Self-Assembly Gets Automated in Reverse of ‘Game of Life’
1 min read
Summary
Alexander Mordvintsev and collaborators at Google Research in Zurich created a type of neural network that learns to behave like a cellular automaton, a long-studied type of miniature computational universe.
The key innovation was using a neural network to define the physics of the cellular automaton, instead of starting with rules and applying them.
The system starts with a desired pattern, which it can recreate even if it gets damaged, and learns from any failure to recreate it.
This enables complexity engineering — engineering the building blocks of a system so that they self-assemble into a desired form — and creates a new model for distributed computers in general.
Changes to the setup enable it to re-create damaged patterns, and model biological evolution.
The work may have lessons for robotics and provide new insight into how regeneration can occur in biology.
