iiinova — where your digital self keeps evolving

Tal Golan @talgolanneuro.bsky.social · 28/08/2026

For a survey of the math behind making models disagree, stay tuned for @kriegeskorte.bsky.social’s thread!

Tal Golan @talgolanneuro.bsky.social · 28/08/2026

These pioneering studies cover only a small part of what is possible when designing stimulus sets to discriminate among models. In our review, we map out this broader space and highlight directions we think are especially promising. rdcu.be/ah9QhWNMqG1a 16/16

rdcu.be

120

Tal Golan @talgolanneuro.bsky.social · 28/08/2026

More recently, @jfeather.bsky.social, @lipshutz.bsky.social, @sarah-harvey.bsky.social , @itsneuronal.bsky.social & @eerosim.bsky.social generalized this approach to more than two models with a method they call principal distortions. 15/16

110

Tal Golan @talgolanneuro.bsky.social · 28/08/2026

In 2024, Jingyang Zhou, @neurogramming.bsky.social, @ajaysub110.bsky.social & @eerosim.bsky.social developed an approach to compare the local geometry of two models, in this case, L1- vs. L2-regularized autoencoders. The L1 AE better predicted human sensitivity to the resulting perturbations. 14/16

120

Tal Golan @talgolanneuro.bsky.social · 28/08/2026

The last four studies above targeted global geometry, the set of distances among representations of different stimuli. A different family of approaches targets local geometry: how each model's representation of a stimulus changes when that stimulus is slightly perturbed. 13/16

110

Tal Golan @talgolanneuro.bsky.social · 28/08/2026

In the Review, we go into more detail on why decorrelating model predictions does not necessarily lead to experiments that are optimal for model discrimination. 12/16

110

Tal Golan @talgolanneuro.bsky.social · 28/08/2026

At the 2022 SVRHM workshop, Wenxuan Guo, @heikoschuett.bsky.social, @kriegeskorte.bsky.social, and I proposed going beyond RDM decorrelation by optimizing stimulus sets for reliable discrimination between representational models. We applied the approach to models of face perception. 11/16

120

Tal Golan @talgolanneuro.bsky.social · 28/08/2026

In 2024, @eghbal-hosseini.bsky.social , @coltoncasto.bsky.social, @nogazs.bsky.social, @colinconwell.link, Robert M Richardson, and @evfedorenko.bsky.social selected sentences that elicited decorrelated RDMs across language models and used them in a human fMRI experiment. 10/16

130

Tal Golan @talgolanneuro.bsky.social · 28/08/2026

In 2022, @kamilajozwik.bsky.social, @njarroyo.bsky.social , Jasper van den Bosch, Ian Charest, Radoslaw Cichy, and @kriegeskorte.bsky.social used a genetic algorithm to disentangle five properties related to animacy: being alive, looking like an animal, mobility, agency, and unpredictability. 9/16

110

Tal Golan @talgolanneuro.bsky.social · 28/08/2026

... and then tested these stimuli in behavioral and fMRI experiments. 8/16

110

Tal Golan @talgolanneuro.bsky.social · 28/08/2026

In 2018, @irisgroen.bsky.social , @mgreenephd.bsky.social, @chrisbaldassano.bsky.social , @drfeifei.bsky.social, Diane Beck, and @cibaker.bsky.social selected scenes so that RDMs based on action labels, object labels, and DNN features were as decorrelated as possible, 7/16

120

Tal Golan @talgolanneuro.bsky.social · 28/08/2026

Models can disagree about more than discrete choices. They can also disagree about which stimuli are (dis)similar, predicting different representational geometries. This suggests another strategy: choose stimuli that pull those geometries apart. 6/16

120

Tal Golan @talgolanneuro.bsky.social · 28/08/2026

The same principle can also be applied to odd-one-out judgments. In 2025, @mgreenephd.bsky.social and Bruce C. Hansen selected scene triplets so that three models, based on affordances, materials, or surfaces, each predicted a different odd one out. Human judgments favored affordances. 5/16

130

Tal Golan @talgolanneuro.bsky.social · 28/08/2026

Controversial stimuli are stimuli optimized to elicit distinct predictions from different models (Golan et al. 2020). Here, each image looks like a 3 to the column model and like a 7 to the row model. Human judgments aligned better with generative than discriminative classifiers. 4/16

110

Tal Golan @talgolanneuro.bsky.social · 28/08/2026

Maximum Differentiation (MAD) competition was introduced in 2008 by Zhou Wang and @eerosim.bsky.social . A pair of perturbed images is optimized so one model predicts a large difference in perceived distortion while the other stays fixed. Human judgments favored SSIM over MSE. 3/16

110

Tal Golan @talgolanneuro.bsky.social · 28/08/2026

Full text link: rdcu.be/ah9QhWNMqG1a 2/16

rdcu.be

120

Tal Golan @talgolanneuro.bsky.social · 28/08/2026

How can we design experiments that make computational models disagree? One section of our new @natrevneuro.nature.com Review with @kriegeskorte.bsky.social and @heikoschuett.bsky.social examines studies that used stimulus sets designed to elicit distinct predictions from competing models. 1/16

114156

Tal Golan @talgolanneuro.bsky.social · 10/07/2026

Obviously, these capabilities are enabled not only by the agents but also by the excellent infrastructure that @plotly.com provides to the scientific and technical community. 5/5

000

Tal Golan @talgolanneuro.bsky.social · 10/07/2026

And here are my 3D versions of Figures 3.8a and 3.8b: brainsandmachines.org/demos/artifi... brainsandmachines.org/demos/artifi... If you don't like a design choice I made, just download the HTML and use Codex or Claude Code to make changes; there's no backend. 4/5

brainsandmachines.org

100

Tal Golan @talgolanneuro.bsky.social · 10/07/2026

I estimate that creating and refining a new visualization now takes no more than an hour, and it can then really make a difference in the classroom for several years. Here are Prince's original UDL interactive figures: udlbook.github.io/udlfigures/ 3/5

udlbook.github.io

UDL Interactive Figures

Web site created using create-react-app

100

Tal Golan @talgolanneuro.bsky.social · 10/07/2026

So I asked Codex to create browser-based, interactive 3D versions of these figures using Plotly. After a few refinement prompts, it works very well. 2/5

110

Tal Golan @talgolanneuro.bsky.social · 10/07/2026

I find agentic coding a powerful amplifier for teaching. This term, I taught with @simonprinceai.bsky.social 's excellent book, Understanding Deep Learning. The book's interactive figures are brilliant, but some students struggle to decipher surfaces presented top-down as 2D heat maps. 1/5

141

Reposted by Tal Golan

Tim Kietzmann @timkietzmann.bsky.social · 06/05/2026

Happy to announce the 3rd iteration of NEAT (Neuro-AI-Talks), which will take place in Osnabrück September 14th-15th 2026. NEAT is a (deliberately small scale) NeuroAI workshop that brings together researchers from neuroscience and AI. www.kietzmannlab.org/neat2026/ More information below 👇

kietzmannlab.org

NEAT 2026

24424

Tal Golan @talgolanneuro.bsky.social · 29/01/2026

Israeli STEM postdocs abroad, have a look at this fellowship: www.gov.il/he/pages/rfp... 24 months x 13,500 NIS, Deadline: March 24, 2026 If you are working at the intersection of deep learning and cognitive (neuro)science, I'd be happy to host.

gov.il

010

Reposted by Tal Golan

CogCompNeuro @cogcompneuro.bsky.social · 03/12/2025

We are excited to announce that the Cognitive Computational Neuroscience meeting (CCN 2026) will be held at New York University from August 3–6, 2026. 2026.ccneuro.org

17428

Tal Golan @talgolanneuro.bsky.social · 20/11/2025

Proud to share our lab’s first paper! Itamar’s work highlights an implicit but impactful tension between maximizing fit to behavioral data and staying sensitive to meaningful differences between neural network models.

071

Reposted by Tal Golan

Jörn Diedrichsen @diedrichsenjorn.bsky.social · 10/09/2025

Variance partitioning is used to quantify the overlap of two models. Over the years, I have found that this can be a very confusing and misleading concept. So we finally we decided to write a short blog to explain why. @martinhebart.bsky.social @gallantlab.org diedrichsenlab.org/BrainDataSci...

26622

Tal Golan @talgolanneuro.bsky.social · 14/09/2025

We’re hiring in my dept. (IEM @ Ben-Gurion Univ.)! Tenure-track faculty opening in Robotics, Data/ML, Info Systems, Applied Stats, Production Systems, or Human Factors. www.bgu.ac.il/media/wwfbuc... Please repost 🙏

021

Reposted by Tal Golan

Dirk Bernhardt-Walther @dirkbwalther.bsky.social · 09/01/2024

Shape-Based Measures Improve Scene Categorization by DNNs. Morteza Rezanejad's magnus opus finally out in IEEE PAMI. doi.org/10.1109/TPAM... A super productive and fun collaboration with John Wilder, Allan D. Jepson, Sven Dickinson, and Kaleem Siddiqi. Read on for a quick summary.

An illustration of our approach using an example from a database of line drawings of natural scenes. In this pipeline, we first extract line drawings of the input image, then its Medial Axis Transform (MAT) is computed and then the skeletal points on the MAT are scored based upon one of our importance score measures. Finally, they are projected back onto the contours. The figure on the top left shows an example scene used in our pipeline, and the figure below presents the jet colormap visualization of its contours using our medial axis-based contour importance measures.

324