I reached out to the ML Collective about tackling the ARC challenge, a dataset of grid-based abstraction and reasoning puzzles.
I ended up building a model based on neural cellular automata and presenting the results to other researchers in the community.
The model solved 14 out of 262 tasks, while a baseline convolutional model with residual connections solved 12. These models contrast in the number of parameters - while the neural cellular automata based model has 18 062, the baseline has 446 410.
After this, I started working on a domain specific language that I hoped a language model could be trained with, but did not see it through. I also wrote some documentation for it.
In late 2022, the ARC challenge remains unsolved.