The tremendous advancements in artificial intelligence remain focused in narrow applications. As described in a recent article authored by Ben Dickson, these narrow systems have been designed to perform specific tasks instead of having general problem-solving abilities. The quest for general problem-solving ability has long been pursued, with many focused-on replicating aspects of human intelligence like vision, language, reasoning, and motor skills. Now, a new paper submitted to the peer-reviewed Artificial Intelligence journal describes an argument put forward by scientists at U.K.-based AI lab DeepMind. They argue that intelligence and its associated abilities likely emerge by rewarding maximization versus formulating and solving complicated problems.
The authors suggest that reward maximization and trial-and-error experience are enough to develop behavior that exhibits the kind of abilities associated with intelligence. And from this, they conclude that reinforcement learning, a branch of AI that is based on reward maximization, can lead to the development of artificial general intelligence.Reward is Enough – DeepMind
The article goes on to describe two potential paths for AI. One path takes multiple narrow artificial intelligence systems and assembles them in a way that solves complicated problems requiring a multitude of skills. The other path takes a page out of nature. Intelligence evolved over time through interaction with the environment. Researchers at DeepMind argue that the most general and scalable way to maximize reward is through agents that learn through similar interactions with their environment. The paper (and the article) explores examples from nature to support their argument. It then makes the connection between that argument and reinforcement learning. Driven by an environment, agents, and rewards, reinforcement learning can replicate reward maximization as seen in nature and can eventually lead to artificial general intelligence.
The paper identified by the author provides several examples of how reinforcement learning acquired general skills. However, like every other emerging scenario, there are obstacles. The author points to the amount of data required to master a given domain, and they still have not figured out how to generalize their learnings across several domains (the whole point of artificial general intelligence). The huge energy demands associated with the process is another potential gating factor, as is its impact on our climate. To the authors credit, they identified many of the obstacles yet to be overcome. The discussion regarding when we will realize artificial general intelligence is still open to speculation – as are the societal challenges it may represent.