1: \begin{abstract}

2:       In this paper, we propose a new approach for building cellular automata to solve real-world segmentation problems.     We design and train a cellular automaton that can successfully segment high-resolution images.    We consider a colony that densely inhabits the pixel grid, and all cells are governed by a randomized update that uses the current state, the color, and the state of the $3\times 3$ neighborhood. The space of possible rules is defined by a small neural network.    The update rule is applied repeatedly in parallel to a large random subset of cells and after convergence is used to produce segmentation masks that are then back-propagated to learn the optimal update rules using standard gradient descent methods. 

3:     

4:     We demonstrate that such models can be learned efficiently with only limited trajectory length    and that they show remarkable ability to organize the information to produce a globally consistent segmentation result, using only local information exchange. 

5:     

6:     From a practical perspective, our approach allows us to build  efficient models -- our smallest automaton uses less than 10,000 parameters to solve complex segmentation tasks.

7:    

8:     \keywords{Image segmentation, cellular automata, deep learning}

9: \end{abstract}

10: