Weeks after cherry-picking reinforcement learning, Meta AI chief Yann LeCun went on Twitter again to express his views and remind everyone how he had predicted the reach and impact of reinforcement learning. LeCun has been a big proponent of self-supervised learning and also said that innovations using SSL have been working better than he anticipated.
The discussion started when Dan Becker, former Google data scientist and founder of Decision AI, tweeted saying that after AlphaGo beat Lee Sedol, a Go professional from South Korea, many researchers started to believe that reinforcement learning will be a game changer. But the impact now seems smaller because of its limitations in real world problems.
During an exclusive interaction with Analytics India Magazine, LeCun said that though reinforcement learning is very efficient, it requires multiple trials. The machine has to learn from the responses that it receives from the world, and this proves to be highly inefficient and unreliable.
Read: Angels & Demons of AI
LeCun adds to his point saying that though RL is inevitable in machine learning, the purpose behind incorporating it in algorithms should be to eventually minimise its use. Agreeing to this, Sebastian Raschka, an AI researcher, agreed with LeCun and said that for large language models to work, self-supervised learning is a must.
Further in the discussion, LeCun said that even ChatGPT uses SSL more than RL but there are only two obstacles – defining explicit objectives and planning abilities.
This is not the first time that LeCun has criticised reinforcement learning. Recently, presenting for SSL at NeurIPS 2022, he suggested that researchers abandon reinforcement learning to further the development in AI.
