Will Dorrell

New pre-print with Peter Latham, @behrenstimb & @jcrwhittington: arxiv.org/abs/2209.15563 We formalise path-integrating representations of space, explain why the optimal representation is multiple modules of grid cells, make testable predictions, and apply to non-2D (1/12)

Cosyne Viewing Parties Visas, costs, care, & environmental concerns all limit Cosyne attendance. Luckily, the talks are livestreamed; but watching alone is the high road to an aneurism. So: viewing parties! Gather regionally to watch Cosyne talks! Info: shorturl.at/3DHZX

📢 Apply now for the 7-week #Mathematics #SummerSchool in London! You will develop mathematics skills and intuition necessary to enter the #TheoreticalNeuroscience or #MachineLearning field. ℹ️ ucl.ac.uk/gatsby/study-a… Please help spread the word!

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🧵

New @tweeprint! Compact deep neural network models of visual cortex B. Cowley, P. Stan, J. Pillow*, M. Smith* tiny.cc/dmxhvz Task-driven DNN models nicely predict neural responses but have millions of params---next to impossible to explain. Do they need to be so large?

Original thread here: twitter.com/dorrell_will/s… Thanks to help from at @RobertTLange we have now added black box-metalearning! This means we can have a plausible method to extract inductive bias from living animals!!! Crazy experimentalist collaborators wanted...

Yes! Excited to say our work is at ICML, even if we aren't (too far, too expensive...) But don't worry! Read all about it: arxiv.org/abs/2211.13544 Or watch: recorder-v3.slideslive.com/?share=84188&s… All with new added bonus content (see below) do get in touch if you have questions/complaints!

Wild! These flies wander around on the snow. They're able to avoid freezing by self-amputating legs as they begin to crystallise!!!

Woop let's disentangle things! We modify an autoencoder's latent space to be compositional (using discretisation & regularisation) leading to networks that learn to distill-out meaningful factors of variation in datasets Super project led by the enormously talented @kylehkhsu!

@yixin_lin_ @TrentonBricken Ah damn! Okay, smart move - look forward to hearing what it's like!

@TrentonBricken @dorrell_will thanks @TrentonBricken ! @dorrell_will, I'm going to Uganda the weekend before the conference, but happy to meet up at the conference, London, etc :)

Anyone going to ICLR looking for a gorilla trecking buddy?? :) Planning to head to Uganda straight after the conference, and looking for a posse.

Excited to have this "tiny paper" accepted into ICLR... In it I distil a theory for the functional role of theta sequences as biologically plausible manifestations of eligibility traces, speeding up learning. You should read it, it's literally two pages🧵openreview.net/pdf?id=vd16AYb…

I think this might be the most beautiful science I've seen in a while

In sum, how design choices affect generalisation is a promising angle for understanding biological circuits, and our tool broadens the use-cases for this method. Hopefully it will be useful! Thanks to Peter Latham and Maria Yuffa (an undergrad!) for pivotal contributions! /end

The bottleneck on our tool is interpreting the easy-to-generalise functions, which is hard in high-dimensional spaces. Low-dimensional toy models, or carefully chosen function have been useful, but in general more tools are needed to interpret the results. (11/12)

New pre-print!! With Maria Yuffa & Peter Latham: arxiv.org/abs/2211.13544 We develop a tool to meta-learn functions that neural networks find easy to generalise and use it to assign a normative role to features (connectome, learning rules, etc.) of biological circuits. (1/12)