This year at #CCN25 we showed the importance of OOD evaluation to adjudicate between brain models. Our results demonstrate these trivial but key facts :
- high encoding accuracy ≠ functional convergence
- human brain ≠ NES console ≠ 4-layers CNN
- videogames are cool

w/ @lune-bellec.bsky.social 🙌

In neuroscience, we often try to understand systems by analyzing their representations — using tools like regression or RSA. But are these analyses biased towards discovering a subset of what a system represents? If you're interested in this question, check out our new commentary! Thread:

🚨 New preprint alert!

Excited to share our latest work on alpha/beta activity, eye movements, and memory.

Across 4 experiments combining scalp EEG/iEEG with eye tracking, we show that alpha/beta activity directly reflects eye movements, and only indirectly relates to memory.

👇 Highlights (1/7):

🧠 Paper out!

We investigated how hippocampal and cortical ripples support memory during movie watching. We found that:

🎬 Hippocampal ripples mark event boundaries
🧩 Cortical ripples predict later recall

Ripples may help transform real-life experiences into lasting memories!

rdcu.be/eui9l

In summary, this shared representational geometry provides a powerful framework to study the brain's functional organization and trace how information is routed through the cortex.

We welcome your questions!

💻 Code: github.com/memory-forma...

(8/8)

Diving deeper into the LOTC hub's social vs non-social component:

Alignment across brains along the lateral stream (EVC→LOTC) is present only when viewing social scenes (with people or animals).

This supports its proposed role as a specialized "third visual pathway" for social perception.

⬇️ (7/8)

So what information does each hub actually encode?

Using KMCCA, we studied the primary dimension that organizes each hub's information:

👁️ EVC: Low-level visual features
🏞️ Ventral Hub: Scene & object structure
👨‍👩‍👧‍👦 LOTC Hub: Social vs. non-social content

⬇️ (6/8)

Vision DNNs capture this shared geometry, with each brain hub showing a different layer alignment profile:

🧠 Early visual ↔️ Shallow DNN layers
🧠 Ventral hub ↔️ Mixed DNN layers
🧠 LOTC ↔️ Deep DNN layers

Language Models only align with the high-level LOTC hub.

⬇️ (5/8)

This shared representational geometry is so consistent across people that we could map a whole-brain connectivity network based on it, revealing interactions between visual, memory and prefrontal areas.

⬇️ (4/8)

We identified 3 cortical hubs with highly consistent representations across all individuals:
📍 Early visual cortex (V1–V4)
📍 Ventral hub (scene/object areas ~PPA)
📍 LOTC Hub (hMT+/TPOJ)

These hubs form two pathways:
- Classical ventral stream (EVC → Ventral)
- Lateral stream (EVC → LOTC)

⬇️ (3/8)

Using Representational Similarity Analysis (RSA) on fMRI data from people viewing diverse scenes, we measure:

- Inter-subject RSA: Are visual representations shared across individuals?
- Brain-Model RSA: Is this shared information low-level (visual) or high-level (semantic)?

Methods ⬇️ (2/8)

🧠🚨 How does the brain represent what we see? Is visual input transformed to form these representations in similar ways across people and even AI models like DNNs?

We explore these questions using fMRI and large-scale representational alignment analyses.

🔗 arxiv.org/abs/2507.13941

Thread👇 (1/8)

Deep learning models and brains share fascinating parallels in their ability to process and instantly integrate new knowledge.

Join us this year at ICON 2025 to discuss how sudden learning emerges across artificial and biological systems! 🧠🤖

Very happy to announce that the latest work "Anticipating multisensory environments: Evidence for a supra-modal predictive system" from @alepebel.bsky.social @fuentemilla.bsky.social and myself has been published in Cognition!

www.sciencedirect.com/science/arti...