meet artificial intelligence that is ‘jack of all trades’

The most promising models of artificial intelligence (AI) of the moment, like the ones that appear in this special, have in common the fact that they are very good at specific tasks. There are AIs that specialize in text, images or robot control. Even though they are highly sophisticated in their fields, they are almost useless for other applications. it is not the case of Catmodel released in May this year by DeepMind.

The system is a jack of all trades: it can play video games, create photo captions, control chatbots and command a robotic arm. In total, it is capable of performing 604 different tasks. The multiple functions originate from the diversity of information that fed the system (the technique used to train the model is the same as for systems focused on generating text). The Gato was prepared with several open databases, from plays to atari even image packs.

“The Gato is a foundation model, which intends to absorb knowledge from several areas that go beyond language. The idea is that, when adopting it for a given application, it brings information from other fields”, explains Fábio Cozman, director of the Center for Artificial Intelligence (C4AI) at the University of São Paulo (USP).

The images below were generated by AI, using the DALL-E 2, Midjourney and Stable Diffusion models)

Images created by AIs (from left to right) DALL-E 2, MidJourney and Stable Diffusion with command "A colorful three-dimensional rendering of a pink cat, writing code on a computer, at her violet office desk"
Images created by AIs (left to right) DALL-E 2, MidJourney and Stable Diffusion with command “A three-dimensional color rendering of a pink cat, writing code on a computer, at its violet office desk” Photograph: DALL-E 2 / MidJourney / Stable Diffusion / Estadão

AIs of this type gain importance as they move towards solving more complex problems. For example, a traffic security system that identifies license plates and scans vehicle databases works better in the Gato scheme than a system that adopts two different algorithms (one for image detection and the other for search). of words). “Communication between different AIs is complex and increases the chance of errors in the systems”, says Anderson Soares, coordinator of the Center of Excellence in Artificial Intelligence at the Federal University of Goiás (UFG).

Generalist models are nothing new among researchers – it is from this idea that AIs are represented in fiction most of the time. In the real world, however, the Cat is the first to show reasonable results. “She is the first to point to a new class of AI systems,” explains Soares.

Continues after advertising

The Cat is the first AI to point to a new class of systems

Anderson Soares, coordinator of the Center of Excellence in Artificial Intelligence at the Federal University of Goiás (UFG)

The excitement was captured in the words of Nando de Freitas, one of the researchers behind Gato. “Game over!” he wrote on Twitter in May, in reference to the growth of generalist systems. There is, however, a long way to go.

In the same message, the researcher admits that it is necessary to make the models bigger, safer and more efficient. Soares also says that they need to prove themselves efficient in commercial applications. For this, the Cat needs to improve its results.

The model multitasks but is not exceptional at any of them. Of the 604 tasks capable of fulfilling, the system achieved a level of excellence more than 50% of the time in just 450. In the world of AI, it is not enough to be a student with a grade of 6 in all subjects. You have to have much higher rates to make a lasting impact.

Source link

About Admin

Check Also

Man steals his own cell phone from a store and charges R $ 3,000 for ‘loss’

A man was arrested in flagrante delicto for the crime of embezzlement after trying to …

Leave a Reply

Your email address will not be published. Required fields are marked *