Compact deep neural network models of the visual cortex Nature - Parsimonious deep neural network models can be used for prediction of visual neuron responses.

DNN models of the brain are getting bigger. Are we replacing one complicated system in vivo with another in silico?

In new work, we seek the *smallest* DNN models of visual cortex, balancing prediction with parsimony.

It turns out these compact models are surprisingly small!

rdcu.be/e5H8G

Now accepted at #ICLR 🎊🙂 See you in Brazil! 🇧🇷

Happy 2nd science birthday to the #KempnerInstitute! 🧠In our 2nd full year we reached 420+ members, 350+ papers, 3,200+ event attendees, and 3M+ GPU hours. See our strides in #intelligence research!

Read the 2025 Annual Report: bit.ly/3LPSVmB
@shamkakade.bsky.social @blsabatini.bsky.social

@tuthill.bsky.social @jan-ache.bsky.social @srinituraga.bsky.social @dddavi.bsky.social @bingbrunton.bsky.social

Finally, I cannot end my words without thanking the open-source ecosystem that made this possible: OpenSim, MuJoCo, MyoSuite, and many more.

Code links:
• OpenSim optimization: github.com/gizemozd/neu...
• MuJoCo imitation learning: github.com/gizemozd/Fly...

Huge thanks to my co-first author Chuanfang Ning, whose master’s thesis sparked this project, and to all co-authors: Jasper S. Phelps, Sibo Wang-Chen (@wangchen.bsky.social), Guy Elisha, Alexander Blanke, Auke Ijspeert, Pavan Ramdya (@ramdya.bsky.social).

While we only optimized front-leg muscles, we provide mid & hind-leg scaffolds so their parameters can be fit next using the same pipeline. Our model bridges motor neurons and joints, paving the way for plugging more realistic neural networks derived from the connectome into embodied agents.

Muscles don’t act alone; passive joint properties matter too. We ported the model to MuJoCo (via MyoConverter) for large-scale, fast simulation and set up imitation learning with muscle-driven control. We found that passive properties in the joints stabilize control and speed up learning.

We've got moments arms covered—what about muscle force? Unlike in humans, most fly muscle parameters are unknown. So we built an optimization pipeline in OpenSim to identify Hill-type parameters to produce measured fly kinematics. This revealed coordinated, testable muscle synergies across behaviors

We began by tracing muscle fibers, origins/insertions, and paths from high-resolution X-ray scans across specimens. This allowed us to recover moment arms around each joint center and cross-sectional area as a prior on muscle strength.

In plain terms: torque = force x moment arm.

🪰 How do dozens of tiny fly muscles cooperate to move a leg?

We’re excited to share the first 3D, data-driven musculoskeletal model of Drosophila legs based on Hill-type muscles, running in OpenSim and MuJoCo simulation environments.

Preprint: arxiv.org/abs/2509.06426

Thanks, Ben! Looking forward to being there :)

Thank you so much!! I hope everything is going well for you! :)

Big life update: I’m super excited to be joining the Kempner Institute as a Research Fellow!
If you’re curious about my research plans or just want to connect, please reach out! 😊

Please RT🙏

Reach out if you want to help understand cognition by modelling, analyzing and/or collect large scale intracortical data from 👩🐒🐁

We're a friendly, diverse group (n>25) w/ this terrace 😎 in the center of Paris! See👇 for + info about the lab

We have funding to support your application!

📽️Recordings from our
@cosynemeeting.bsky.social
#COSYNE2025 workshop on “Agent-Based Models in Neuroscience: Complex Planning, Embodiment, and Beyond" are now online: neuro-agent-models.github.io
🧠🤖

COSYNE 2025 Rajan lab posters. Friday, March 28 - Poster Session 2: 2-043: Emergent small-group foraging under variable group size, food scarcity, and sensory capabilities by Zhouyang (Hanson) Lu, Satpreet H Singh, Sonja Johnson-Yu, Aaron Walsman, Kanaka Rajan 2-058: 'Modeling rapid neuromodulation in the cortex-basal ganglia-thalamus loop' by Julia Costacurta and Yu Duan (co-first), John Assad, Kanaka Rajan and Scott Linderman (co-senior) 2-060: 'Measuring and Controlling Solution Degeneracy across Task-Trained RNNs' by Ann Huang, Satpreet Singh, Kanaka Rajan Saturday, March 29 - Poster Session 3: 3-020: 'ForageWorld: RL agents in complex foraging arenas develop internal maps for navigation and planning' by Ryan Badman, Riley Simmons-Edler, Joshua Lunger, John Vastola, William Qian, Kanaka Rajan 3-109: 'Inhibition-stabilized disordered dynamics in mouse cortex during navigational decision-making' by Siyan Zhou, Ryan Badman, Charlotte Arlt, Kanaka Rajan, Christopher Harvey

Big showing from the Rajan Lab at @cosynemeeting.bsky.social!

We have posters on everything from multi-agent social foraging to neuromodulated neural networks. Catch us in Poster Sessions 2 & 3 🧠🤖

#Cosyne2025 #NeuroAI #CompSci #neuroskyence

Sadly, I won’t be there in person this year because of visa issues :(( (yep, they’re real and they suck)... But two of my amazing co-organizers will be there: @satpreetsingh.bsky.social and @chingfang.bsky.social

Only 4 days to go until our workshop!! 🪰🐁🤖
If you're at COSYNE, don't miss out on incredible talks and inspiring panel discussions at "Agent-Based Models in Neuroscience: Complex Planning, Embodiment, and Beyond" on March 31 :)

Check out the latest schedule: neuro-agent-models.github.io

🚨 SUPER excited to announce our Cosyne workshop on Neuro Agents! 🤖🐁🪰🧠 We have an incredible lineup of speakers, check out the program! neuro-agent-models.github.io

See you in Canada! 🇨🇦

#Cosyne #Cosyne2025 #NeuroAI

Congrats!! I will be there as well, presenting another biomechanics paper ☺️

I wrote an introduction to RL for neuroscience last year that was just published in NBDT: tinyurl.com/5f58zdy3

This review aims to provide some intuition for and derivations of RL methods commonly used in systems neuroscience, ranging from TD learning through the SR to deep and distributional RL!

Thank you so much, John! I would love to hear your thoughts!

As you unwrap your holiday presents, consider how you coordinate your fingers and limbs.
@gzmozd.bsky.social identified fly brain networks for body part coordination through experiments, biomechanical modeling, connectomics, and neural network simulations ! 🤖
www.biorxiv.org/content/10.1...

10/ Big thanks to our amazing collaborators and the incredible fly community for creating the open-source tools that made this work possible. 🙌 #Neuroscience #MotorControl #Drosophila #Connectome @neuroxepfl.bsky.social @fly-eds.bsky.social @flywire.bsky.social

Centralized brain networks underlie body part coordination during grooming Animals must coordinate multiple body parts to perform important tasks such as grooming, or locomotion. How this movement synchronization is achieved by the nervous system remains largely unknown. Her...

9/ So next time you see a fly grooming itself or you try multitasking, take a moment to appreciate the magic of coordination. Check out our preprint! 🪰🧠 www.biorxiv.org/content/10.1...

8/ The fly’s strategy enables robustness yet flexibility, thus it may be a common blueprint for movement across species—or even for other behaviors in flies. 🐁🐱🦎

7/ Recurrent excitation: Drives non-groomed antennal pitch movements and keeps other motor networks in sync. ⚡️
Broadcast inhibition: Suppresses targeted antennal movement to prevent conflicting actions. ⛔️

6/ To understand this better, we simulated the grooming network and ran a computational neural activation screen. Two key circuit motifs emerged as the stars of this coordination process:

5/ Think of it as an elegant engineering solution: these central neurons enable flexibility, allowing any brain region to initiate or stop the behavior. 🛠️