China develops an AI-censorship algorithm; a Japanese robot learns to paint with a brush, 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 tech fields—cybersecurity, and now in the world of artificial intelligence (AI).

ForkLog has gathered the most important AI news from the past week.

Microsoft Announces Acquisition of Siri Developer

Software giant Microsoft announced the acquisition of Nuance for $19.7bn, the company that helped Apple create Siri.

Nuance specialises in the development of voice-recognition technologies for business, including healthcare and customer service.

This is the software giant’s second-largest acquisition after LinkedIn for $26.2bn in 2016.

The deal now awaits antitrust approval. If no issues arise, Microsoft expects to close the deal by the end of 2021.

In China, censorship search engine based on Google’s technology

Chinese researchers said they have developed an AI censor that can find “harmful information” on the internet with an accuracy of over 91%.

Developers say they were aided by the open-source Google BERT language model, which has underpinned the company’s search engine since 2019.

Since BERT cannot analyse texts longer than 512 words, the developers created an algorithm that splits long texts into smaller chunks, runs them through the model, and then reintegrates the results.

According to researchers, their “search engine” can understand hidden subtexts and altered words that Chinese users often employ to circumvent censorship.

They also added that their system could ease the work for censors who currently search for banned information on the internet.

Intel responds to Nvidia Grace processor reveal

Intel chief Pat Gelsinger commented on the announcement of Nvidia Grace, Nvidia’s first server CPU.

He argues that his company is the leader in the market. Gelsinger noted that last week they unveiled a new generation of server processors, designed in part for artificial intelligence tasks.

He also referenced Habana AI processors, whose maker Intel bought in 2019, and noted that close collaboration with Amazon in cloud technology signals success in the space.

“I’d say the idea behind processors is Intel’s foundation. We are now embedding AI into it, and we expect this to be a field where we are on the offensive, not on the defensive,” said Gelsinger.

Twitter to study its algorithms for bias

Twitter announced the launch of the Responsible Machine Learning initiative.

Under the programme, data scientists and engineers at the company will study potential “unintended harm” that its AI could cause and publish the research findings in open access.

First, developers will review the image-cropping algorithm on the network, which has been criticised as “biased toward light-skinned people”.

Twitter also examines the Home timeline, including assessing “fairness of Home Timeline recommendations by racial subgroups” and analysing “recommendations for content for different political ideologies in seven countries”.

It remains unclear what impact this will have on the service. Twitter noted that in some cases it may adjust aspects of the platform based on the research, while other findings may simply lead to “important discussions”.

In Texas, a robo-delivery vehicle will deliver pizza

Domino’s Pizza and autonomous-vehicle maker Nuro began delivering pizzas to residents in one of Houston’s neighborhoods using robo-delivery vehicles.

On certain days and times, customers who place an order online can request delivery by a driverless vehicle that uses radar, 360-degree cameras and thermal imaging to control movement.

Upon arrival, the user will receive a notification on their phone, as well as a PIN code required to access the pizza.

In 2019, Nuro supplied its robo-delivery cars for Kroger deliveries in Houston and Phoenix. In 2020, California authorities allowed the company to test autonomous vehicles on public roads and deliver orders from Walmart and CVS Pharmacy.

Intel plans to release 35,000 autonomous delivery vans by 2028

Mobileye, Intel’s autonomous-delivery unit, and startup Udelv will create 35,000 autonomous delivery vans by 2028.

The first batch of 1,000 robo-delivery vehicles will be produced by 2023. Donlen, which will operate the fleet under a pre-order, has already signed.

All delivery vans produced under the collaboration will be equipped with the Mobileye Drive system, which includes 13 cameras, nine short- and long-range lidars, six radars, and the EyeQ single-chip computing system that manages the sensors.

It remains unclear in which cities Donlen will deploy the autonomous couriers.

Researchers have found serious errors in popular open-source datasets used to train neural networks

MIT researchers have found that ten popular open-source datasets contain serious errors.

In their assessment, the highest rate of inaccuracies is in Google’s QuickDraw hand-drawn-drawings dataset — 10.12% of total labels, and in the ImageNet dataset for testing computer-vision algorithms — 5.8%.

The researchers described the exact labeling mistakes. For example, a mushroom may be labeled as a spoon, a frog as a cat, and a high note by a popular pop singer in an audio file labelled as a whistle.

They added that such slip-ups in test datasets affect the performance of AI algorithms. They urged AI developers to observe stricter “data hygiene” when building their models.

Study: people trust algorithms more than each other

Researchers from the University of Georgia conducted an experiment, which found that people tend to trust an algorithm more to solve a difficult problem than another person or themselves.

In the study, 1,500 participants were asked to view a series of images and determine how many people were in each. As the number of people in the photo increased, participants lost confidence in their answers. They were offered a hint: one of them was generated by a group of a thousand people, the other by an algorithm. Participants predominantly chose the latter.

According to the researchers, the experiment demonstrates that people probably do not understand AI’s real capabilities. They also believe this is a bad sign when a person decides in advance that the algorithm’s answer is better simply because it was generated by a computer.

Scientists present a new method for colouring black-and-white photographs

The team of researchers from the University of Washington, UC Berkeley and Google Research created a new Time-Travel Rephotography technique for colouring black-and-white photographs.

According to the developers, the algorithm does not merely add colour to old images, as dozens of other solutions do, but makes skin appearance more realistic. To achieve this, they trained the model on portraits taken with modern digital cameras.

The researchers noted that a hundred years ago cameras captured a limited spectrum and “filtered” red hues. As a result, black-and-white portraits looked more wrinkled and unnatural than they actually were, distorting our sense of how people looked in that era.

In Japan, a robot has learned to paint paintings

A group of researchers from IBM Japan, the University of Tokyo and Yamaha Motors created an AI-powered robot artist that can paint with a movable arm and a set of brushes, changing them autonomously.

The robot can be programmed to create a painting based on the number of brushstrokes, determining its complexity.

For example, if asked to paint a butterfly in 30 strokes, it will create an abstract image. If the number is increased to 300, a more realistic painting will emerge.

The robot can work with watercolour and acrylics, and it can mix colours and dilute paint with water.

In the future, the developers aim to equip the robot with computer vision so it can assess progress as the painting is being created.

Also on Forklog:

• IBM released a set of practical Qiskit Machine Learning modules, which form part of its open-source quantum software.
• Nvidia introduced new AI processors and development tools, and spoke of record revenue for the incomplete first quarter of the 2022 financial year.
• The Russian Bar Association proposed regulating the AI market in the country.
• In the EU, they will consider a ban on the use of AI for citizen surveillance.
• Experts said that the number of deepfakes online doubles every six months.
• Elon Musk stated that Tesla will become one of the largest AI companies in the near future.

Subscribe to ForkLog news on Telegram: ForkLog AI — all the news from the AI world!