Finishing 2nd in Kaggle’s Abstraction and Reasoning Challenge

Original article was published by Alejandro de Miquel Bleier on Artificial Intelligence on Medium


Some of these preprocessing ideas, which also included rotating matrices when necessary, or augmenting the set of samples, ended up being a key step in improving our final score.

Team Merging and End of Competition

As we were adding functions to our function basis, we started improving our score in the leaderboard. When there were only around ten days to go, we were quite confident about us ending up in the top 10. At that point, Yuji, a participant who had a decent score too, suggested us to merge teams. We considered it a good opportunity to finish in the top 3, so we decided to accept his offer, which in the end helped us solve the extra tasks that we needed to finish in the 2nd position.

Final leaderboard

Final Thoughts

Taking part in this competition was an amazing experience. I believe that there is a lot of exciting work ahead in the world of AI coming from the branch of AGI. There are just too many simple things that DL algorithms are not capable of doing, and I’m sure that future reasearch in this domain will significantly contribute to humanity’s progress. Being able to identify powerful abstractions has always been an extremely powerful tool in mathematics (that’s what mathematics is all about really), and I think that the right abstractions will have the potential to become game-changers in the field of AI.

However, this competition has shown that this might be quite difficult to achieve. No team out of the 914 participants found a satisfying, AI-focused solution for this problem. Our solution is quite deterministic (even if some of our operations included ML techniques), and so are the solutions of all the other top teams; the winner’s solution is actually quite similar to ours. Whereas this might be a bit upsetting, one needs to keep in mind that the competition lasted only for around 3 months, and if the top solutions managed to solve around 20% of the tasks in the test set, I’m sure that some long-term collaborative research could lead to the discovery of techniques that bring computers closer to understanding and mimicking our way of reasoning.