Regulating AI | The Juice

Zumo Labs presents The Juice, a weekly newsletter focused on computer vision problems (and sometimes just regular problems). Get it while it’s fresh.

Week of April 19–23, 2021

____

The halcyon days of the algorithm as a convenient scapegoat are quickly drawing to a close. This week the FTC issued a warning to companies building AI systems, explaining that there are actually several ways those companies may already be in violation of federal law. Whether a company is overpromising their technology’s capabilities or they’ve unintentionally deployed a biased algorithm, the FTC makes it clear (in an article best described as lightly threatening) that they would prefer they didn’t.

Meanwhile, the EU published a new framework proposal intended to regulate the use of AI across the union’s nations. It is both exhaustive and riddled with exceptions. A ban on the use of real-time biometric ID tech in public places by police, for example, makes an exception for locating suspects for crimes that carry a 3 year minimum sentence. Even specifying “real-time” leaves open the option of police working directly with contentious firms such as Clearview AI. We’re planning on doing a deep dive on it, but in the meantime Dr. Gabriela Zanfir-Fortuna has a lengthy Twitter thread breaking it down.

As always, Like, Subscribe, and share The Juice with your smartest friends.

____

#EarthDay

Lawns (for many a millennial an unachievable dream) are environmentally problematic. Beyond the vast amounts of water they consume and native plants they displace, emissions caused by lawnmowers — not required to have catalytic converters like cars — contribute significantly to global warming. But new robotic lawnmowers, which are basically weatherproofed Roombas with computer vision systems, might help make a dent in that problem.

Robotic Lawnmowers Could Cut a Huge Swath in Air Pollution, via Inverse

#FacialRecognition

Often when we worry about facial recognition technology, we’re thinking about the government or other bad actors compiling a database of faces and names they might use against us in some way. Meanwhile, we willingly provide those exact data to social media platforms. That came back to bite one suspected Capitol rioter, whose face was ID’d through his girlfriend’s Instagram posts. In other news this week, a bipartisan bill has been introduced that would essentially ban law enforcement and intelligence agencies from buying data from Clearview AI.

FBI used facial recognition to identify a Capitol rioter from his girlfriend’s Instagram posts, via The Verge.

Senators want to block government agencies from buying Clearview AI data, via Engadget.

#Amazon

Where do you draw the line on sharing your biometric data? Amazon is betting that consumers will be willing to share their palm prints for the convenience of a faster checkout at Whole Foods. Meanwhile, Amazon India also announced this week that they’re planning to roll out computer vision tech that will grade the quality of their produce. They estimate that the switch away from a manual grading system to the AI alternative will reduce costs by 78%.

Amazon brings its palm-scanning payment tech to Whole Foods, via The Washington Post.

Amazon building computer vision tech to check defects in fruits, vegetables, via Business Standard.

#BadData

“Alongside other minority groups, people with disabilities have long been harmed by flawed data and data tools. Disabilities are diverse, nuanced and dynamic; they don’t fit within the formulaic structure of AI, which is programmed to find patterns and form groups. Because AI treats any outlier data as “noise” and disregards it, too often people with disabilities are excluded from its conclusions.”

Flawed data is putting people with disabilities at risk, via TechCrunch.

#EthicalAI

We’ve been highlighting news about the breakdown of Google’s Ethical AI team since December. But this report from Bloomberg suggests that, while that’s when things became visible, there were actually systemic issues inside the organization going back years. It’s a great piece, and may be the most comprehensive overview of the whole debacle if you’re trying to get up to speed on it.

Google Turmoil Exposes Cracks Long in Making for Top AI Watchdog, via Bloomberg.

____

📄 Paper of the Week

GANcraft: Unsupervised 3D Neural Rendering of Minecraft Worlds

Minecraft worlds are composed of thousands of cubes, which are conveniently labeled according to their class (grass, water, rock, etc.). Combined with the procedural generation of Minecraft worlds, they really are the perfect test bed for a neural volume rendering paper. Hats off to this research team for picking a great dataset and a great problem. Neural volume rendering is possible because a 3D world can be represented as a continuous function which is differentiable (differentiable means you can perform gradient descent). This rendering approach is becoming increasingly popular, and it is becoming increasingly apparent that the way we render things now (running shader programs on meshes made of triangles) will eventually be replaced by some form of volumetric rendering. This could have huge implications in the world of 3D: those who don’t keep up will get left behind.

____

Think The Juice was worth the squeeze? Sign up here to receive it weekly.