Research lays honest AI groundwork: these neural networks are able to warn when their own analyses are unreliable: "We need to have the ability to rely on high-performance models, but we also need to understand when we will not be able to rely on them."

This is what a researcher from the Massachusetts Institute of Technology, MIT, explains in Science Alert. Alexander Amini works in the institution's artificial intelligence and Computer Science Laboratory, and is part of a team that has just developed neural networks capable of predicting when one should not be trusted... in themselves.

Artificial neural networks are an indispensable element of an artificial intelligence. They articulate the computational processes that allow machines to adopt proposals, solutions and conclusions. As they detail in Science Alert, its purpose is to imitate a human brain taking into account a multitude of factors and putting them on a balance with as many criteria.

Given that AI already makes decisions that can directly affect the lives of human beings —autonomous driving, medical diagnoses— it is essential that these decisions can also be measured with indicators on the confidence that can be given on them.

For this reason, Amini and his team at MIT have developed a model called Deep Evidential Regression, with which neural networks are able to yield a percentage: the degree of confidence that their conclusions yield.

Research lays honest AI groundwork

Neural networks will be able to measure the confidence of their proposals by assessing several issues: how much data the AI has received to process, how trained the AI is, and how accurate and clear the data it has received.

According to Science Alert, this system could be implemented in autonomous vehicles, when the AI of the on-board computer is at a junction and the car does not know whether to cross or not. You may conclude that you can cross, but if it is at night and the cameras are unable to receive well-lit images, the AI itself would be able to warn that it has made a decision with a low confidence index.

At the moment the research has not been peer reviewed and is pending to be presented at the NeurIPS Congress in December. "We are seeing many AI models with neural networks come out of laboratories and land in the real world, in situations that potentially put the lives of human beings at risk," warns Amini, the lead researcher. "Any user of this technology, whether they are doctors or autonomous car passengers, should be aware of the risks they run."

And it is still left so that confidence about AI can be blind. Although a neural network can get it right 99% of the time, the remaining 1% can have fatal consequences.

Research lays honest AI groundwork


Canyoning, Water Trekking, Coasteering, Via Ferrata, Trekking and Orientation Circuits: ADVENTURE GIRONA


More news:

The 5 best-selling mobile models on Amazon Black Friday that are still on sale

If you are looking for a new smartphone that meets that premise of good, nice and cheap you have come to the right place, since here you will find the best-selling phones this Black Friday.

It is clear that being a best seller is a clear thermometer of the quality of the device in question, but it is that here, in addition, not only do the numbers accompany, but they are really competent smartphones in everything they are proposed.

Among the selected ones you will find features for every taste: fast charging, autonomy or performance, but they offer much more.

Huawei P Smart Pro

First of all, the Huawei P Smart Pro stands out for an offer that leaves it below the 200 euro range and stands out for a very interesting photographic section, as well as the absence of notch or hole on the screen to house the selfie camera since it comes out as a Periscope.

It is also a device with a very particular and youthful design that is accompanied by a 4,000 mAh battery, which will be worth you to endure all day.

Little NFC X3

The second exponent is the Poco X3 NFC that will be your perfect ally if you are looking for autonomy, since it is able to reach 3 days with normal daily use. In addition, it offers a lot of customization even though it lacks some polishing and a design with a lot of personality.

Xiaomi Mi 10t Lite 5G

In third place appears a Xiaomi, specifically the Mi 10t Lite 5G which, as its name suggests offers you connectivity to the latest and also has a very interesting design, excellent performance and a fairly rocky autonomy capable of reaching you in two days.

Samsung Galaxy M51

A tough competitor for the little one, since this Galaxy M51 keeps inside a 7.000 mAh battery, able to withstand everything you throw at it. Of course, it is a little more expensive.

Samsung Galaxy S10 Lite

And for the end, the crown jewel simply for the discount it has on it: 260 euros less.

Of course, the Galaxy S10 Lite does not stay alone in that, but delivers a very interesting terminal that you will love if you are looking for fast charging, as well as a more than remarkable photographic section.

You may also find interesting: