Nvidia's A100 GPU coming to a cloud near you, DARPA details AI war games, Intel wants to help scan your brain

Plus: Zuck wants machines to spot bad memes on Facebook

6 Reg comments Got Tips?

Roundup The world of artificial intelligence keeps on turning: here's a summary of what's been going down.

Nvidia A100 in the cloud: Nvidia's GPU Technology Conference (GTC) was particularly eventful this year as CEO Jensen Huang unveiled Ampere, the architecture powering its next-gen hardware, while speaking from his kitchen at home.

Nvidia’s first Ampere graphics processor is the A100. It’s 20 times more powerful than the previous Tesla V100 GPU, Huang claimed. If you missed the whole affair, we covered the main announcements and the specs in more detail here.

Unfortunately, the silicon is not available for purchase as a standalone product for you and I, unless you want to buy eight of them together packaged in the DGX A100 system for $199,000.

If you can’t afford that, you can still use A100s to train your machine-learning models in the not-too-distant future, if you run the code on the cloud. Nvidia has partnered up with numerous platforms, including Alibaba Cloud, AWS, Baidu Cloud, Cisco, Dell, Google Cloud, HPE, Microsoft Azure, and Oracle, to offer A100s as-a-service.

“As AI model complexity continues to rise, the number of model parameters has gone from 26 million with ResNet-50 just a few years ago to 17 billion today,” AWS said.

“To increase performance and lower cost-to-train for models, AWS is pleased to announce our plans to offer EC2 instances based on the new NVIDIA A100 Tensor Core GPUs.” So, you can’t use A100s right away, but if you promise to be a big spender, you may be granted early access if you sign up here.

Google Cloud also gushed about collaborating with Nvidia. “We’ll be making the A100 GPUs available via Google Compute Engine, Google Kubernetes Engine, and Cloud AI Platform, allowing customers to scale up and out with control, portability, and ease of use,” it said.

Other GTC announcements include a couple of software packages, and a serious update to Nvidia's CUDA platform.

CUDA 11 is out and supports optimizations for the Ampere architecture. There are new features and libraries to accelerate matrix math operations. You can read about that in more detail here.

Nvidia also published two new frameworks: Jarvis, to build virtual cartoon chatbots, and Merlin, for scaling up recommendation systems.

Layoffs at Cruise: Self-driving startup Cruise Automation, owned by General Motors, is laying off around 160 employees amid the coronavirus pandemic.

“These are very difficult decisions to make, and we do not make them lightly,” CEO Dan Amman said in a staff memo. “These changes are the right choice for the mission.”

Cruise, based in San Francisco, was acquired by GM in 2016 to head the US automaker's self-driving efforts. Its goal to develop an autonomous robo-taxi service has been difficult to achieve.

Although Cruise had aimed to get its self-driving vehicles on the road for public use in 2019, it's had to revise its timeline. The technology simply isn't good enough yet.

Facing dwindling funds and an unstable economic climate, Cruise is getting rid of eight per cent of its staff to cut costs. It's not the only self-driving upstart to face financial trouble: Starksky Robotics completely shut up shop in March. The autonomous truck startup ran out of money and its CEO Stefan Seltz-Axmacher warned the industry was only going to get tougher over time with more startups dropping out of the race to build autonomous vehicles.

Help Facebook get rid of ‘hateful memes’: The antisocial networking giant has announced a competition to encourage machine-learning developers to solve the tricky problem of automatically identifying offensive memes.

The goal of the Hateful Memes Challenge is to develop tools that are able to spot whether a particular meme is harmful or innocent. Algorithms are able to detect hate speech or violence or nudity, but it’s a little trickier for memes since they’re “multimodal,” meaning there’s multiple types of information to analyse at once. The automated tools have to be able to understand the whole meme to work out its intent.

To help, Facebook has released a dataset containing 10,000 memes, some benign and others hateful, and has partnered with crowdsourced competition biz DrivenData to give away a chunk of cash to the best AI meme terminator.

“We are releasing this Hateful Memes data set to the broader research community and launching an associated competition, hosted by DrivenData with a $100,000 prize pool,” Facebook said this week.

It’ll be interesting if the winning tools from the competition will work on real memes. Facebook didn’t publicly reveal any examples from its dataset; you have to sign up for the competition to see them. You can also read more about the challenge in a paper here.

Intel teams up with academics to sniff out brain tumors using AI: Chipzilla and the Perelman School of Medicine at the University of Pennsylvania (Penn Medicine) in the US have been awarded $1.2m to build an AI system that analyzes medical scans of brain tumors in a secure manner.

It’s difficult to get a hold of large clinical datasets in America since medical records are sensitive and legally protected. The grant, given by the Informatics Technology for Cancer Research (ITCR) program of the government's National Institutes of Health (NIH), funds development of new methods to train neural networks on data more privately.

Intel and Penn Medicine will tackle this issue with a technique known as federated learning to train neural networks capable of detecting brain tumors. If successful, such a technique would allow hospitals to train their models without having to share patients' personal data as one big centralized system – for example, on the same cloud server.

“AI shows great promise for the early detection of brain tumors, but it will require more data than any single medical center holds to reach its full potential,” Intel Labs said this week.

“Using Intel software and hardware and support from some of Intel Labs’ brightest minds, we are working with the University of Pennsylvania and a federation of 29 collaborating medical centers to advance the identification of brain tumors while protecting sensitive patient data.”

You can read more about their work here.

AI war games: The US Department of Defense’s research arm DARPA wants to experiment with AI bots in simulated war games as part of its Gamebreaker Artificial Intelligence Exploration program.

Nine teams from academia and industry have been selected to build the best game-playing agents in different environments. Some of the most common games include battle strategy games like StarCraft II, OpenRA, others are simple shoot-em ups like 1944, and some are tactical war games like Modern Operation, or race car driving ones like TORCS.

Some notable names taking part in the program, include top defense and aerospace companies including BAE, Northrup-Grumman, and Lockheed Martin. There are a few universities such as UC Santa Barbara and Purdue University, and a a handful of machine learning startups like Blue Wave AI labs and EpiSci.

“If we can figure out a generic method to assess and then manipulate balance in commercial video games, my hope is that we might then apply those AI algorithms to create imbalance in DoD simulated war games used to train warfighters for real-world battle,” said Lt. Col. Dan “Animal” Javorsek, the Gamebreaker program manager in DARPA’s Strategic Technology Office.

The teams aren’t competing directly against with one another, since they are all experimenting with different games. You can find out more details about each team here. ®

SUBSCRIBE TO OUR WEEKLY TECH NEWSLETTER


Biting the hand that feeds IT © 1998–2020