AI in the news:
Simone Scardapane
Facebook Developer Circle Roma, 28 Maggio
what's new from F8 2018 (& Google I/O)
Something about me
Post-doc fellow (Sapienza), with a strong interest in ML at large, and in the promotion of developers' events.
Focus on: Cambridge Analytica
Focus on: Cambridge Analytica
"It's easier to build an AI system to detect a nipple than what is hate speech," Zuckerberg said on a conference call with analysts following the company's first-quarter earnings report.
Using AI, Facebook [removes] 99% of terrorism-related content without a user having to notify the company first, he said. By contrast, it will likely take years for AI to reliably recognize hate speech he said. That difference is "frustrating," he acknowledged.
AI Spotlight: Hashtags
Image classification!
- The dataset contains approx. 1.5 million images annotated in 1000 classes.
- Can we scale up image classification to tens of millions of images?
- Who is going to label all these images?
Scaling up image classification
Scaling up image classification
- 3.5 billion public images from Instagram.
-
17 thousands different hashtags.
- 336 GPUs on 42 machines running for 22 days.
AI Spotlight: Frameworks
Deep learning fragmentation
Academia vs. Industry
Announcing PyTorch 1.0
AI Spotlight:
Playing GO
Deep Mind vs. Lee Sedol
Open sourcing Alpha Go
AI Spotlight
(beyond F8):
Audio Generation
