Abstraction and Reasoning Challenge

My colleagues Alejandro and Yugi and I placed second in the Abstraction and Reasoning Challenge on Kaggle (Spring 2020). Hosted by Francois Chollet, the aim was to construct an AI able to solve the so-called tasks, consisting of transformations on pixel matrices with colors 0-9. These have very few training samples, making ordinary machine learning techniques not accurate enough. There were a total of over 900 teams and a total of $20000 in prize money.

On the picture to the right is an example of a task solved by our method. The task can be broken down into the following steps. First, it is necessary to identify black as background color. Once this is done, different shapes can be determined via isolation. Among the computed shapes, the largest must be then reconstructed on each of the other shapes, which in this case are subshapes of the largest one. To complicate this even more, the reconstructed shape may have rotated and/or flipped. Our algorithm generates the three candidate solutions displayed, which are then submitted to the final solution. Since the third candidate perfectly matches the output, the task is considered as solved.

Our solution and documentation are available in this public repository.

Project Euler

Currently level 14, with more than 350 solved problems. I keep a repository with links to the relevant theoretical background of each problem I solve.