3.3.3 Learn from the algorithm itself about it’s decisions

Interpretability can help humans to learn from machines too. In 2016, AlphaGo played a mysterious move against Lee Sedol, the world’s top player for AplhaGo, which completely stunned everyone. The move didn’t make sense to any human present in the room. 1

That’s a very strange move! I thought maybe it was a mistake. - Lee Sedol

Three times European Go Champion, Fan Hui was gobsmacked by the move too.

It’s not a human move, I have never seen a human play the move like this! So beautiful! - Fan Hui

Google DeepMind AlphaGo vs Lee Sedol
credits: bbci

This historic incident concludes that human can learn from machines too provided we know how the system works and and reason for its success. In areas like Reinforcement Learning, interpretability can help a lot in understanding the decisions made by algorithms.

RL based Hide-and-Seek Game by OpenAI credits: linkedin

Recently, OpenAI released a clever hide and seek game based on 2 teams of baby RL agents. One of the team needs to find the next team. The arena has random objects placed. Through Reinforcement Learning, the agents learned to exploit the objects by moving them and made use of physics during their time in the game. Most of the moves we not taught to the agent and through self understanding and learning, it learnt to exploit the bugs in the game and break the rules. 2

Citations

1

https://www.wired.com/2016/03/googles-ai-viewed-move-no-human-understand/

2

https://techcrunch.com/2019/09/17/clever-hide-and-seek-ais-learn-to-use-tools-and-break-the-rules/