Yann LeCun proposes to endow AI systems with ‘common sense’

“Autonomous” AI capable of exploring and perceiving the surrounding world as humans do requires more effective training. This conclusion was reached by Meta vice president and chief AI scientist Yann LeCun, according to Motherboard.

Compared with humans and animals, modern AI systems lack common sense. This concept is essential for developing ‘autonomous’ AI algorithms capable of learning on the fly by observing the real world, rather than through prolonged training aimed at a single specific task.

LeCun published a исследование in which he proposed a way to address the problem. He argues that algorithms require more effective training, because AI \”isn’t very good\” at planning and forecasting changes in the real world. By contrast, humans and animals can acquire a vast amount of knowledge through observation and mundane physical interactions.

Teenagers can learn to drive after tens of hours of practice without getting into an accident. For the same task, AI systems require enormous amounts of data.

“A self-driving car would need to drive off a cliff several times before it understands that it\’s a bad idea, and thousands more attempts to learn not to fall off,” LeCun noted during the доклада at the University of California, Berkeley.

According to him, the difference lies in the presence of ‘common sense’ among humans and animals.

In the paper, the scientist describes this concept as a set of models that help a living being draw a distinction between what is probable, feasible, and impossible. Such a capability enables people to explore their surroundings, fill in missing information, and offer new solutions to unknown problems.

LeCun said that many modern training processes, such as reinforcement learning, are not up to the mark when it comes to matching human reliability in real-world tasks.

“This is a practical problem, because people genuinely need machines with common sense. We need driverless cars, home robots, and intelligent virtual assistants,” he said.

The specialist proposed an architecture designed to minimise the actions an AI must perform to learn and accomplish the designated task.

LeCun outlined a model for building an ‘autonomous’ intelligence, consisting of five separate but configurable modules. One of the most challenging components of the architecture would be a simulator-like module that assesses the state of the world and forecasts imagined actions and other sequences. This would allow knowledge about the environment to be applied to a range of tasks.

In July, Google fired an engineer after he stated о «зарождении разума» у ИИ-алгоритма LaMDA.

In January, Elon Musk пообещал предотвратить бунт машин против людей.

In June 2021, DeepMind scientists заявили, что для достижения общего искусственного интеллекта достаточно обучения с подкреплением.

Follow ForkLog’s news on Telegram: ForkLog AI — all the news from the world of AI!