Ayush Patravali

Search the site

Jump to a page, project, paper or journey step

AI / MLNational Institute of Advanced StudiesMay 2025 to Mar 2026

Neurochaos learning, extended to graph data

Neurochaos learning learns from small samples on modest compute, but had only been tested on flat and time-series data. We extended it to graph data for the first time.

My roleResearch intern, second author

Preprint on arXiv (Feb 2026). Macro-F1 0.844 on Cora, falling to 0.265 on Actor as neighbours' labels diverge.

  1. Cora0.844
  2. PubMed0.76
  3. CiteSeer0.732
  4. Wisconsin0.57
  5. Cornell0.43
  6. Squirrel0.34
  7. Actor0.265
Test macro-F1 on seven graph datasets, from Table 2 of the paper. Strong where neighbours share labels, weaker as they don't.

The question

Can neurochaos learning, a brain-inspired method that works on small data and ordinary compute, classify nodes in a graph?

What we did

Mean neighbour aggregation turns each node’s features and its neighbourhood into one vector. ChaosNet passes it through chaotic GLS neurons and extracts firing time, firing rate, energy and entropy, and a cosine-similarity classifier assigns the label. We tested seven datasets, from Cornell (183 nodes) to PubMed (19,717 nodes), and swept the q, b and epsilon hyperparameters.

Results (test macro-F1)

Cora 0.844, PubMed 0.76, CiteSeer 0.732, Wisconsin 0.57, Cornell 0.43, Squirrel 0.34, Actor 0.265. Strong where neighbours tend to share a label, weaker as that fades.

Paper

Honna, Patravali, Nagaraj and Narendra, “Linked Data Classification using Neurochaos Learning”, arXiv:2602.16204, 2026.

arXiv:2602.16204CodeBack to the journey