A new study from Facebook AI Research evaluates common machine-learning classifiers' ability to label photos of objects found in households in rich countries versus household objects from poor countries and finds that the models' performance lags significantly when being asked to classify the possessions of poor people. Read the rest

In Energy and Policy Considerations for Deep Learning in NLP, three UMass Amherst computer science researchers investigate the carbon budget of training machine learning models for natural language processing, and come back with the eyepopping headline figure of 78,468lbs to do a basic training-and-refinement operation. Read the rest

"The Training Commission" is Ingrid Burrington and Brendan C Byrne's serialized science fiction tale, taking the form of an email newsletter that lets you eavesdrop on the correspondence between the story's principal characters: it's set after a civil war ("the Shitstorm"), sparked by misbehaving and easily abused machine-learning systems, and which was resolved after a protracted and catastrophic digital blackout. Read the rest

Update: I've been emailed twice by Ever PR person Doug Aley, who incorrectly claimed that Ever's signup notice informed users that their data was going to be used to train an AI that would be marketed for military applications. It's true that during the signup process, users are asked whether they want to "use" facial recognition (that is, to label their images), but not whether they consent to having their images used to train that system, and especially not for commercial military applications.

Ever is an app that promises that you can "capture your memories" with unlimited photo storage, with sample albums featuring sentimental photos of grandparents and their grandkids; but Ever's parent company has another product, Ever AI, a facial recognition system pitched at military agencies to conduct population-scale surveillance. Though Ever's users' photos were used to train Ever AI, Ever AI's sales material omits this fact -- and the only way for Ever users to discover that their photos have become AI training data is to plough through a 2,500 "privacy policy." Read the rest

An adversarial preturbation is a small, human-imperceptible change to a piece of data that flummoxes an otherwise well-behaved machine learning classifier: for example, there's a really accurate ML model that guesses which full-sized image corresponds to a small thumbnail, but if you change just one pixel in the thumbnail, the classifier stops working almost entirely. Read the rest

Software developer Chris Harris is experimenting with machine learning to remove cars from video footage; while the software isn't quite seamless, the results are pure, glorious glitch aesthetic. Read the rest

Facebook gave "as many as" 260 contractors at Wipro, Ltd in Hyderabad, India access to users' private messages and private Instagram posts so that the contractors could label them prior to their inclusion in an AI training-data set. Read the rest

Next weekend, PEN America is throwing its World Voices Festival, including a McSweeney's-sponsored panel on algorithmic bias called The Bigot in the Machine, featuring poet/media activist Malkia Cyril, and Equality Labs founder Thenmozhi Soundararajan, moderated by investigative journalist Adrianne Jeffries: it's on May 11 at 2:30 at Cooper Union's Frederick P. Rose Auditorium. Tickets are $20. Read the rest

Researchers from KU Leuven have published a paper showing how they can create a 40cm x 40cm "patch" that fools a convoluted neural network classifier that is otherwise a good tool for identifying humans into thinking that a person is not a person -- something that could be used to defeat AI-based security camera systems. They theorize that the could just print the patch on a t-shirt and get the same result. Read the rest

Pete Warden (previously) is one of my favorite commentators on machine learning and computer science; yesterday he gave a keynote at the IEEE Custom Integrated Circuits Conference, on the ways that hardware specialization could improve machine learning: his main point is that though there's a wealth of hardware specialized for creating models, we need more hardware optimized for running models. Read the rest

Ford CEO Jim Hackett -- formerly head of the company's autonomous vehicle division -- publicly announced that the company had "overestimated the arrival of autonomous vehicles" and that the vehicles, when they did arrive, their "applications will be narrow, what we call geo-fenced, because the problem is so complex." Read the rest

Ben Snell's sculpture Dio was created by training a machine learning system on a corpus of 1,000+ sculptures, tweaked in some unspecified way by Snell, who then 3D printed a mold based on the final shape: he filled the mold with a resin impregnated with the computer that ran the algorithm, which Snell had ground to powder. Read the rest

Bloomberg reporters learned that -- despite public pronouncements to the contrary -- Amazon has an "annotation team" of thousands of people all over the world, charged with reviewing recordings made by Alexa devices in the field, with both staffers and contractors listening to conversations that Alexa owners have had with and near their devices. Read the rest

ChinAI Jeff Ding's weekly newsletter reporting on the Chinese AI scene; on the occasion of the newsletter's first anniversary, Ding has posted a roundup of things about the Chinese AI scene that the rest of the world doesn't know about, or harbors incorrect beliefs about. Read the rest

This week, thousands of googlers and many others (including me) signed an open letter objecting to the inclusion of Heritage Foundation president Kay Coles James on the company's Advanced Technology External Advisory Council (ATEAC), on the the grounds that James had frequently evinced viciously transphobic, racist, anti-immigrant sentiments. Read the rest

Researchers from Tencent Keen Security Lab have published a report detailing their successful attacks on Tesla firmware, including remote control over the steering, and an adversarial example attack on the autopilot that confuses the car into driving into the oncoming traffic lane. Read the rest

People.cn is a publicly listed subsidiary of The People's Daily, the official newspaper of the Chinese Communist Party; its fortunes are rising and rising with no end in sight as it markets itself as an outsource censorship provider who combine AI and a vast army of human censors to detect and block attempts to circumvent censorship through irony, memes, and metaphors. Read the rest