Abstract
This paper focuses on a data-centric approach to machine learningapplications in the context of Conway's Game of Life. Specifically, we considerthe task of training a minimal architecture network to learn the transitionrules of Game of Life for a given number of steps ahead, which is known to bechallenging due to restrictions on the allowed number of trainable parameters.An extensive quantitative analysis showcases the benefits of utilizing astrategically designed training dataset, with its advantages persistingregardless of other parameters of the learning configuration, such as networkinitialization weights or optimization algorithm. Importantly, our findingshighlight the integral role of domain expert insights in creating effectivemachine learning applications for constrained real-world scenarios.