
Our “Neuro-AI” paper has just been published in Cell. It describes a new deep learning-based classifier for identifying different cell types in Neuropixels recordings: doi.org/10.1016/j.cell.2025.…


Our “Neuro-AI” paper has just been published in Cell. It describes a new deep learning-based classifier for identifying different cell types in Neuropixels recordings: doi.org/10.1016/j.cell.2025.…

Neuropixels Opto is out! 🎉🥳🍾🚨 Congratulations to a fantastic team! nature.com/articles/s41592-0…

If you’re at Cosyne, catch @maxime_beau at the workshop on “Bridging the gap between cell types and spike trains"! Monday 16th March: celltypestospikesworkshop.gi…

The culmination of 8 years of teamwork: the IBL brain-wide map and prior papers have been published in @Nature today! We are so happy to be part of this great collaboration 🎉🎉🎉 doi.org/10.1038/s41586-025-0… doi.org/10.1038/s41586-025-0…

Open-access link: urldefense.proofpoint.com/v2…

This is real team science: 5 labs expert in recordings from mouse and monkey cerebellum joined forces to make the project happen. Thanks to a great team, led by @1DanaCohen1, @Steve_Lisberger, @dkneurolab , @theHullLab, @Neuralcomplab, and Javier Medina.

Every step of our approach is generalizable to other systems. We hope that our work will inspire other groups to acquire their own ground-truth database to crack cell type classification in other brain regions.

This is a game-changer, as we can now simultaneously track the activity of 5 different neuron types with silicon probes such as Neuropixels during behaviour, without the need for genetic engineering.

The algorithm is open source and freely available on our website (c4-database.com) so you can use it for analysing your own recordings. The site also provides free access to our entire dataset and a set of tools for visualizing neurons in the library.

The VAE’s latents were then used to train a supervised multi-layer perceptron. The final algorithm is highly accurate, yielding > 95% reliable prediction of cell types in both mouse and monkey cerebellar recordings.

Key to success was the use of a hybrid network. Self-supervised variational autoencoders trained on unlabelled neurons were used to encode the waveform shape and firing statistics (captured by autocorrelograms) into a low-dimensional latent space (the VAE’s “bottlenecks”).

We created a carefully validated “ground truth” library of over 200 optogenetically tagged cerebellar neurons and used this to train a classifier with semi-supervised generative modelling to recognize neuron types based on their electrophysiological signatures.

We tackled a decades-old problem in systems neuroscience: the identification of neurons in extracellular recordings in vivo.

Congratulations to the team! @omgauld, @paqio, @llerussell, @hwpdalgleish, Maya Iuga, @Fran_Sacadura, @ArndRoth & @BeverleyAClark

This result suggests that non-coding neurons can still convey important information, which may have significance for builders of brain-machine interfaces as it widens the pool of neurons available to enhance behavioral performance.

Second, the recruitment of this non-coding pool by optogenetic stimulation can enhance perception. This means that this latent pool of neurons (the “dark matter”) can actually be useful for sensory processing!