Developers teach artificial intelligence to predict people’s feelings

Scientists have long been discussing the possibility of the emergence of artificial intelligence (AI), able to recognize human emotions. Many tools of this kind have already been created – but they are usually based on recognizing emotions in a person’s face, and the accuracy of this method raises many questions. Now, researchers from Massachusetts have developed a model that can not only determine, but predict emotions, and be guided by completely different criteria.

Emotion recognition can be useful in a variety of areas, from detecting emotions when contacting business partners, colleagues or students exclusively online, to using technology in wearable gadgets and medicine. However, existing models of this kind – primarily those that determine emotions from facial expressions – cause quite a lot of criticism.

This week, neuroscientists at the Massachusetts Institute of Technology (MIT) published in a British scientific journal Philosophical Transactions of the Royal Society A article devoted to the construction of a computer model of this kind. It can predict people’s emotions – including joy, gratitude, confusion, embarrassment, regret.

As MIT cognitive neuroscience professor Rebecca Sacks, a participant in the study, notes that such developments often focus on recognizing emotions from facial expressions, but this skill itself is not the most important part of human emotional intelligence.

“The most important thing about understanding other people’s emotions is to predict what people will feel before it happens. If all of our emotional intelligence were exclusively reactive, it would be a disaster,” says Ms. Sachs.

In other words, the model should not determine already expressed emotions, but predict the emotional reaction to certain events.

The task set by the scientists was to make the model predict the emotions of people in situations described by the so-called prisoner’s dilemma. This is an example from game theory, when two “prisoners” independently of each other must decide whether to cooperate or betray.

As the authors of the study note, when building the model, they included several factors that can influence people’s emotions. Among these factors are people’s desires, their expectations in particular situations, the fact that someone is watching them, and so on.

To build the model, the scientists took the scenario of the British television game Golden Balls. It allocates $100,000 for each pair of participants. After a discussion in a pair, each participant must decide whether to share the award or play only for themselves. If both choose to split, each gets $50,000. If one is willing to split and the other decides to play for themselves, the latter gets the full $100,000. If both players refuse to cooperate, neither gets anything.

During the game, participants experience different emotions – from joy and anger to embarrassment. The computer model should act as an observer of the game – and restore those steps that lead the observer to a conclusion about the emotions of the participants. To create a model that predicts these emotions, the researchers identified three separate modules.

The first of these was taught to draw conclusions about a person’s preferences and beliefs based on a person’s actions. For this, the so-called reverse planning was used: conclusions about what a person wanted and expected based on his subsequent actions. The second module compared the results of the game with what each player wanted and expected. Finally, the third module was taught to draw conclusions about the emotions of the players based on their expectations and the outcome of the game.

“From this data, the model learns that, for example, feeling intense joy in such a situation means getting what you wanted, getting it honestly and without deception,” notes Rebecca Sacks.

According to her, building a model shows that people want not only to get money, but also to do it in such a way that they do not look in the eyes of others either as a deceiver or as someone who is easily deceived.

Next, the scientists compared the evaluation of emotions demonstrated by a computer model trained on such examples with the evaluation of people watching the game. According to the authors of the study, the level of accuracy of the Massachusetts model was higher than that of any other model created to determine human emotions. Going forward, the MIT researchers plan to develop the model in such a way that it can make predictions about people’s emotions in more complex situations than an artificially created game situation.

“Researchers have helped build a deeper understanding of how emotions affect our actions; and then, by inverting that model, they show how we can use people’s actions to understand the emotions that underlie those actions. This line of work helps us to see emotions not just as “feelings”, but also as something that plays an essential and at the same time difficult to distinguish role in human social behavior, ”commented Nick Chater, a professor at the British University of Warwick, who specializes in the study of human behavior, commented on the study.

There are many systems for recognizing emotions in a person’s face. True, many experts doubt the accuracy of such tools. For example, HR service HireVue has been using facial emotion recognition during job interviews for some time. However, then the company had to refuse due to criticism of the technology for being “pseudo-scientific” and misjudging real emotions.

Similar systems have already been tried and policemen in different countries, which also drew criticism both from the point of view of intrusion into private life and from the point of view of the accuracy of such tools. EU authorities going to ban systems of this kind, considering them violating human rights.

The main questions among experts are not even the accuracy of determining facial expressions, but the very possibility of understanding the real emotions of a person from them.

According to Kate Crawford, co-founder of New York University’s AI Now Institute, which studies the social impact of AI, these systems themselves are built “on apparently shaky foundations.”

“They claim, if you will, to understand our internal emotional states based on micro facial expressions, tone of voice, and even the way we walk,” Ms. Crawford notes. “And at the same time that these technologies are spreading, a large number of studies show that there is no significant evidence that people have a strong relationship between the emotions they experience and the way their face looks.”

Yana Rozhdestvenskaya