Apurva Ratan Murty
banner
apurvaratan.bsky.social
Apurva Ratan Murty
@apurvaratan.bsky.social
Neuroscientist. Assistant Professor of Cognition and Brain Science @ GeorgiaTech Prev: Postdoc @ MIT, PhD @ IISc Bangalore
www.murtylab.com
#VSS2025 Tomorrow in Talk Room 2 at 9.15AM. Mayukh Deb will be presenting our work on TopoNets! You can read more about it at toponets.github.io. But there will be a lot more!
May 18, 2025 at 9:09 PM
Come listen to Ben Lahner present a new combined fMRI dataset - MOSAIC. Happening soon at 1145 in Talk Room 2.
May 18, 2025 at 3:28 PM
and finally. Our experience with
@iclr-conf.bsky.social reviews was pretty great! The reviews were constructive, reviewers actively involved during discussion and encouraging in their response.. How refreshing! Appending a particularly favorite reviewer comment
January 30, 2025 at 3:31 PM
We also find units clustered according to distinct temporal integration windows a la recent results from Ev Fedorenko's lab. (thanks to Sam Normal Haigere's lab for cool methods to find these units in GPTs!)
January 30, 2025 at 3:27 PM
TopoNet-Vision models recapitulate topographic signatures in visual cortex like category selectivity for faces, bodies, scenes and large-scale organizational biases a la @taliakonkle.bsky.social 's work!
January 30, 2025 at 3:26 PM
Even though the focus of this paper was mostly on the engineering and the topo algorithm, we couldn't hold back and had to do some neuro tests as well! :) TopoNets predict responses in monkey neural data (BrainScore) and human brains (NSD, see paper) very well!
January 30, 2025 at 3:25 PM
We also find that inducing topography into neural networks makes the model features way more sparse and parameter efficient. We show this using both L1 pruning and downsampling features!
January 30, 2025 at 3:24 PM
Now equipped with so many topo models (!), we can address a few open Qs. Q1. What is the representational consequence of topography for models (and brains)? We find that topography, not model performance, lowers rep. dimensionality in TopoNets.
January 30, 2025 at 3:24 PM
The result? Toponets! TopoNets are the highest performing supervised topographic models to date! How high performing? TopoRN18s and TopoViTs with > 65% on ImNet, TopoRN50s >70% on ImNet! We have TopoGPT-Neos and Topo-NanoGPTs as well! It is super easy to implement and scale!
January 30, 2025 at 3:23 PM
Hey Bsky friends on #neuroskyence! Very excited to share our
@iclr-conf.bsky.social paper: TopoNets! High-performing vision and language models with brain-like topography! Expertly led by grad student Mayukh and Mainak! A brief thread...
January 30, 2025 at 3:23 PM