Lightnews — Scholar-powered news

Steven Scholte @neurosteven.bsky.social · 28d

An idea to post an expanded version on biorxiv.org?

Steven Scholte @neurosteven.bsky.social · 28d

That is deal with the statistical structure of the world on the basis of learning experience (sampling + evolution). To jump to human behaviour would be, I think, mixing up 2, or potentially 3 levels of observation.

1

Steven Scholte @neurosteven.bsky.social · 28d

DCNNs here confirm our ideas of vision at a neuroscience level and potentially expand these with a broader view of what these filters are and how they can emerge. Beyond that, it suggests a theory of what vision does at this stage.

1 1

Steven Scholte @neurosteven.bsky.social · 28d

I guess the theory that you would abstract from results such as these is that feedforward vision is dominated by features that naturally emerge through hierarchical processing and show these features can emerge in a convolutional tree.

1

Steven Scholte @neurosteven.bsky.social · 29d

A substantial amount of the neural activity that can be explained relates to these processes, and this part (my 5 cents, this data) is (the bulk) of what is explained by DCNNs. But a necessary part of processing for vision outside the lab.

1

Steven Scholte @neurosteven.bsky.social · 29d

For the model to classify objects it does a lot of 'stuff'. For instance, without a background a shallow network suffices (Seijdel et al., 2020, scholar.google.com/citations?vi...). Natural images force these networks to do a lot more than what we would label object recognition.

scholar.google.com

1

Steven Scholte @neurosteven.bsky.social · 29d

It has been surprising for me the last 7 years how easy it is to find signature of texture processing (Loke et al., 2024) and scene segmentation (Seijdel et al., 2020, 2021) in DCNNs and how little of it seems to relate to subsequent steps. We have really been trying.

1

Steven Scholte @neurosteven.bsky.social · 29d

DNNs still capture an impressive amount of variance. The most parsimonious account I think is that DNNs model the initial encoding well but miss, or perform differently, subsequent steps of object recognition.

2 1 1

Steven Scholte @neurosteven.bsky.social · 29d

Great work from my PhD Jessica Loke, together with @lynnkasorensen.bsky.social , @irisgroen.bsky.social and Nathalie Cappaert.

3

Steven Scholte @neurosteven.bsky.social · 29d

Take-home: High DNN-brain correlations may reflect shared texture processing rather than object understanding. This challenges how we interpret brain-AI alignment studies and suggests new directions for computational models. www.biorxiv.org/content/10.1...

Shared texture-like representations, not global form, underlie deep neural network alignment with human visual processing

Deep neural networks (DNNs) are a leading computational framework for understanding neural visual processing. A standard approach for evaluating their similarity to brain function uses DNN activations...

www.biorxiv.org

1 1 9

Steven Scholte @neurosteven.bsky.social · 29d

Trajectories make it visual:
🔵 Texture path → high alignment, low object info (upper-left quadrant)
🔴/🟢 Natural & object-only paths → more object info but no extra alignment.
This explains why better object recognition ≠ better brain prediction.

1 1

Steven Scholte @neurosteven.bsky.social · 29d

Cross-prediction: texture features predict brain responses to natural scenes almost as well as features from the originals themselves.
Local image statistics = the common representational currency between artificial and biological vision.

1 1

Steven Scholte @neurosteven.bsky.social · 29d

The key dissociation:
• EEG encodes object category across all conditions.
• But object info does not drive DNN–brain alignment.
• Peak alignment occurs when object info is minimal (texture condition).

1 1

Steven Scholte @neurosteven.bsky.social · 29d

Three versions of each image:
🔴Natural scenes
🔵Texture-synthesized (global summaries of local stats only; no recognizable objects)
🟢Object-only (objects without backgrounds). Counterintuitive result: strongest DNN–brain alignment for texture-only images!

1 1

Steven Scholte @neurosteven.bsky.social · 29d

🧠 New preprint: Why do deep neural networks predict brain responses so well?
We find a striking dissociation: it’s not shared object recognition. Alignment is driven by sensitivity to texture-like local statistics.
📊 Study: n=57, 624k trials, 5 models doi.org/10.1101/2025...

4 36 110

Steven Scholte @neurosteven.bsky.social · Sep 7

Our response is due in 3 weeks. Pondering.

3

Reposted by Steven Scholte

Matthias Nau @matthiasnau.bsky.social · Aug 18

Datasets like NSD & THINGS offer rich stimuli but often test a single task.

After great conversations at #CCN2025 on multi-task studies & generalization in brains & models, I thought I would repost our perspective for those interested in this topic. We need multiple tasks!👉 doi.org/10.1038/s415...

Centering cognitive neuroscience on task demands and generalization - Nature Neuroscience

Task demands are a primary determiner of behavior and neurophysiology. Here the authors discuss how understanding their influence through multitask studies and tests of generalization is the key to ar...

doi.org

1 14 50

Steven Scholte @neurosteven.bsky.social · Aug 17

Thank you!!!!

9

Steven Scholte @neurosteven.bsky.social · Aug 17

Lynn Flannery, Kerry Miller, Jeff Wilson, Kevin Koenrades, Brenda Klappe and our Volunteers: Nina Fitzmaurice, Ole Jürgensen, Denise Kittelmann, Elif Ayten
Maithe van Noort, Mohanna Hoveyda, Caroline Harbison, Yamil Vidal, Sotirios Panagiotou,Sofie Wahlberg, Danting Meng, Mobina Tousian.

1 9

Steven Scholte @neurosteven.bsky.social · Aug 17

@claires012345.bsky.social @mheilbron.bsky.social Angela Radulescu @neuroprinciplist.bsky.social @dotadotadota.bsky.social Tyler bonnen Sneha Aenugu @hannesmehrer.bsky.social @debyee.bsky.social Julian Kosciessa @anne-urai.bsky.social @mdhk.net @tdado.bsky.social Shauney Wilson, Shawna Lampkin,

mehrer.bsky.social

1 13

Steven Scholte @neurosteven.bsky.social · Aug 17

But could not have run without @lauragwilliams.bsky.social @jaspervdb.bsky.social @achterbrain.bsky.social @niklasmuller.bsky.social @eringrant.me @pebenjamters.bsky.social @shahabbakht.bsky.social @judithfan.bsky.social @jfeather.bsky.social Jiahui Guo @tknapen.bsky.social @lampinen.bsky.social

1 11

Steven Scholte @neurosteven.bsky.social · Aug 17

a reception in Hotel Arena, an epic-party in Ijver (with about 40% of attendees on the dance floor), and above all 929 community members who brought a lot of energy and hopefully had, on multiple dimensions a fantastic conference.
So proud to be, together with @irisgroen.bsky.social chair.

This link will take you to a page that’s not on LinkedIn

lnkd.in

1 1 15

Steven Scholte @neurosteven.bsky.social · Aug 17

#CCN2025 is over. Over 5 days there were 6 fantastic keynotes, 550 posters, 3 community events, 3 keynote & tutorials, 3 generative adversarial collaborations, 8 Satellite events, 1 community lunch meeting, 1 cross-conference hackathon, 1 competition, coffee all day, stroopwafels on day 1,

1 7 62

Reposted by Steven Scholte

CogCompNeuro @cogcompneuro.bsky.social · Aug 15

That's a wrap for CCN2025 -- and so planning for CCN2026 in New York is starting today! Save travels to all participants and remember to fill out the feedback survey sent via email!

1 7 56

Reposted by Steven Scholte

CogCompNeuro @cogcompneuro.bsky.social · Aug 11

The moment we've all been waiting for— #CCN2025 is HERE! Today we start with satellite events, and tomorrow the main conference begins in Amsterdam! We'll be sharing daily updates about each day's program, so follow the CCN account to stay in the loop. Can't wait to see everyone!

1 4 35