Algorithm creates a ‘new Nirvana track’, ‘live’ photos learn to dance, and other AI news

We aim to inform readers not only about developments in the bitcoin industry but also about what is happening in adjacent technological domains—cybersecurity, and now the world of artificial intelligence (AI).

ForkLog has gathered the most important AI news of the week.

US government employees use Clearview AI without their supervisors’ knowledge

Employees at almost two thousand U.S. government agencies, including police departments and public schools, used Clearview AI without authorization from their superiors.

According to BuzzFeed News, they used facial recognition to search for not only participants in recent protests across the country but also their own friends and relatives.

Reporters contacted the heads of several agencies. It emerged that many did not know their staff had access to the system.

This came to light thanks to an anonymous source who provided journalists with a database of search queries indicating which agencies had access to the platform.

Based on this information, journalists created a page, where anyone can check how and when facial recognition from the controversial platform was used.

Clearview AI declined to confirm the information’s reliability and did not respond to questions about the leak.

EU urged to ban facial recognition technology

A coalition of 56 digital-rights groups urged the European Commission to ban facial recognition technologies for mass surveillance without any exemptions.

An open letter on behalf of the European Digital Rights (EDRi) was sent to Justice Commissioner Didier Reynders—just weeks before the Commission planned to publish new rules on ethical AI use on the continent.

Rights advocates urged the Commissioner to back stronger protections for fundamental rights in forthcoming laws.

The group says there are no compelling examples where facial recognition in public spaces could justify the harm to rights such as privacy, data protection, or non-discrimination.

EU’s Vice-President for Digital Technologies Margrethe Vestager previously stated that the use of this technology runs counter to GDPR requirements on consent to data collection. But rights advocates say this is not enough, as the regulation contains many exemptions.

In the open letter, EDRi also urged the Commission to carefully study other potentially dangerous AI applications, particularly those that could impede access to healthcare, social welfare or justice.

Facebook releases open dataset in which annotators self-reported their data

Facebook’s AI team released as open data the Casual Conversations dataset, comprising 45,186 videos with 3,011 participants.

The company said that people who contributed to the dataset and self-identified their gender, age and skin tone received financial compensation.

A team of professional annotators also worked on the dataset. They appended lighting metadata to the videos to improve face recognition accuracy in various lighting conditions.

Facebook says this is the first of its kind visual data set to be publicly available.

Google AI chief Samy Bengio resigns

Google AI chief Samy Bengio resigned after 14 years in the department. His departure is linked to the controversial dismissals of Timnit Gebru and Margaret Mitchell—two leading AI-ethics researchers.

Gebru was fired in December 2020 while she was working on a paper criticizing large language models and warning of their dangers. In February the company sidelined Mitchell, who had argued that her colleague’s firing was unlawful.

Samy Bengio spoke up for the researchers. In a post on Facebook he wrote that he was “shocked” by what happened to Gebru.

In a farewell letter, Bengio told colleagues that it was hard for him to leave the division. He did not disclose his future plans or the reasons for his departure.

US government contractor creates name-by-face prediction algorithm

The American nonprofit Mitre Corp. developed technology for predicting a person’s name from facial features.

It builds on work by researchers at the Hebrew University of Jerusalem, who trained the algorithm to determine a person’s name from two given options. The model they developed correctly guessed the name in 59% of cases, above the 50% expected from random guesses.

However, Mitre employees refined the technology and raised the accuracy from two options to 80.5%.

They suggest their algorithm could also be used in reverse: selecting the most suitable photos of people for a given name.

Journalists also found that the sole client of Mitre is the U.S. government. They previously developed facial-recognition technologies for the FBI. Over the past few years the company has received up to $2 billion in federal funding for research.

US startup sold a non-existent AI technology to Utah’s government for $20.7 million

The U.S. startup Banjo sold an AI platform to the state of Utah for $20.7 million, later admitting that it had never developed AI technologies.

Local authorities launched an internal investigation and suspended the contract. An audit confirmed there were no AI capabilities whatsoever, and deemed the platform unsafe for privacy.

AI technology was a key selling point for the startup, influencing the state’s decision.

Following the win in the contract, the company was granted access to surveillance cameras, 911 video footage, and traffic data.

The controversial company continues to operate under a new name. Journalists found that to avoid heightened scrutiny, the startup rebranded as safeXai.

Waymo chief resigns

Waymo’s chief executive, John Krafcik, stepped down after 14 years with the unit. His departure is linked to the controversial layoffs of Timnit Gebru and Margaret Mitchell—two leading AI-ethics researchers.

Gebru was fired in December 2020 while she was working on a paper criticizing large language models and warning of their dangers. In February the company sidelined Mitchell, who had argued that her colleague’s firing was unlawful.

Samy Bengio spoke up for the researchers. In a post on Facebook he wrote that he was “shocked” by what happened to Gebru.

In a farewell note, Bengio told colleagues that it was hard to leave the division. He did not disclose his future plans or the reasons for his departure.

AI wrote a ‘new Nirvana track’

The mental health charity Over the Bridge used Google’s Magenta tool and a convolutional neural network to create a melody and lyrics for a Nirvana-style song.

First, Magenta studied certain notes and harmonies from MIDI files of 20 Nirvana songs to study its repertoire.

Then researchers selected the best AI-generated samples and layered various instrumental effects on them. When the track was ready, it was performed by Nirvana’s cover-band frontman, Eric Hogan.

Earlier the organization had created tracks using AI and other musicians who died at 27. This, in turn, aims to draw attention to mental-health resources people can turn to if they need help.

An enthusiast creates AI for transforming images into the style of famous artists

Digital artist Glenn Marshall developed the Chimera text-to-image transformer, which stylizes any image in the style of famous painters.

The artist built a system based on the Aleph-Image neural network, which relies on CLIP and DALL-E models developed by OpenAI.

The algorithm transforms the input image toward a given textual prompt, such as ‘the art of Salvador Dalí’. In a process of repeated manipulations and frame-by-frame iterations, the AI gradually identifies elements and shapes matching the description until a final composition is created.

Marshall said he had not seen anything like this in 30 years of observing computer art.

“Using any image and any text, the combination of boundless possibilities is simply mind-blowing. No one else is doing anything like this,” the artist added.

He also noted that the technique he created is closer to distortion than to style transfer.

ForkLog also features:

• Researchers developed a game to illustrate the risks of emotion recognition by AI.
• Deep Nostalgia taught ‘live’ photos to dance and blow a kiss.
• Researchers accelerated neural network training on CPUs by 15x.
• The anti-cheating program integrated a ‘racist’ facial recognition algorithm.

Subscribe to ForkLog’s AI news on Telegram: ForkLog AI — all the AI news!