AI Conference

Focus on AI at Facebook F8; PyTorch 1.0 Released (Updated)

Facebook generally uses its F8 Developer Conference to introduce new website features. This year however there was a distinct emphasis on “AI,” with the term mentioned more than ever in the event's keynote speech.

Facebook generally uses its F8 Developer Conference to introduce new website features. This year however there was a distinct emphasis on “AI,” with the term mentioned more than ever in the event’s keynote speech.

Facebook Founder and CEO Mark Zuckerberg unveiled the company’s strategic AI roadmap, pushing R&D in vision technologies, unsupervised learning, natural language processing, reinforcement learning, generative networks, and AI developer tools. Facebook also announced several AI-related updates to elevate the user experience.

image (10).png

Detecting fake accounts and fake news

Russian bots, fake news and the Cambridge Analytical scandal have hit Facebook hard. The seminal social media platform — where billions of users share photos and videos of their lives with family and friends, and exchange thoughts and opinions on events of the day — is increasingly revealing a dark side, as a malignant tool for spreading false information, instigating hatred, and even influencing elections.

Facebook wants to fight back. Zuckerberg told the F8 audience the company is now using AI algorithms to detect “spammers who just want money,” “fake accounts created by bad actors,” and “real people who are sharing fake information.”

Facebook has previously suggested they were deploying cutting-edge AI to recognize fake news and bot-generated accounts. In 2016, AI pioneer and then Facebook Director of AI Research Yann LeCun said AI could be used to identify fake news or violence in live video content on the site. In 2017, Facebook announced that they would use AI to detect terrorism-related posts.

AI researchers are divided on whether machine learning algorithms can effectively detect fake news. Dean Pomerleau is a research scientist and entrepreneur who helped organize a crowdsourcing challenge to develop machine learning solutions to combat fake news. He told The Verge that in his opinion AI couldn’t fix the fake news problem. On the other hand, Aaron Edell, CEO of AI company Machine Box, claims to have produced a machine learning algorithm that can detect fake news with higher than 95% accuracy.

Zuckerberg faced questions about Facebook’s data privacy practices at a congressional hearing last month, where he estimated the company would need five to ten years to build an effective AI-powered fake news detection system that leaves humans out of the loop.

Improving Instagram content recommendations

The Instagram Explore page is a powerful feature that recommends posts based on personal interests and tastes. Users can browse content across topics of interest, even from accounts they don’t follow.

Facebook Data Science & Analytic Manager Tamar Shapiro announced at F8 that Facebook will revamp Instagram Explore to better organize its recommended content into different topic channels, and this new Explore page will be powered by AI.

“In order to deliver cutting-edge experience, we are augmenting AI with content classification and curation signals from our community,” said Shapiro.

Shapiro also introduced a new AI-powered “bullying comment filter,” which can hide content that disturbs or upsets users. Last year, Instagram launched an offensive comment filter, which can automatically hide comments it deems “divisive” or “toxic.”

image (11).png

Messenger bot translating across different languages

Facebook VP of Messaging Products David Marcus announced M Translation, a Facebook Messenger AI-powered machine translation service for Facebook’s buy-and-sell service Marketplace.

M Translation expands on M Suggestion, a pop-up feature launched last year that suggests relevant content and capabilities. Now, when users receive a Marketplace message in a language different from their default, M Translation can translate the message into their default language.

image (12).png

Facebook smart speakers to sell only outside the US?

Rivals Google, Amazon, and Microsoft have already jumped on the smart speaker bandwagon, as have major Chinese tech companies. Now Facebook is rumoured to be readying their own line of smart speakers, although Zuckerberg was tight-lipped about the plan at F8.

Multiple media reports suggest Facebook will launch two devices this July in overseas markets. The unusual marketing plan is said to be due to Facebook’s slipping trustworthiness among users in the US. The smart speaker is expected to be equipped with a touchscreen and camera, and will be powered by the text-based chatbot Facebook Messenger bot, M, which will likely get an upgrade to voice assistant.

UPDATE: PyTorch 1.0 released

image (14).png

On F8’s second day, Facebook announced PyTorch 1.0, the latest version of its open-source AI software framework that guides and supports researchers from research stages to deployment of trained models for various AI applications.

VP of Facebook Infrastructure Bill Jia said “PyTorch 1.0 takes the modular, production-oriented capabilities from Caffe2 and ONNX and combines them with PyTorch’s existing flexible, research-focused design.”

Facebook is pushing the combined PyTorch — Caffe2 framework. Last month, the Caffe2 Github page introductory “readme” document was suddenly replaced with a link: “Source code now lives in the PyTorch repository,” which enabled Caffe2 users to directly check Caffe2 code in PyTorch.

Since its release in October 2016, PyTorch has become a preferred machine learning framework for AI researchers due to its flexibility. Over half of Facebook AI projects run on PyTorch. PyTorch 1.0 will be available to beta users later this summer.

Facebook also announced the open-sourcing of many AI tools, including Translate, a PyTorch Language Library for neural machine translation, and ELF OpenGo, an AI bot based on the ELF (extensive, lightweight and flexible) platform for training gameplay AI with reinforcement learning.

Facebook Chief AI Scientist Yann LeCun tweeted that the ELF OpenGo bot “has attained professional level in two weeks of training and has won 15 games against 4 top professional human players.”


Journalist: Tony Peng| Editor: Michael Sarazen

0 comments on “Focus on AI at Facebook F8; PyTorch 1.0 Released (Updated)

Leave a Reply

Your email address will not be published.

%d bloggers like this: