Mariia Drozdova

PhD student @ University of Geneva | Paradigms of Intelligence Team @ Google

[Mayalen Etcheverry @mayalen_etc]

Can complex multi-step reasoning emerge purely from cells that only talk to immediate neighbors?

Happy to share our paper “Reasoning with Neural Cellular Automata (NCA)”, from our team at Google, Paradigms of Intelligence 🧵👇

[Mariia Drozdova @MariiaD_ML]

Neural Cellular Automata can solve reasoning tasks even though each cell only communicates with its neighbors! Every cell runs the same network, enabling generalization: trained on small mazes, NCAs solve huge ones; trained on easy Sudoku, they solve hard ones! 🎉

[Mayalen Etcheverry @mayalen_etc]

Can complex multi-step reasoning emerge purely from cells that only talk to immediate neighbors?

Happy to share our paper “Reasoning with Neural Cellular Automata (NCA)”, from our team at Google, Paradigms of Intelligence 🧵👇

[Mariia Drozdova @MariiaD_ML]

7/7) All in all, diffusion provides the stochastic training curriculum, recurrence carries the iterative computation for reasoning, and noise allows for test-time exploration.

Paper: https://t.co/FL4Mhi7DuZ

[Mariia Drozdova @MariiaD_ML]

1/7) What happens when we add a persistent memory state and remove timestep conditioning from a denoising model?

We obtain an anytime iterative solver!

With less than 250k-params, we obtain a 99.9% accuracy on Sudoku-Extreme and 98.3% on Maze-Unique.

https://t.co/FL4Mhi7DuZ

[Mariia Drozdova @MariiaD_ML]

6/7) Once the model finds a unique solution, it forms a stable attractor.

When we continue iterating upon the solution after converging, the solution remains stable in more than 99% of trajectories.