Tom George

What are the brain’s “real” tuning curves? Our new preprint "SIMPL: Scalable and hassle-free optimisation of neural representations from behaviour” argues that existing techniques for latent variable discovery are lacking. We suggest a much simpl-er way to do things. 1/21🧵

A little overdue but happy to announce that I 1) Got my PhD🎓 2) Started a postdoc and CNS Fellowship with Guillaume Lajoie and Blake Richards (@g_lajoie_ @tyrell_turing) Come find me at @Mila_Quebec in Montreal working on new things NeuroAI!

A great article from SWC on our neural data analysis technique (now accepted into ICLR). More updates to come soon...plus will be @CosyneMeeting presenting this too!

Application deadline has been extend. Two more weeks to apply...don't miss out!!

🌍🧠💻Applications are well and truly open for the third CaMinA. African nationals studying biology, medicine, engineering, maths etc. can, and should, apply for this summer school in Zambia. Neuro and ML are changing the world and now I the time to get into them, please RT!

@GabyMohamady @OxExpPsy amazing!!!!

Delighted to share that I’ll be joining @OxExpPsy as a group leader/Faculty next spring. My group will investigate how our cells build models of the world and of our internal goals, and how this goes wrong in disorders like schizophrenia.

@CortexClub loving the hype

Our next CorTech talk is scheduled for tomorrow, Monday December 9th at 4pm with @TomNotGeorge in the Sherrington Library. Title: Modelling Locomotion; From Continuous Environments to Hippocampal Neurons with the RatInABox package. Post-talk, we will head to the Lamb and Flag!

@vineettiruvadi ((also I'm not saying it shouldn't still be used - in many cases its still quite useful...just that perhaps we could look a bit deeper))

@vineettiruvadi I think we should! (my experience with a lot of contemporary work in HPC is that tuning curves == spikes fitted to behaviour is still dominant)

CaMinA is back for its 3rd year...this time we're going to beautiful Zambia🇿🇲! I'm proud to see this course grow and bring together the smartest students across Africa with leading neuro institutes like @AllenInstitute and @SWC_Neuro Applications open soon, please share widely!

🌍Exciting news! The 2025 TReND-CaMinA Course will be held in Lusaka, Zambia 🇿🇲 from July 7th–23rd. Dive into computational neuroscience and machine learning with us! 📅 Applications open: December 15th 🔗More info: trendinafrica.org/trend-camina/ Stay tuned & spread the word! 🧠✨

@rory_bedford ^ this👌🏼... 100%

What are the brain’s “real” tuning curves? Our new preprint "SIMPL: Scalable and hassle-free optimisation of neural representations from behaviour” argues that existing techniques for latent variable discovery are lacking. We suggest a much simpl-er way to do things. 1/21🧵

@EliSennesh Good question! In the space of a tweet that might be too much to answer but can I direct you to section 2 of the paper where I think you'll find the answer tomge.org/papers/simpl/s…

@TomNotGeorge What's your prior and model structure by which you connect spikes, behavior, and the latent variable you impute?

@zilong_ji nice I hadn't seen this one! Muller and Kubie 1989 did the same thing with place fields. Was an inspiration for SIMPL (which is like an automatic and much less constrained version of this idea)

@TomNotGeorge This is cool and reminds me of the Ouchi & Fujisawa 2024 paper, where they plot the spikes with future position instead of the current position, they got a nicer and cleaner tuning field of a grid cell.

@BenDichter @NeurodataWB @DANDIarchive Hi Ben, this is exactly the plan so thanks for bringing the @DANDIarchive datasets to my attention. I'd love your help here! We should chat...

@TomNotGeorge I'd love to see this analysis applied to reanalyze the mountain of diverse, open, state-of-the-art hippocampal datasets in @NeurodataWB on the @DANDIarchive from labs of Frank, Buzsaki, Moser, Giocomo, and more. I'd be happy to help someone get started 🙂

@sai_prasanna bsky.app/profile/tomnot… ^^ I just joined. Could you add me to some lists!?

@TomNotGeorge Btw b sky 🦋 is getting popular among AI and neuroscience people, in case you are looking to shift! There are lists to which you can be added and in which you can find other neuroscience folks

@sai_prasanna thanks! Yes SIMPL will be most applicable to relatively "simple"/low-D/continuous latent spaces (of which there are a few obvious contenders in the brain)

@dlevenstein I hate twitter these days. Is it time to switch?

@TomNotGeorge Yup. All makes sense. One last Q: what’s a guy gotta do to get this thread over on bsky? 😜😜😇

@dlevenstein • Ultimately its a trade off…we assume Gaussian (therefore unimodal) posteriors which therefore allows Kalman filtering which is a SUPER fast E-step. If you want multimodal posteriors (ie encode multiple positions) then expect big slow-downs :(((

@TomNotGeorge I do wonder about the foundational assumption here: that there's a single represented variable in the network and that variable corresponds to a spatial position. I see why you would make that assumption but it's a strong one.

At the risk of rambling I'll end the thread here and perhaps do a deeper dive in the future. Give it a read (or better, try it on your data) and let us know your thoughts! tomge.org/papers/simpl/ 21/21

This isn’t cheating, behaviour has always been there for the taking and we should exploit it. If we ignore behaviour and initialise randomly SIMPL still works but the latent space isn’t smooth and “identifiable”. This is certainly something to consider…. 20/21