iiinova — where your digital self keeps evolving

Fabian Sinz

@sinzlab.bsky.social

534 followers 753 following 40 posts

NeuroAI, Deep Learning for neuroscience, visual system in mice and monkeys, computational lab based in Göttingen (Germany), sinzlab.org

Posts Replies Media

Fabian Sinz @sinzlab.bsky.social · 24/04/2026

🏆 OmniMouse sets a new state of the art in the Sensorium 2022/2023 competitions (sensorium-competition.net) and matches or exceeds specialized baselines across nearly every other evaluation regime: • Neural response prediction • Neural forecasting • Behavioral decoding

110

Fabian Sinz @sinzlab.bsky.social · 24/04/2026

Data scaling tells the opposite story. Nested subsets of 8 → 16 → 32 → 64 → 323 sessions: performance improves monotonically for every task and every model size.

210

Fabian Sinz @sinzlab.bsky.social · 24/04/2026

Across decoding of gaze position, pupil size, and running speed, performance improved smoothly with compute budget, reminiscent of classic scaling-law behavior. Larger models consistently achieved higher single-trial correlations, with saturation at the largest model scale.

100

Fabian Sinz @sinzlab.bsky.social · 24/04/2026

For neural response prediction (forecasting, population prediction, stimulus-conditioned encoding), gains level off around ~80M parameters. Loss curves saturate or start to overfit. At this data scale, model capacity doesn't appear to be the limiting factor.

110

Fabian Sinz @sinzlab.bsky.social · 24/04/2026

We trained a family of OmniMouse models, ranging from 1M to 300M parameters on nested subsets of 8 → 323 sessions, looking at both model and data scaling. Neural encoding and behavioral decoding show quite different scaling behavior.

100

Fabian Sinz @sinzlab.bsky.social · 24/04/2026

OmniMouse is a multi-modal, multi-task transformer: single neuron tokenization, a ViT for videos, and behavior traces are fed into a unified model. Through flexible masking, our model supports neuronal encoding & decoding, forecasting, or any combination of the three.

110

Fabian Sinz @sinzlab.bsky.social · 24/04/2026

Scaling data and model size has driven progress in language and vision. We wanted to know whether the same scaling holds for neural activity, and which axis is the bottleneck. So we trained a single model on all of the data.

100

Fabian Sinz @sinzlab.bsky.social · 24/04/2026

The dataset that we’re releasing today is among the largest ever released in neuroscience: • 3.1M segmented neurons • 78 mice • 328 sessions • 150B+ neural tokens • Naturalistic + parametric visual stimuli • Behavioral data

152

Fabian Sinz @sinzlab.bsky.social · 24/04/2026

🧠Introducing OmniMouse: One of the largest datasets in neuroscience along with a systematic study of scaling properties of brain models Co-led by 🤩 K. Willeke, @pollytur.bsky.social and @alexrgil14.bsky.social Models trained on up to 3M neurons, >150B tokens

14326

Fabian Sinz @sinzlab.bsky.social · 21/11/2025

🔬 Exciting PostDoc Opportunity! 🐁🧠 We - the @sinzlab.bsky.social (sinzlab.org) and @trose-neuro.bsky.social (troselab.de) Labs - are seeking an experimental postdoc to work at @unibonn.bsky.social with cutting-edge miniature 2-photon microscopy and gaze tracking in freely behaving mice.

2166