Alarmingly, Facebook needs more first-person shooter footage, US Energy dept buys AI-training chips, and more

Plus: Watch cute bot agents master hide-and-seek

Far Cry 4
No, not the good kind

Roundup Let's kick your week off with the latest happenings in the world of AI and machine learning.

We don't have enough first-person shooter videos

Facebook has admitted it couldn't stop the Christchurch mosque shootings because it didn't have enough "first-person footage of violent events" to train its algorithms.

A gunman livestreamed the first 17 minutes of an attack in the Al Noor Mosque in Christchurch, New Zealand, on Facebook in March. The attack left over 50 people dead and several injured. The social media giant was criticised for failing to detect and remove the video for several hours, leading to over 300,000 copies that continued to proliferate the platform.

"The video of the attack in Christchurch did not prompt our automatic detection systems because we did not have enough content depicting first-person footage of violent events to effectively train our machine learning technology," Facebook said this week. So it's turning to police training videos to boost first-person footage.

"That's why we're working with government and law enforcement officials in the US and UK to obtain camera footage from their firearms training programs – providing a valuable source of data to train our systems."

Facebook continues to use a mixture of machine learning technology and humans to flag, review content, and remove content that violates its policies.

Aw, watch some cute bots play hide-and-seek

Yes, the simple childhood game might not be exciting or hard – after all, these days AI software can now master chess, Go or poker. Folks over at OpenAI, however, reckon that the game environment, made up of four agents split evenly into hiders and seekers, breeds interesting strategies.

The hiders are given a reward of +1 score if both hiders remain hidden and a -1 score if any of the hiders are spotted by a seeker. Conversely, the seekers are given a +1 if all hiders are found, and a -1 if they aren't. Random objects like boxes and ramps are littered around the game. Over time, the hiders have learnt to drag boxes to obscure gaps and seekers have noticed they can use ramps to jump over the walls to get to the hiders. After several million games during training, the seekers learn to drag the ramp into the rooms with them and obscure any open spaces with the boxes to prevent the seekers getting to them.

Youtube Video

Though it's not the most impressive thing to come out of OpenAI, the video of the bots playing is pretty adorable.

US Energy department to buy Cerebras chips

Cerebras, the AI hardware startup known for crafting the world's largest chip, has signed a contract with the US Department of Energy to build computers for deep learning.

The multi-year collaboration involves the Argonne and Lawrence Livermore national laboratories, according to HPCWire. Rick Stevens, head of computing at Argonne National Laboratory in Washington DC, said the Cerebras chips will help researchers "invent and test more algorithms, to more rapidly explore ideas, and to more quickly identify opportunities for scientific progress."

Details about the 46,000mm2 silicon die packed with a whopping 1.2 trillion transistors were revealed at Hot Chips in August. The 18GB of on-chip memory may help companies keep all their data on the chip rather than using servers so it's ideal for sensitive customers like the US government. It is, however, a nightmare to cool and requires custom-built infrastructure to prevent the chips from getting too toasty.

Google opens research lab in India

The Bangalore lab will be led by Manish Gupta, an ACM fellow with a background in computer vision and deep learning. Milind Tambe, a Professor of Computer Science at Harvard University, will also join as the facility's Director of AI for Social Good.

You can read more about that here. ®

Sponsored: From CDO to CEO

SUBSCRIBE TO OUR WEEKLY TECH NEWSLETTER




Biting the hand that feeds IT © 1998–2019