoding models, we used BOLD fMRI responses to a large set of naturalistic movie clips from Huth et al. [9]. This consisted of video clips de- picting a wide variety of dynamic scenes...
[24]

Adaptive Spatial Downsampling: To manage the high dimensionality of activations from convolutional layers, we employ adaptive spatial pooling

FC This selection spans from low-level features to high-level semantic representations. Adaptive Spatial Downsampling: To manage the high dimensionality of activations from convolutional layers, we employ adaptive spatial pooling. We use PyTorch’s adaptive pooling functions (s...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Pouya Bashivan, Kohitij Kar, and James J. DiCarlo. Neural population control via deep image synthesis. Science, 364 (6439):eaav9436, 2019. 1

work page 2019

[2] [2]

Nunez- Elizalde, and Jack L

Tom Dupr ´e La Tour, Michael Eickenberg, Anwar O. Nunez- Elizalde, and Jack L. Gallant. Feature-space selection with banded ridge regression. NeuroImage, 264:119728, 2022. 3, 5

work page 2022

[3] [3]

Visualizing higher-layer features of a deep network

Dumitru Erhan, Yoshua Bengio, Aaron Courville, and Pascal Vincent. Visualizing higher-layer features of a deep network. University of Montreal, 1341(3):1, 2009. 2

work page 2009

[4] [4]

Gao, Alexander G

James S. Gao, Alexander G. Huth, Mark D. Lescroart, and Jack L. Gallant. Pycortex: an interactive surface visualizer for fmri. Frontiers in Neuroinformatics, 9, 2015. 5

work page 2015

[5] [5]

Allen, Yihan Wu, Ghislain St-Yves, Thomas Nase- laris, Kendrick Kay, Mert R

Zijin Gu, Keith Wakefield Jamison, Meenakshi Khosla, Emily J. Allen, Yihan Wu, Ghislain St-Yves, Thomas Nase- laris, Kendrick Kay, Mert R. Sabuncu, and Amy Kuceyeski. NeuroGen: Activation optimized image synthesis for discov- ery neuroscience. NeuroImage, 247:118812, 2022. 1, 4

work page 2022

[6] [6]

Guclu and M

U. Guclu and M. A. J. Van Gerven. Deep neural networks reveal a gradient in the complexity of neural representations across the ventral stream. Journal of Neuroscience, 35(27): 10005–10014, 2015. 1

work page 2015

[7] [7]

Hansen, Kendrick N

Kathleen A. Hansen, Kendrick N. Kay, and Jack L. Gallant. Topographic organization in and near human visual area v4. The Journal of Neuroscience, 27(44):11896–11911, 2007. 3

work page 2007

[8] [8]

Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition , pages 346–361

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition , pages 346–361. Springer Interna- tional Publishing, 2014. 3

work page 2014