summary: Researchers have developed a computational model that encapsulates attributes of social intelligence typically associated with human “theory of mind” and can predict human emotions.
This model predicts emotions such as joy, regret, and bewilderment based on the prisoner’s dilemma game theory scenario. The system uses factors such as a person’s desires, expectations, and whether their behavior is being observed to predict their emotional response.
According to the researchers, the model outperformed previous models in predicting emotions, representing a significant advance in emotional artificial intelligence.
Important facts:
- A computational model designed by MIT neuroscientists predicts human emotions based on whether human desires, expectations, and behaviors are observed.
- The AI system was trained using a scenario from a British game show called ‘Golden Ball’, which operates on the principles of prisoner’s dilemma game theory.
- The model’s success depends on its ability to express emotional predictions more accurately than previous models, incorporating the core intuitions that the human brain uses to predict the emotional responses of others.
sauce: Massachusetts Institute of Technology
When interacting with others, you probably spend part of your time anticipating how they might feel about what you say or do. This task requires a cognitive skill called theory of mind that helps us infer the beliefs, desires, intentions and emotions of others.
Neuroscientists at MIT have now designed a computational model that can approximate the social intelligence of human observers to predict the emotions of others, including joy, gratitude, confusion, regret and bewilderment.
The model was designed to predict the emotions of people caught in a situation based on the prisoner’s dilemma, a classic game theory scenario in which two people decide whether to cooperate or betray their partners. .
To build the model, the researchers looked at a number of things that are hypothesized to influence people’s emotional responses, such as their desires, their expectations in a particular situation, and whether someone is watching their behavior. Incorporated some factors.
“These are very common basic intuitions, and what we’ve said is that we’re taking advantage of that very basic grammar to create models that learn how to predict emotions from those features. It’s possible,” says Rebecca Sacks, a professor of brain science at John W. Jaab. Dr. Cognitive Sciences is a member of the Massachusetts Institute of Technology McGovern Brain Institute and a senior author on this study.
Sean Dae Houlihan PhD ’22 is a postdoctoral fellow at the Newcomb Institute for Computational Sciences at Dartmouth College and the lead author of this paper. Philosophical Deal A.
Other authors include Max Kleiman-Weiner PhD ’18, postdoctoral fellow at MIT and Harvard. Luke Hewitt PhD ’22, Visiting Scholar at Stanford University. Joshua Tenenbaum is Professor of Computational Cognitive Science at MIT and a member of the Center for Brain, Mind and Machine and he is a member of MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL).
emotion prediction
Much research has been done on training computer models to infer emotional states based on facial expressions, but it’s not the most important aspect of human emotional intelligence, Sachs says. Much more important is being able to predict someone’s emotional reaction to an event before it happens.
“The most important thing about understanding other people’s emotions is predicting how they’ll feel before things happen,” she says. “If all our emotional intelligence were reactive, it would be a disaster.”
To model how human observers would make these predictions, the researchers used a scenario taken from a British game show called ‘Golden Ball’. On the show, contestants bet her $100,000 pot to pair up.
After negotiating with their partners, each contestant secretly decides whether to split the pool or try to steal it. If both decide to split, each will receive his $50,000. If one person splits and one person steals, the person who stole gets the whole pot. If both try to steal, no one gets anything.
Depending on the outcome, contestants may experience different emotions. It’s a mixture of joy and relief when both contestants split, surprise and anger when one opponent steals the pot, and perhaps guilt and excitement when one of them succeeds in stealing. .
To create a computational model that can predict these emotions, the researchers designed three separate modules. The first module is trained to infer a person’s preferences and beliefs based on their behavior through a process called reverse planning.
“The idea is that by looking at someone’s behavior for a bit, you can probabilistically infer what they wanted and expected in that situation,” Sachs said.
Using this approach, the first module can predict a contestant’s motivation based on their in-game behavior. For example, if someone was trying to share a pot and decided to split it, we could infer that that person expected others to split as well. If someone decides to steal, they may expect the other person to do so and not want to be deceived. Or maybe they expected the other to split and wanted to take advantage of it.
The model can also integrate knowledge about a given player, such as the player’s occupation, to infer the player’s most likely motivations.
The second module compares the outcome of the game with what each player wants and expects to happen.
The third module then predicts how the participants will feel based on what we know about the outcome and their expectations. This third module of his was trained to predict emotions based on predictions from human observers of how a contestant would feel after a particular outcome.
The authors argue that this is a model of human social intelligence, designed to mimic the way observers reason about each other’s emotions and how people actually feel. It emphasizes that it does not model
“The model learns from data that, say, feeling great pleasure in this situation means getting what you want, doing it fairly, and doing it without taking advantage of it. It learns,” says Sachs.
core intuition
Once the three modules were up and running, the researchers used them on a new dataset from the game show to determine how the model’s emotion predictions compared to those made by human observers. Did. This model performed much better than previous emotion prediction models on that task.
The model’s success comes from incorporating key elements that the human brain also uses to predict how others will react to certain situations, Sachs said. says. These include calculations of how a person assesses situations and reacts emotionally based on their desires and expectations. This concerns not only material gains, but also how they are viewed by others.
“In our model, the underlying mental state of emotion is the core intuition of what you wanted, what you were expecting, what happened, and who saw it. And people don’t just want things, they don’t just want money, they want to be fair, but they also want to be duped and cheated. No,” she says.
In future work, the researchers hope to adapt the model to be able to make more general predictions based on situations other than the game show scenarios used in this study. In addition, we are working on creating a model that can predict the development of the game based on the facial expressions of the contestants after the announcement of the results.
Funding: This study was funded by the McGovern Institute. Paul E. and Lyra Newton Brain Science Award. The center of the brain, mind and machine. MIT-IBM Watson AI Lab. and interdisciplinary university research initiatives.
About this AI and emotion research news
author: Sarah McDonnell
sauce: Massachusetts Institute of Technology
contact: Sarah McDonnell – MIT
image: Image credited to Neuroscience News
Original research: open access.
“Emotion Prediction as Computation for Generative Theory of Mind” by Rebecca Sachs et al. Philosophical Deal A
overview
Emotion Prediction as a Computation of Generative Theory of Mind
Observers can systematically and subtly predict what emotions participants will experience from sparse descriptions of events.
We propose a formal model of sentiment prediction in the context of public high-stakes social dilemmas. This model uses inverse planning to infer a person’s beliefs and preferences, including social preferences for fairness and maintaining a good reputation.
The model then combines these inferred mental contents with events to compute a “rating” of whether the situation fits expectations and meets preferences.
We learned a function that maps computed ratings to emotion labels and matched the model with human observer quantitative predictions for 20 emotions, including joy, relief, guilt, and envy. I can. A comparison of models shows that estimated monetary preferences alone are not sufficient to explain the observer’s sentimental expectations. Inferred social preferences are incorporated into nearly all emotion predictions.
Both human observers and models use minimal personalization information to tailor their predictions of how different people will react to the same event.
Our framework therefore integrates the concepts of reverse planning, event evaluation, and emotion into a single computational model to reverse engineer people’s intuitive emotional theories.
