This article is more than 1 year old

Facial recog biz denies its software identified 'antifa members' among mob that stormed Capitol Hill

Plus: US ban on selling AI code to China renewed for 2021, and Jim Keller hired by Canadian AI chip startup

In brief The facial recognition company said to have identified antifa members among rioters who ransacked Capitol Hill last week denied that its technology had ever done such a thing.

The rumour started in an article (since deleted) in Washington Times and was later spread by Matt Gaetz, a representative from Florida's fourth district and Trump supporter, after armed insurgents stormed the legislative halls.

In what appears to be a bizarre attempt to shift the blame from extreme Trump followers, the conservative-leaning news publisher said XRVision had spotted antifa members in footage filmed from the disaster. But the company told Buzzfeed the Times article was completely false. Instead, it said its software had actually spotted neo-Nazi and QAnon supporters among the crowd.

New York City is thinking of auditing AI algorithms used in recruitment

Officials working at the The Big Apple's city council have put forward a bill that would require employers to notify candidates when their applications have been analyzed by software, and to subject those algorithms to yearly audits.

The audits are designed to make sure the algorithms are checked for any potential biases. Companies will not be allowed to sell their software onto employers without having been audited first. If they violate any of the proposed rules, they will have to pay a penalty to the council.

The bill was introduced last year and first reported by Wired's Tom Simonite this week. Many experts thought that the bill was not strict enough and should make the auditing process tougher to ensure that vendors were being fairly assessed.

OpenAI's zero-shot image model, CLIP

OpenAI revealed two new AI models, but the image recognition model named CLIP appears to have made less of a splash than the GPT-3-based DALL·E.

The company's researchers claim CLIP is more efficient at learning how to recognize and classify images compared to traditional computer vision models. Off-the-shelf systems have to be retrained to identify objects that they weren't originally trained to recognize. But developers don't have to go through all that trouble with CLIP, however.

Instead of being taught to just recognize objects from specific labels, CLIP is pretrained using pairs of images and sentences of text. This means the model is able to learn a "wide variety of visual concepts in images and associate them with their names.

"As a result, CLIP models can then be applied to nearly arbitrary visual classification tasks," OpenAI explained in a blog post.

Jim Keller left Intel, two years later he pops up at an AI startup

Chip veteran Jim Keller – best known for his work at Apple, AMD, and most recently Intel (which he left in 2018) – has joined a Canadian silicon startup.

Toronto-based Tenstorrent, founded in 2016 by ex-AMD and ex-Nvidia chip architect Ljubisa Bajic, has hired Keller to be the company's president and CTO. Designing and manufacturing new chip architectures takes years, and the upstart doesn't have any products on the market yet.

The chip upstart wants to design the whole stack, the hardware and software to support machine learning models in the data center to the edge. "Tenstorrent was founded on the belief that the ongoing shift towards ML-centric software necessitates a corresponding transformation in computational capabilities," CEO Bajic said in a statement.

"There is nobody more capable of executing this vision than Jim Keller, a leader who is equally great at designing computers, cultures, and organizations. I am thrilled to be working with Jim and beyond excited about the possibilities our partnership unlocks."

US ban on selling geospatial AI software to China extended to 2021

The US government has renewed a ban that prevents American companies from selling China software capable of training neural networks to analyze satellite imagery for 2021.

The initial decision came into effect last year on 6 January, and will remain in place for the remainder of this year, a spokesperson from the Bureau of Industry and Security confirmed to El Reg.

Under the rules stated in the latest publication of the Federal Register, the department is able to extend the temporary ban for three years. It is in the process of determining whether the decision should be permanent, but officials have not been able to meet to discuss the issue safely due to COVID.

"However, due to the pandemic, the regime did not convene and therefore did not consider acceptance of the proposal. An extension of time is appropriate in order for the US Government to continue its effort at the Wassenaar Arrangement in 2021," the journal [PDF] said. ®

More about

TIP US OFF

Send us news


Other stories you might like