iiinova — where your digital self keeps evolving

Maxence Pajot @maxencepajot.bsky.social · 21/09/2026

It's possible yes! We go in more detail in the discussion of the paper, but even if the successes are impressive, models might still lack something to fully model human perception of geometric shapes.

010

Maxence Pajot @maxencepajot.bsky.social · 20/09/2026

It's an interesting suggestion! I ran a quick test with the available weights on the quadrilateral task. The DVD seems to improve alignment over the baseline within your pipeline. However all models trained with this pipeline were less aligned with human behavior than the classically trained resnet.

111

Maxence Pajot @maxencepajot.bsky.social · 02/09/2026

The third task is the most demanding: matching a photograph with a minimalist drawing of the same item, for instance a fish drawn as an ellipse and a triangle. Humans find this trivial. No model does, although the best ones get close to human performance. 7/8

120

Maxence Pajot @maxencepajot.bsky.social · 02/09/2026

In the second task, we go beyond quadrilaterals, looking at the representations of a broader set of 68 shapes. Alignment with humans was again mainly driven by training set size. However, adding symbolic information did bring the best models closer to humans. 6/8

110

Maxence Pajot @maxencepajot.bsky.social · 02/09/2026

In the first experiment, we model an outlier detection task on quadrilaterals. Best models now correlate highly with human behaviour, above the symbolic model. Adding symbolic features did not improve alignment. What drives this is training set size, not architecture or number of parameters. 5/8

120

Maxence Pajot @maxencepajot.bsky.social · 02/09/2026

We compared human behaviour on three tasks to a wide range of models: CNNs and Vision Transformers, from 20M to 7B parameters, trained on datasets ranging from ImageNet-1k to LAION-2B, using classification, CLIP or DINO. 4/8

110

Maxence Pajot @maxencepajot.bsky.social · 02/09/2026

A lot has changed since. Campbell et al. (2024) found that recent large-scale models seem to behave in a more human-like way on geometric tasks. We revisit this, while also looking at the factors driving these abilities. 3/8

110

Maxence Pajot @maxencepajot.bsky.social · 02/09/2026

In 2021, Sablé-Meyer et al. showed that vision models did not model human behaviour on quadrilaterals accurately: their representations were in fact closer to those of baboons than of humans. Human behaviour was best captured by a symbolic model, encoding right angles, parallel sides, etc. 2/8

110

Maxence Pajot @maxencepajot.bsky.social · 02/09/2026

New AI models now show incredible math skill, solving problems that had stumped mathematicians for years. But what about basic geometric abilities? In our new paper in PLOS Computational Biology, we investigate this in vision models: doi.org/10.1371/jour... 1/8

doi.org

Can neural networks model the human perception of geometric shapes?

Author summary Recent neural networks have achieved remarkable success across a wide range of visual tasks, often rivaling or surpassing human performance. Yet, it remains unclear whether these system...

1356

Reposted by Maxence Pajot

Lucas Benjamin @lucaswbenjamin.bsky.social · 24/08/2026

Our new paper is now out in @pnas.org! How does the (baby) brain extract regularities from sequences from what-follows-what up to large-scale network structure? What if it all came from a single mechanism? www.pnas.org/doi/abs/10.1...

15216

Reposted by Maxence Pajot

Barbu Revencu @barbu-revencu.bsky.social · 12/06/2026

Representations of geometric shapes have syntactic structure w/ @maxencepajot.bsky.social and @standehaene.bsky.social is out & open-access in JEP:General doi.org/10.1037/xge0.... For an overview, see thread below!

doi.org

APA PsycNet

02711

Reposted by Maxence Pajot

Alireza Karami | علیرضا کرمی @alirezakr.bsky.social · 04/05/2026

Our new study (Debray*, Karami* et al.) explores how the brain encodes basic math concepts (integers, fractions, shapes). Samuel Debray led the 7T fMRI & behavioral work; I led the MEG. @danielavalerio.bsky.social @chrplr.bsky.social @standehaene.bsky.social www.biorxiv.org/content/10.6...

14118

Maxence Pajot @maxencepajot.bsky.social · 17/07/2025

Definitely, thanks for the suggestion!

000

Maxence Pajot @maxencepajot.bsky.social · 15/07/2025

In short: number frequencies show consistent patterns across languages. A simple model grounded in a Language of Thought—recursively using +, ×, and 1—captures these patterns remarkably well. This supports the idea that number concepts are built compositionally in the mind. doi.org/10.1016/j.co...

doi.org

Redirecting

040

Maxence Pajot @maxencepajot.bsky.social · 15/07/2025

With the cumulative model, we get an added benefit: it yields precise hypotheses about the mental representation of each number. For instance, 24 is most likely represented as 4×6, but also as 2×12, 3×8, or the successor of 23—in that order.

130

Maxence Pajot @maxencepajot.bsky.social · 15/07/2025

Our model was actually missing a crucial ingredient: approximation. Locutors experience a tradeoff between accuracy of the expressed quantity and length of the expression. Adding a parameter to account for this accurately models the whole curve.

120

Maxence Pajot @maxencepajot.bsky.social · 15/07/2025

Fitting our models to the data, we find that both models accurately captured the downward trend as well as some local peaks, yet it failed to do so for decade numbers. Why?

110

Maxence Pajot @maxencepajot.bsky.social · 15/07/2025

We developed two models to explain how often a number appears: - Shortest-path: only the simplest construction determines frequency - Cumulative: all valid constructions contribute to frequency, weighted by their complexity Here's how:

120

Maxence Pajot @maxencepajot.bsky.social · 15/07/2025

Building on this, we tested a new cognitive model based on a Language of Thought (LoT) introduced by Dehaene et al., 2025. The idea: number concepts are constructed in the mind using simple building blocks—1, +, and ×.

120

Maxence Pajot @maxencepajot.bsky.social · 15/07/2025

Some of this isn’t new. Dehaene & Mehler (1992) already showed that number word frequencies follow a ~1/n² law. They suggested that while cultural or environmental factors could explain some of the frequency curve, the psychological organization of number concepts must play a major role.

120

Maxence Pajot @maxencepajot.bsky.social · 15/07/2025

Using the Google N-grams database, we looked at the frequency of number words from 1 to 99 in 6 languages. Across the board, we found two patterns: 🔻 Frequency drops with size 📈 Local spikes at certain numbers—especially round ones and those with many small divisors

130

Maxence Pajot @maxencepajot.bsky.social · 15/07/2025

Some numbers seem to show up everywhere. Think of 10, 12, 24, 36... Others—like 26 or 34—don’t get the same attention.Why? In our new paper with @standehaene.bsky.social and @mathiassablemeyer.bsky.social, we argue it's because of how the mind builds number concepts. 🧵 doi.org/10.1016/j.co...

doi.org

Redirecting

23210

Reposted by Maxence Pajot

Barbu Revencu @barbu-revencu.bsky.social · 28/02/2025

Preprint alert (link below)! @maxencepajot.bsky.social, @standehaene.bsky.social, and I show that human adults, but not convolutional or transformer networks, encode geometric shapes in hierarchically structured representations. TL;DR: Geometric-shape representations have internal syntax! 1/8

22812