At one of his first public appearances since he joined Facebook to lead his AI initiatives, Vice President Jerome Pesenti expressed concern over the increasing usage. the computing power needed to create powerful AI systems.
"I can tell you that it keeps me awake at night," Pesenti said. "The advanced computing for companies like Facebook and Google can afford an experience we have already achieved."
It will take more software innovation so that artificial intelligence can develop without being hindered, and hardware and software optimization rather than brute force calculation could be critical for the computer. IA in the years to come.
Pluribus, an artificial intelligence system developed by Facebook AI Research and the Carnegie Mellon University, introduced today, is an example of a system less dependent on computational technologies for breakthroughs. It can be used by world-class poker players. In an article of Sciencethe researchers said Pluribus needed only $ 150 in cloud computing to train.
The end of Moore's Law means that the calculation needed to create the most advanced AI increases.
Pesenti quoted an OpenAI analysis that the calculation required to create state-of-the-art systems has been multiplied by 10 per year since 2012.
"We are still seeing gains with the increase in the number of calculations, but the pressure of the problem will only increase," Pesenti said. "I think we will continue to use more computing resources, you will still have a net profit, but it will go slower because you can not keep up with 10 times a year. It's just not possible.
The analysis presented last month revealed that the costs of training systems such as OpenAI's GPT-2 can exceed the carbon emissions of the lives of five cars.
Pesenti, who heads artificial intelligence at Facebook, on stage at the VentureBeat Transform conference, spoke of the unique challenges faced by Facebook when deploying artificial intelligence systems for $ 2.8 billion. unique users around the world. is simply modified or a deekfake.
Companies that encounter obstacles on their way when deploying artificial intelligence can be cultural or logistical, or simply do not recognize that the artificial intelligence stack is not identical to the typical technical stack.
Artificial intelligence plays a role in virtually every aspect of Facebook's services, ranging from ads to postings to recommendations on Facebook or Instagram, to moderation of content, as well as new customer experiences such as the smart camera. Portal.
Many Facebook services use Intel processors, said last year the Facebook technical manager, Kim Hazelwood.
Pesenti – like executives of Google, Microsoft and Airbnb in their discussions on Transform – also spoke about the importance of diversity in hiring and the need to ensure that AI works the same way for all .
He thinks the bias usually comes from datasets rather than AI creators.
"We are making progress. It is still very far from what we should be, "he said. "We must do everything in our power to increase diversity on the ground."
Facebook has shared new business diversity statistics earlier this week, but has not released any statistics on racial or gender diversity in divisions such as AI Research's fully-dedicated Facebook division. artificial.
An analysis conducted by a Data & Society colleague and co-author of the Algorithmic Accountability Act, Mutale Nkonde, revealed that Facebook AI Research currently employs 146 people, none of whom are of African descent.
An inclusive AI program created by Lade Obamehinti, head of the AR / VR activity of Facebook, is currently being used internally to control product biases. Obamehinti created the program a year ago after discovering that Portal's Smart Camera AI was not working with people like her with a darker pigment.
However, the measure of team diversity for AI may soon be out of date. Pesenti wants Facebook developers to be part of every team and division of the company.
"My goal is to make every engineer in the business a ML engineer, and that number has tripled in the last year. So you talk about thousands and thousands of engineers who are not part of my team and who are not actually ML engineers, he said.