Researchers at Britain’s DeepMindstated that for achieving general artificial intelligence reinforcement learning is sufficient. Maximising rewards will help develop the machine’s intellectual capacities, they argue.
“[We] consider an alternative hypothesis: a general goal of maximising rewards is sufficient to govern the abilities studied by natural and artificial intelligence,” the researchers said.
According to the researchers, to develop general AI it is necessary to recreate the process of evolution rather than merge a multitude of narrow applications with algorithms into a single system.
They argue that this is an effective approach that has led to the development of living beings with the skills and abilities to perceive, orient themselves, modify their environment, and communicate with one another.
“The natural world, where animals and humans interact, as well as the environment for artificial agents by their very nature are so complex that success in them requires sophisticated abilities,” the study says.
The researchers also believe that the abilities related to intelligence that arose from the goal of maximise rewards will provide a deeper understanding of them.
“Conversely, when each ability is created to solve a specialised goal, the question ‘why’ is pushed aside,” they added.
The DeepMind artificial intelligence laboratory was founded in 2010 and acquired by Google in 2014. The organisation conducts research in natural language processing, computer vision and neural-network training algorithms.
One of DeepMind’s best-known developments is the Go-playing program AlphaGo. In March 2016 it defeated the world’s strongest Go player Lee Sedol by 4–1.
In May 2021, Herman Gref, chairman of Sberbank’s board, said that artificial intelligence will force people to change professions.
In May, Nobel laureate Daniel Kahneman expressed the view that artificial intelligence will surpass humans, but not soon.
In March, OpenAI president Sam Altman said that in the coming decade every American could earn $13,500 a year thanks to artificial intelligence.
Subscribe to ForkLog AI on Telegram:ForkLog AI — all the news from the world of AI!