I’ll be at @iclr-conf.bsky.social presenting our work “Credit-based self organizing maps: training deep topographic networks with minimal performance degradation” — if you're into #NeuroAI, cortical topography, vision, or just up for a good convo, drop by and DM me! 🧠✨

9/ Finally, we show that compared to other baselines including classic SOM (AB-SOM) and TDANN, CB-SOM shows substantially higher alignment with neural activity in the visual cortex of both macaques (Brain-Score) and humans (NSD).

8/ Similarly, the model’s deeper layers replicated the organization of the higher-level visual cortical regions according to category-selectivity, resulting in category-selective filter clusters in these layers.

7/ We also observed an exponentially decaying pairwise correlation between filters as a function of their distance on the simulated cortical sheet.

6/ Our model replicated the topographical organization seen in the primary visual cortex topographical organization, including smooth transitions in selectivity according to stimulus orientation and spatial frequency.

5/ With this approach, our ResNet-18 model achieved a substantial performance boost while preserving topographical organization.

4/ We proposed an alternative approach based on self-organizing maps, Credit-Based Self-Organizing Maps(CB-SOMs), where the competition mechanism is guided by the importance of the assigned credit to each weight – i.e. filters with the highest gradient norm in each conv layer.

I am excited to share that my undergraduate research has now been accepted to #ICLR2025.
📎: openreview.net/pdf?id=wMgr7...
🧑‍💻: github.com/BashivanLab/...

Tweeprint🧵⬇️