Several companies offering technology services in Spain have acknowledged that their Artificial Intelligence (AI) algorithms have experienced malfunctions due to sudden changes in people’s behavior caused by the coronavirus pandemic, just as developers have done before. AI in the rest of the world, and that is seen in the errors of facial recognition with masks or in fraud detection systems.
Last week, an investigation by the Massachusetts Institute of Technology (MIT, United States) warned about this phenomenon. According to international technology providers, AI failures have affected very different sectors, from recommendations of streaming platforms to automated inventory systems in stores.
Now, companies that offer IT services in Spain such as Atos Iberia and Fujitsu have confirmed to Europa Press that these problems have also been registered locally. Its impact has focused on facial recognition mechanisms when people wear masks, and more generally they have also affected sectors such as commerce, transport and logistics.
“All AI algorithms base their ‘intelligence’ on the data that was used for their training,” Carlos Cordero, CTO of Fujitsu, assured this agency. In the case of facial recognition mechanisms, these have been trained with full-face images, so the use of masks as protection against the Covid-19 implies problems.
To avoid them, there are already some facial recognition systems that work with a mask, such as those developed by the Barcelona company Herta and recently by the SPEC Group, among others, the latter also capable of measuring temperature at a distance.
For this same reason, other members of the industry at an international level, such as the American startup Workaround and the University of Wuhan (China), have begun to develop databases with photographs of the faces of people with masks collected through the Internet. , in some cases from images of social networks like Instagram. In both cases, these data have been published openly on the Github platform.
The problems have also affected iPhone mobiles and its Face Unlock, which does not work with a mask. For this reason, the latest version of its mobile operating system, iOS 13.5, has automatically started showing the password screen.
This measure speeds up the access process for people wearing coronavirus masks, as it avoids waiting seconds while the Face ID system tries, unsuccessfully, to recognize the user’s covered face.
OTHER ERRORS: DETECTION OF FRAUD
Errors in AI due to human behavior change have also affected other different applications, since they are “extrapolated to practically all AI algorithms that have been trained with information that has changed substantially in a few weeks” , explained the Fujitsu CTO.
Among the AI mechanisms that have failed during this crisis are automatic fraud detection systems such as the use of credit cards, risk calculations in the financial sector, demand predictions and other predictive algorithms that “have been seen little or nothing effective due to the sudden change in the behavior of the society confined to their houses “, as Fujitsu has admitted.
HOW TO ADAPT THE AI
In this situation, the industry is faced with the difficulty of adapting AI systems to continue operating despite the exceptional context of a pandemic. “The algorithms can be adapted by retraining them with current data, representative of the new situation,” José Esteban, director of Innovation at Atos at Iberia, said in statements to Europa Press.
For this process, according to Atos it is necessary that humans supervise the results obtained by the automatic algorithms, reviewing their decisions and annulling them when necessary.
Not all AI mechanisms, however, have received the same impact from the crisis. Although “there is no AI that reliably predicts” phenomena such as the coronavirus, “so that AI does not fail when the environment suddenly changes, a possible solution is systems that continually learn from data,” Esteban said.
In solutions that use unsupervised continuous learning, systems can adapt themselves to abrupt change like today, with massive use of technology for teleworking, consumption and entertainment.
However, these systems can also have problems, since “they cannot always learn quickly enough,” according to Esteban.
Another drawback derived from unsupervised algorithms is that “if they work very well and learn continuously quickly, they can lead to systems starting to behave”, since, for example, their actions are no longer predictable, as the manager of Atos in Iberia.
Portaltic / EP