Quantum machine learning hits a ceiling, Microsoft creates ‘endless remix’ and other AI news

We aim to inform readers not only about developments in the Bitcoin industry, but also to report on what is happening in adjacent tech spheres — cybersecurity, and now the world of artificial intelligence (AI).

ForkLog has gathered the week’s most important AI news.

Google unveiled a suite of AI features at the I/O conference

The tech giant unveiled a suite of AI-powered features and products.

The tech giant showed its language model Multitask Unified Model, which developers plan to integrate into the search engine. The company says MUM is 1000 times more powerful than BERT, trained on 75 languages, understands the context of a query, and searches for information across multiple sources.

Google introduced the neural network LaMDA, which communicates like a living person and maintains a natural dialogue on a range of topics. In a demo shown by CEO Sundar Pichai, the algorithm played the role of a dwarf planet Pluto and answered questions about itself.

For now LaMDA understands only text, but the development team is working to enable the model to understand images, video and audio. In the future Google plans to integrate it into Assistant and Search.

The Photos app will soon gain AI-powered features, including cinematic photos, automatic sorting of photos by items depicted, and memories albums.

Google plans to add a skin, hair and nail state detection feature based on photos uploaded by patients to Search or other apps.

The system can recognise 288 skin conditions. It was trained on a dataset of 65,000 images of diagnosed conditions, millions of tagged photos that have previously sparked concerns, and thousands of photos of healthy skin across all skin tones.

According to developers, the app is not intended to replace medical diagnosis and treatment. Its goal is to help doctors offer more personalised care.

Google also announced a new generation of TPUv4 processors. They are nearly three times as powerful as the previous version, and a 4096-chip configuration will deliver more than one exaFLOP of performance, the company said. TPUv4 will form part of Google’s cloud platform by the end of 2021.

In addition, Google plans to build the first commercial quantum computer by 2029.

California regulators to scrutinise Tesla’s autonomous advertising

The California Department of Motor Vehicles began an investigation into Tesla. The agency intends to determine whether the company’s advertising for Full Self-Driving is misleading.

The probe will cover the following Tesla features:

• navigation;
• automatic lane change;
• autoparking;
• warning to take control;
• Autopilot computer;
• detection of traffic signals and STOP signs.

If it turns out Tesla violated the rules, the regulator could suspend or revoke approvals to deploy autonomous vehicles, and revoke the manufacturer and dealer licenses.

Earlier, a self-driving Tesla crashed into a parked police car in a Seattle suburb.

According to the driver, he was merely watching the situation and expected the autopilot to detect an obstacle and avoid the car, not ram it.

Police fined the Tesla owner.

Waymo autonomous vehicle blocked the highway and tried to flee from rescue crews

The Waymo self-driving car blocked a lane in Arizona and then attempted to flee from a rescue crew. A blogger who tests the car and posts trips on YouTube reported this.

The car was guided by orange cones laid out by the road service.

Waymo said the vehicle did experience difficulties and called for support. It’s likely the operator gave wrong instructions, exacerbating the situation.

A rescue crew arrived at the scene. While it was on the way, the car did start moving again but soon stopped, blocking a two-lane road for several minutes.

The problem was resolved only after the road service removed the cones. However, a few minutes later more cones forced the autonomous vehicle to stop again. In the end the driver had to take control and drive the car to its destination.

Facebook AI researchers develop neural network that forgets irrelevant information

The engineers from Facebook AI Research developed the AI technique Expire-Span, which helps neural networks forget irrelevant information.

The algorithm first predicts which information will be most useful to the network in a given context, then assigns an expiry date to it. Neural networks will be able to retain relevant information longer, continuously clearing memory to “forget” irrelevant data points.

According to the researchers, the algorithm is still in early stages. In the future they hope to create a system that resembles human memory.

Scientists claim to have reached the limit of quantum machine learning

American scientists say they have identified a limit that quantum machine learning cannot surpass.

According to them, the discovered gap in the conventional understanding of information encryption will prevent reconstructing chaotic and random information, for example from black holes.

They considered a classic thought experiment by Hayden-Preskill, in which Alice throws some information into a black hole, and Bob tries to decode it and obtain the original records with the help of quantum machine learning.

The new theorem shows that reconstructing it will be practically impossible due to the learnability problem, which exponentially complicates the task as the studied system grows.

Nevertheless, the researchers noted that most physical processes of potential interest are simple and structured. This means the scenario described in the study will be extremely rare in practice.

The world is not ready for AI hackers, says security expert

Security expert Bruce Schneier said humanity is not ready for hackers armed with artificial intelligence.

“Any good AI system will, of course, find breaches,” he said.

Schneier added that the main problem is that people “do not close all paths for hacking.” Once AI starts to look for them, vulnerabilities of that scale will emerge that practitioners simply cannot deal with.

Nevertheless the expert is confident that the edge will eventually lie with defense tools, as AI can be used to detect and close gaps.

Russia to offer state data to AI developers

The Ministry of Digital Development, Communications and Mass Media proposed selling access to state data to AI developers. This follows a project presented at a thematic expert session of the ANO “Digital Economy.”

According to reports, the current version of the document includes datasets such as regional price levels, medical examination results and their diagnoses, soil composition for agricultural land, and other information.

The sets will be prepared for the state operator, accredited organisations, or the ministries themselves.

Prices for datasets will be set depending on their complexity and volume. Access to classified data will be granted only to organisations accredited by relevant agencies.

Voice cloning and monetisation platform for celebrities unveiled

Veritone launched the Marvel.AI platform for cloning voices, enabling celebrities to monetise their speech.

According to developers, anyone can create a digital copy of their voice and put it up for sale. Samples can then be used to voice news reads, commercials and other content.

The platform also allows digitising the voices of deceased individuals. For this, archived voice recordings are required.

Researchers build AI that recognises tiger mosquitoes

The creators of the Mosquito Alert platform and Budapest University developed an AI algorithm that recognises tiger mosquitoes in photographs.

Users of the service between 2015 and 2019 sent scientists 7,168 images of insects. With their help, the algorithm learned to correctly classify tiger mosquitoes in images in 96% of cases.

To date, insects in photographs are identified by entomology experts, which is time-consuming and requires years of experience. AI could speed up the process and improve control of tiger mosquitoes.

The researchers noted that the insect threatens human health. It transmits diseases for which there is no vaccine. The only preventive measure is combating vectors.

To this end, Mosquito Alert asked users in tropical countries to photograph and send mosquito photos to the platform so scientists could analyse and monitor its spread in a timely manner.

Also on ForkLog:

• Researchers created AI, predicting the violin’s timbre based on its pattern.
• Opinion: AI will surpass humans, but not soon.
• Amazon extended the moratorium on facial recognition technology by the police.
• Microsoft used AI to create an “endless remix”.
• TSMC announced a breakthrough in 1-nanometer chip production.

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