Are rewards enough for general artificial intelligence? We looked at this idea a few days ago; this paper is a response and criticism based mostly on conceptual difficulties with the idea. For example, as summarized in Reddit, the author argues that "the reward hypothesis is very much like behaviorism (B. F. Skinner) and past-tense learning (Rumelhart and McClelland), both of which suffered from confirmation bias and have failed. It is a circulary hypothesis." Also: "Reinforcement learning is a purely selective process... actions must already exist for them to be selected." I've seen numerous arguments of this form over the years and what they have in common is that neural networks have always eventually shown they can perform the tasks critics say were conceptually impossible.
Today: 5 Total: 97 [Direct link] [Share]
