Evaluating Neural Data Tokenizers: A Framework for Assessing Learned Representations of Spiking Activity
Federico D'Agostino*, Alex Gilbert*, Susanne Keller*, Jaivardhan Kapoor, Nicolas Reategui, Vedang Lad, Matthias Kümmerer, Jai Bhagat, Emin Orhan, Hasan Atakan Bedel, Nikos Karantzas, Fabian H. Sinz, Surya Ganguli, Jakob H. Macke, Sophia Sanborn, Alexander S. Ecker, Katrin Franke, Matthias Bethge, Andreas S. Tolias, Konstantin Friedrich Willeke
* Equal contribution
Neural tokenization and the dimensions used to evaluate what a representation preserves.
A brain foundation model is only as useful as the information its inputs preserve. Across ten datasets, this work builds and evaluates neural tokenizers along multiple axes, showing that accurate reconstruction does not necessarily produce useful downstream representations. It establishes a framework for treating tokenization as a central scientific and engineering decision in building models of neural activity.
02Generative modeling
In review
SpikeDiT: Generative neural system identification at spike-train resolution
Federico D'Agostino, Jai Bhagat, Jaivardhan Kapoor, Alex Gilbert, Zhuokun Ding, Anaya Gaelle Pouget, Surya Ganguli, Edgar Y. Walker, Fabian H. Sinz, Katrin Franke, Sophia Sanborn, Andreas S. Tolias, Matthias Bethge, Konstantin Friedrich Willeke
SpikeDiT combines neural tokens and stimulus conditioning to generate population spike trains.
SpikeDiT is a novel flow matching diffusion transformer model for generating neural spike trains at single millisecond resolution, combining a neural tokenizer with stimulus conditioning to capture both average neuronal responses and structured trial-to-trial variability. Across visual and auditory datasets, it offers a route toward richer models of neural populations, including ‘neural inpainting’ — completion of partially observed activity over missing time or missing neurons, without retraining.
Understanding computation & behavior
How intelligence is organized, learned, and expressed.
A pipeline for interpretable neural latent discovery
Jai Bhagat, Anaya Gaelle Pouget, Sara Molas-Medina
Sparse encoder–decoder models recover neural features, with hierarchical representations and extensions across time windows.
Large neural recordings contain rich structure, but useful predictions alone do not explain what a representation means. NLDisco uses sparse encoder–decoder models to discover latent features that can be related to behavior, the environment, and their contributing neurons. Tests on synthetic and real data show how this approach can turn high-dimensional activity into interpretable building blocks for studying neural computation.
Compressed Computation is not Computation in Superposition
Jai Bhagat*, Sara Molas-Medina*, Giorgi Giglemiani, Stefan Heimersheim
* Equal contribution
The original and simplified compressed-computation models, with loss across input sparsity.
Can an artificial neural network genuinely perform more independent computations than it has neurons? This work re-examines a proposed example and finds that its apparent advantage depends on an unintended mixing of inputs in the training targets. By separating that effect from the computation of interest, it sharpens the experimental standards needed to understand what neural networks actually learn.
05Learning & social behavior
In review
Reinforcement learning models of ethological mouse behavior reveals static and dynamic social preferences and policies
Anaya Gaelle Pouget*, Jai Bhagat*, Dario Campagner, Tiago Branco
* Equal contribution
Learned state values change with the other mouse’s location; example mouse, averaged across three model seeds.
How does sharing an environment change an animal’s decisions? Using long-duration recordings of freely behaving mice, this work combines imitation learning and offline reinforcement learning to describe movement policies and behavioral attractors in a changing foraging habitat. The models reveal how a partner’s current location shapes movement and shifts the balance between nesting and foraging, connecting natural social behavior to quantitative descriptions of action.
Making discovery possible
Tools and shared foundations for studying the brain.
Aeon: An open-source platform to study the neural basis of ethological behaviors over naturalistic timescales
D. Campagner*, Jai Bhagat*, G. Lopes*, L. Calcaterra, A. G. Pouget, A. Almeida, T. T. Nguyen, C. H. Lo, T. Ryan, B. Cruz, F. J. Carvalho, Z. Li, A. Erskine, J. Rapela, O. Folsz, M. Marin, J. Ahn, S. Nierwetberg, S. C. Lenzi, J. D. S. Reggiani, SGEN group – SWC GCNU Experimental Neuroethology Group
* Equal contribution
Aeon integrates modular habitats, synchronized hardware, and continuous data acquisition and analysis.
Understanding the brain requires observing behavior on the timescales over which animals actually live and adapt. Aeon brings together customizable habitats, synchronized acquisition, and integrated analysis to study self-guided behavior and neural activity continuously over extended periods. By combining experimental control with foraging, nesting, and social interaction, it opens a path toward testing theories of brain function in richer environments.
Reproducibility of in vivo electrophysiological measurements in mice
International Brain Laboratory, Kush Banga, Julius Benson, Jai Bhagat, Dan Biderman, Daniel Birman, Niccolò Bonacchi, Sebastian A Bruijns, Kelly Buchanan, Robert A A Campbell, Matteo Carandini, Gaelle A Chapuis, Anne K Churchland, M Felicia Davatolhagh, Hyun Dong Lee, Mayo Faulkner, Berk Gerçek, Fei Hu, Julia Huntenburg, Cole Lincoln Hurwitz, Anup Khanal, Christopher Krasniak, Petrina Lau, Christopher Langfield, Nancy Mackenzie, Guido T Meijer, Nathaniel J Miska, Zeinab Mohammadi, Jean-Paul Noel, Liam Paninski, Alejandro Pan-Vazquez, Cyrille Rossant, Noam Roth, Michael Schartner, Karolina Z Socha, Nicholas A Steinmetz, Karel Svoboda, Marsa Taheri, Anne E Urai, Shuqi Wang, Miles Wells, Steven J West, Matthew R Whiteway, Olivier Winter, Ilana B Witten, Yizi Zhang
A shared experimental pipeline, apparatus, and recording target make cross-laboratory comparisons possible.
A science of the brain depends on results that hold beyond a single laboratory. Across 121 experimental replicates in ten labs, this collaboration examines where standardized neural recordings agree and where variability persists. It identifies electrode targeting and analysis choices as important sources of variation, and develops quality-control criteria that make shared, cumulative neuroscience more reliable.
Neuropixels 2.0: A miniaturized high-density probe for stable, long-term brain recordings
Nicholas A Steinmetz*, Cagatay Aydin*, Anna Lebedeva*, Michael Okun*, Marius Pachitariu*, Marius Bauza, Maxime Beau, Jai Bhagat, Claudia Böhm, Martijn Broux, Susu Chen, Jennifer Colonell, Richard J Gardner, Bill Karsh, Fabian Kloosterman, Dimitar Kostadinov, Carolina Mora-Lopez, John O'Callaghan, Junchol Park, Jan Putzeys, Britton Sauerbrei, Rik J J van Daal, Abraham Z Vollan, Shiwei Wang, Marleen Welkenhuysen, Zhiwen Ye, Joshua T Dudman, Barundeb Dutta, Adam W Hantman, Kenneth D Harris, Albert K Lee, Edvard I Moser, John O'Keefe, Alfonso Renart, Karel Svoboda, Michael Häusser, Sebastian Haesler, Matteo Carandini, Timothy D Harris
* Equal contribution
Smaller probes and headstages, with denser recording sites and support for two probes per headstage.
Learning and memory unfold over days and months, demanding tools that can follow the same neurons through time. Neuropixels 2.0 combines miniaturized, high-density probes with motion-correction methods to enable stable recordings during freely moving behavior. Demonstrations across six laboratories show how these advances extend the scale and duration at which neural activity can be studied.
Rigbox: An open-source toolbox for probing neurons and behavior
Jai Bhagat*, Miles J Wells*, Kenneth D Harris, Matteo Carandini, Christopher P Burgess
* Equal contribution
Signals connects experimental timing and hardware inputs to stimuli and device outputs.
Good experiments need software that makes complex ideas practical to test. Rigbox provides a modular toolkit for behavioral neuroscience, bringing together hardware control, synchronized data streams, stimulus presentation, and experiment management. Its Signals framework makes behavioral tasks easier to express and extend, supporting more flexible and reproducible experimental work.
LSTM neural networks for LFP event detection and classification in the rodent hippocampal-cortical network
Jai Bhagat, Hector Penagos, Matthew A. Wilson
Training curves and precision–recall comparisons for spindle detection and spindle–ripple classification.
Brief hippocampal ripples and cortical spindles offer a window into the coordination of memory-related activity. This work uses LSTM networks to detect and classify these events from their temporal signatures, reducing reliance on labor-intensive manual scoring. Preliminary results point towards meaningful classification of these LFP events from trained LSTM networks.