杏吧原创

DeepMind鈥檚 virtual psychology lab seeks flaws in digital minds

Google鈥檚 AI company has released a simulated 3D environment in which machines can pit themselves against cognitive tests designed for humans
Coloured spheres
AIs and humans may find a given symbol in totally different ways
Getty

Are you thinking what I鈥檓 thinking? It鈥檚 a question researchers have been asking artificial intelligence from the start. Now, a team at Google鈥檚 DeepMind has developed a virtual 3D laboratory called Psychlab in which both humans and machines can take a range of simple tests and compare their cognitive abilities.

The tests were originally designed by psychologists to isolate and evaluate specific mental faculties in people, such as the ability to , for example. Now it is AI鈥檚 turn.

In initial tests, DeepMind鈥檚 own machine learning software was better at some tasks than others. The AI excelled at visual search 鈥 finding a given symbol in a group of others 鈥 but failed miserably when asked to track the position of multiple symbols on a screen. 鈥淧eople can do this effortlessly and instantly,鈥 says at Florida State University.

But one point of the project is to expose weaknesses in AIs that might otherwise go unnoticed. The team hopes that Psychlab will help developers improve their own systems. To enable this, DeepMind has released it as an open-source project so anyone can use and adapt it to their needs.

Task master?

The project could lead to more versatile AI, says at the University of Washington. 鈥淚t can broaden the set of tasks that AIs are tested on,鈥 he says.

But both Domingos and Boot agree there may be few similarities between how an AI tackles a test and the way we do. 鈥淓ven if the AI performance matches the human performance, it could be doing the task in a completely different way to a human,鈥 says Boot.

Take visual search, for example. When asked to locate a particular symbol in a group of symbols with different colours and shapes, humans take more time to complete the task with bigger groups of symbols. The DeepMind AI solved the problem in roughly the same amount of time no matter how many symbols were shown.

But DeepMind is unusual in that its co-founder, Demis Hassabis, has a research background in neuroscience. Comparing AI cognition with human cognition is still tantalising, says at the University of Oxford. 鈥淧sychlab is in this spirit.鈥

Reference:

Topics: Artificial intelligence / DeepMind / Google / Machine learning / Psychology