This article is more than 1 year old

Who likes role-playing? OK, OpenAI puts on its robe and wizard hat... Game bot code released, plus Apple robo-car project blow, and more

Quick catch-up on machine-learning news

Roundup Here's your rapid-fire guide to what's been happening lately in the world of machine learning.

Multiagent madness: OpenAI has released materials to train a large number of neural-network-based software agents using reinforcement learning in multiplayer game environments.

The tech platform, dubbed Neural MMO, creates tile-based environments, in which bots move around trying to collect food and water to stay alive. There are a limited number of resources at any given time, and teams compete by attacking one another. Destroying opponents means there is less competition for food and water, so there's a higher chance of survival. The agents are rewarded for staying alive in the game for as long as possible.

Neural MMO supports up to 100 million lifetimes of 128 concurrent bots in 100 concurrent game servers. Developers can use it to craft their reinforcement learning algorithms for use in bots for massively multiplayer online (MMO) games.

"In recent years, multiagent settings have become an effective platform for deep reinforcement learning research," OpenAI explained this week. "Despite this progress, there are still two main challenges for multiagent reinforcement learning. We need to create open-ended tasks with a high complexity ceiling: current environments are either complex but too narrow or open-ended but too simple.

"Properties such as persistence and large population scale are key, but we also need more benchmark environments to quantify learning progress in the presence of large population scales and persistence. The game genre of Massively Multiplayer Online Games (MMOs) simulates a large ecosystem of a variable number of players competing in persistent and extensive environments."

See the above links for code, info, and an academic paper going into more detail.

Apple slashes robo-ride project: Apple confirmed it is laying off 190 employees from its AI-powered self-driving car team.

Rumors that its autonomous car division – a highly secretive internal effort dubbed “Project Titan” – was shrinking began swirling in January. Since then, Silicon Valley-based Apple has filed a letter to the California Employment Development Department, detailing that 190 unlucky bods will be jobless by mid-April.

Engineers are the hardest hit, according to the San Francisco Chronicle: at least 33 hardware engineers, 31 product design engineers, 22 software engineers, and 38 engineering program managers will be laid off.

It’s not looking too good for Apple in this arena. A recent Department of Motor Vehicles report in California showed that the company's robo-rides had one of the worst “disengagements per mile” rates in the US state. That metric is a fancy way of describing the average distance an autonomous vehicle can travel before its human driver had to disengage the AI and take over the controls.

Apple’s cars can only travel 1.1 miles before human intervention is needed. That number is staggeringly low compared to Waymo’s 11,154 miles and Cruise’s 5,204 miles. The only company even worse than Apple appears to be Uber at 0.4 miles.

You can get a lowdown of the report here or look at all the submissions here.

Australia invests a few million in building ethical AI: The Australian government’s Defence Department is forking out a modest AU$9m ($6.4m, £4.8m) to fight off “killer robots.”

The money will be given to researchers at the University of New South Wales (UNSW) and the University of Queensland, who are tasked with inventing ways to prevent machines taking over. Jai Galliott, a researcher at the Australian Defence Force Academy at UNSW, said, according to The Sydney Morning Herald: "A lot of people speak about ethics and law of autonomous weapons, killer robots and whatever else, but not a lot of people are actually trying to provide the solutions."

Dr Galliott envisions nightmarish scenarios such as autonomous weapons destroying Red Cross ambulances and dangerous factory robots working alongside humans, all of which needs tackling. “Asimov’s laws can be formulated in a way that basically represents the laws of armed conflict, actually," Dr Galliott added.

Autopilot may have been responsible for Model 3 crash: A man driving a Tesla Model 3 was killed this month when his car smashed into a semi truck in Florida, USA, on Friday. Jeremy Banner, 50, of Lake Worth died at the scene, and the flash motor's Autopilot software – a form of super-cruise-control – may be to blame, if it was engaged.

The deadly accident was described as a “side underride”, or in other words, the roof of the car was sheared off as it passed underneath the truck. A statement from the Palm Beach County Sheriff’s office reads:

Vehicle 1 (V-1) was a tractor/trailer combination vehicle traveling eastbound on the driveway access to 14095 SR 7 (Pero Farms) preparing to turn left onto SR 7. Vehicle 2 (V-2) was traveling southbound on SR 7 within the outside lane approaching Pero Farms. After V-1 came to a brief stop at a stop sign, V-1 entered the southbound lanes of SR 7 pulling into the path of V-2. V-2 struck the driver side of V-1’s trailer resulting in the roof being sheared off as it passed underneath the trailer. V-2 continued southbound and came to a final rest approx 3/10th of a mile south of the collision. The driver of V-2 was pronounced deceased on scene.

It is unknown right now if Tesla’s Autopilot function was enabled during the crash, though two federal watchdog – the National Transportation Safety Board, and the National Highway Traffic Safety Administration – are investigating. If the software was active, it would not be a dissimilar situation to the one in 2016 when a driver died after he and his Tesla Model S's Autopilot system failed to notice a truck and smashed right into it.

We have reached out to Tesla for comment.

Massive Mozilla voice data dump: If you’re looking to train a neural network on speech recognition, there are some new voice datasets that were released by Mozilla last week.

The datasets are collected under Mozilla’s “Common Voice” project. In it, you can find samples of people’s voices in different languages. It is, apparently, the largest available public-domain voice dataset with 1,087 hours of speech data in 18 languages, ranging from English to Esperanto to Hakha Chin, a language predominantly spoken in Myanmar.

“The Common Voice dataset is unique not only in its size and licence model but also in its diversity, representing a global community of voice contributors,” Moz's George Roter explained. "Contributors can opt-in to provide metadata like their age, sex, and accent so that their voice clips are tagged with information useful in training speech engines."

Mozilla has been updating its Common Voice dataset since the project began in June 2018. You can download it here.

Microsoft wants to build a medical chatbot for clinical trials: Microsoft is hoping to partner with pharmaceutical giants to find patients looking for clinical trials to help flesh out its medical chatbot.

Coders working at a Microsoft hackathon project in Israel crafted a bot capable of matching patients to clinical trials. Now, Microsoft wants to expand the clinical trials bot project into something bigger, apparently.

But Microsoft's not trying to come up with a new product to sell. Instead, its turning to pharma corps for more data. It wants to test the bot to find new trial patients so that it can eventually pitch the idea to other companies that can use the tool to create a fully-fledged product for drug companies to use.

Users type in queries such as: “trials for a 52-year old California female with breast cancer.” And, the bot will throw up questions querying the current health status of the patient, such as if he or she has gone through chemotherapy before and how far they’re willing to travel before spitting out a possible trial, according to Bloomberg. ®

More about


Send us news

Other stories you might like