AI learns in new ways by exploring virtual worlds

In 2009, Feifei Li , a computer scientist at Princeton University at the time, created a dataset that would change the history of artificial intelligence. The dataset, called ImageNet, contains millions of labeled images that can be used to train sophisticated machine-learning models to recognize what’s in the images. In 2015, these machines surpassed human recognition abilities. Before long, Fei-Fei Li began searching for what she called another “North Star” — one that would push AI into true intelligence in a completely different way.

She looks back on the Cambrian Explosion 530 million years ago, when many terrestrial species first appeared, and she draws inspiration from it. One influential theory holds that the explosion of new species is due in part to the emergence of eyes that can see the world around them for the first time. Feifei Li realized that animal vision never emerges on its own, but is “deeply rooted in a whole body that needs to move, navigate, survive, manipulate and change in a rapidly changing environment.” She said: “That’s why I There will be a natural shift to a more aggressive vision in AI.”

Fei-Fei Li’s work today focuses on artificial intelligence agents that can not only accept static images from datasets, but also move around and interact with the environment in a simulated environment in a 3D virtual world. That’s the broad goal of a new field known as embodied AI, and Fei-Fei Li isn’t alone in the field. The field overlaps with robotics, as robots can be the physical equivalent of an embodied AI agent in the real world, while reinforcement learning—always training interactive agents to learn to use long-term rewards as incentives. But Fei-Fei Li and others believe that embodied AI could drive a major shift from machine learning of direct capabilities, such as recognizing images, to learning how to perform complex human-like tasks, such as making an omelet, in multiple steps.

This article is reprinted from: https://www.solidot.org/story?sid=71950
This site is for inclusion only, and the copyright belongs to the original author.

Leave a Comment