COVID-19 diagnosis from cough sounds, computer vision fooled by stickers, and other AI news

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

ForkLog has assembled the week7s most important AI news.

OpenAI7s computer vision fooled by handwritten notes

Researchers from OpenAI’s machine-learning lab discovered that their modern computer-vision system can be fooled with tools no more sophisticated than a pen and notebook.

To fool the program into misidentifying what it sees, it is enough to write the object7s name and paste it on another object.

“We call these attacks typographic. By leveraging the model7s ability to reliably read text, we found that photographs with handwritten text can fool the model,” the OpenAI researchers wrote in the blog.

They noted that such attacks resemble ‘adversarial images’ that can fool commercial machine-vision systems, but they are much easier to create.

Adversarial images pose a real danger to systems relying on machine vision. For example, researchers have shown that they can trick Tesla autonomous driving software into changing lanes without warning simply by sticking certain stickers on the road. Such attacks pose a serious threat to a wide range of AI applications, from medical to military.

OpenAI developers noted that this is an experimental model and not used for commercial purposes. In the future, researchers will work on eliminating such vulnerabilities.

British scientists have created an AI tool for diagnosing COVID-19 from cough sounds

Researchers at the University of Essex have created a COVID-19 screening tool that can diagnose the virus by analysing the sound of a cough. The app, named DeepCough3D, uses AI to analyze the cough audio at frequencies inaudible to humans, The Next Web reports, citing the full study in the IEEE Transactions on Service Computing.

Researchers tested the algorithm on more than 8,000 audio samples, from coughing patients in Spanish and Mexican hospitals. Of them, about 2,000 people had the coronavirus infection, and the rest did not.

According to the developers, DeepCough3D achieves 98% accuracy in determining whether cough samples belong to people with COVID-19. Lead researcher Dr. Javier Andreu-Perez says the tool could fundamentally change the fight against the pandemic.

“It [the tool] is far less invasive than most other COVID-19 tests, and also offers rapid results, opening the way to solutions for preliminary screening,” Andreu-Perez said.

The Essex University team states that their study differs from others in that it demonstrated high accuracy in detecting infection across thousands of clinically confirmed samples tested in certified laboratories.

The researchers also used this tool to classify coughs into three severity levels, which could help health professionals allocate resources such as ventilators.

In the future, scientists plan to conduct interventional studies with the technology and pursue certification of the tool so it can be used in hospitals to diagnose coronavirus infection.

Activists file suit to ban Clearview AI in California

A group of activists filed a lawsuit in California to halt Clearview AI7s operations in the state. The advocates seek to prohibit the company from collecting any biometric data and to delete all Californians7 personal information in its possession.

🧵Today @ConMijente @NorCalResist and 4 activists filed a lawsuit in CA against Clearview AI.

These activists have criticized police and ICE policies and have openly supported Black Lives Matter. JFL and @braunhagey are proud to represent them. 1/https://t.co/ZQLPABiOQ7

— Just Futures Law (@JustFuturesLaw) March 9, 2021

Plaintiffs contend that Clearview AI violates California residents7 right to privacy by collecting their photos from the web and extracting biometric data from them without permission.

The lawsuit also alleges that the company7s facial-recognition technology helps the government monitor protesters, immigrants, and people of color.

To date, Clearview AI software is used by more than 2,400 police departments across the United States. Plaintiffs contend that California law enforcement agencies continue to use it to locate alleged offenders even as several cities in the state have banned facial-recognition technologies.

AI created a psychedelic clip from album covers

Spanish artist Bruno Lopez created a psychedelic music clip by training the algorithm on thousands of Spotify album covers.

First, he built a Python script that pulled album covers for each track featured in Spotify7s official editorial playlists. The dataset totaled about 50,000 images.

These data were used as training data for Nvidia7s StyleGAN2 architecture. After several days of training, the model learned to generate its own album covers.

Next, Lopez built another Python script to organise the results by colours and faces on the covers. He then animated the images to the beat of Sound Stabs7 song “Interference”.

Ericsson taps AI to help hospitals treat COVID-19 patients

Ericsson announced a collaboration with Telia and Sahlgrenska University Hospital to create planning and forecasting models to help the hospital better prepare to admit patients.

The project uses advanced analytics and AI alongside medical information to help the hospital decide what resources will be required. This should help them cope better with demand for their services from COVID-19 patients.

Ericsson said the data could be used to improve estimates of how many people with COVID-19 will require hospitalization. The forecast could then be used to verify the availability of adequate equipment and hospital beds for treatment.

Optibus launches AI platform to plan urban public transport routes

The Israeli startup Optibus unveiled an AI-powered SaaS platform designed to build public transport schedules.

The service can account for demographic data showing how route changes affect residents7 access to public transit, handles workloads such as timetable creation and route planning, which, according to the developers, reduce costs and save time.

Optibus also visualises geographic information when planning routes for cyclists to help them reach to or from public transport.

The company7s solution has already helped optimise transport in more than 450 cities worldwide, including New York, Los Angeles, London, Melbourne, Brasilia, Hong Kong and Singapore. It is now available as a cloud-based app to all comers.

Google Fit adds heart-rate and breathing-rate measurement features using smartphone camera

Google began rolling out heart-rate and breathing-rate tracking via the camera in the Fit app. Both features rely on computer-vision algorithms and enable measuring heart rate and breathing rate without wearable sensors.

To use the heart-rate tracker, users must place a finger on the rear camera. The app will immediately begin counting heartbeats by monitoring subtle changes in skin color.

To use the breathing-rate tracker, users must ensure the front camera has a good view of the torso and remains still for 30 seconds. The feature will then calculate breaths per minute by measuring small chest movements.

The new Google Fit features are currently available only to Pixel smartphones; however the company plans to roll out trackers to more devices soon.

Also on ForkLog:

• The Chinese government unveiled a five-year plan to accelerate the development of advanced technologies, including artificial intelligence and quantum computing.
• Researchers from NVIDIA and Harvard University developed the AtacWorks toolkit for deep learning that speeds up human genome analysis from two days to half an hour.
• A KPMG study showed that the pandemic accelerated AI adoption in enterprises.

Subscribe to ForkLog news on Telegram: ForkLog Feed — the full news feed, ForkLog — the most important news and polls.