You probably attempt to guess how others will react to what you say or do. Theory of mind is needed to infer others’ thoughts, feelings, and intentions for this work.

MIT neuroscientists have created a computational model that can anticipate other people’s emotions, including joy, appreciation, bewilderment, regret, and shame, matching human observers’ social intelligence.

“These are very common, basic intuitions, and what we said is, we can take that very basic grammar and make a model that will learn to predict emotions from those features,” says Rebecca Saxe, the John W. Jarve Professor of Brain and Cognitive Sciences at MIT’s McGovern Institute for Brain Research and the study’s senior author.

Emotion prediction

Saxe adds that while much study has gone into training computer models to deduce someone’s emotional state from their facial expression, it is not the most significant component of human emotional intelligence. Predicting an event’s emotional response is more crucial.

“Anticipating what others will feel before the thing happens is the most important thing about understanding other people’s emotions,” she explains. It would be disastrous if all our emotional intelligence was reactive.

The researchers utilized “Golden Balls” situations from a British game show to mimic how humans make these guesses. Contestants compete for $100,000. Each competitor secretly chooses to divide or steal the pool after bargaining with their companion. If they divide, each gets $50,000. Splitting and stealing gives the stealer the pot. Both thieves lose.

Contestants may feel delight and relief if both split, astonishment and rage if one steals the pot, or guilt and thrill if one steals.

Three modules were created to anticipate these emotions with a computer model. Inverse planning trains the first module to deduce a person’s preferences and beliefs from their actions.

“This idea says if you see just a little bit of somebody’s behavior, you can probabilistically infer things about what they wanted and expected in that situation,” Saxe adds.

Based on game activities, the first module may forecast participants’ motivations. If someone splits to divide the pot, they likely anticipated the other person to split too. Someone who steals may have expected the other person to steal and wanted to avoid being deceived. They may have anticipated the breakup and tried to take advantage of them.

The model may use information about particular participants, such as their employment, to determine their motivation.

The second module compares game results to player expectations. Using the outcome and competitors’ predictions, a third module predicts their feelings. Human observers predicted participants’ emotions following a given outcome to train this third module. This is a model of human social intelligence, not a description of how individuals feel.

“From the data, the model learns that what it means, for example, to feel a lot of joy in this situation is to get what you wanted, to do it fair, and to do it without taking advantage,” Saxe adds.

Basic beliefs

The researchers employed the three modules on a fresh game show dataset to compare the algorithms’ mood predictions to human observers’. This emotion prediction model outperformed all others.

Saxe believes the model’s effectiveness comes from its combination of crucial characteristics that the human brain utilizes to predict how others would behave. These include calculations of how a person would appraise and emotionally react to a situation based on their wants and expectations, which include money gain and social status.

Our model’s key intuitions are that emotion is about what you intended, expected, happened, and who witnessed. People desire more than goods. She said they want fairness and not to be tricked.

By inverting their model, the researchers explain how we might use people’s actions to deduce their underlying emotions. “This line of work helps us see emotions not just as ‘feelings’ but as playing a crucial, and subtle, role in human social behavior,” says Nick Chater, a behavioral science professor at the University of Warwick who was not involved in the study.

The researchers intend to expand the model’s predictions beyond the game-show context in future work. They’re also developing algorithms that can anticipate the game’s outcome based purely on competitors’ expressions once the results are announced.