Learning and Generalizing the Structure of Emotion-Based Reward Associations

• In Event: Probabilistic Emotion Signals Guide Inference and Behavior

Abstract

Humans attend to social signals from others and use those signals to make (and generalize) inferences about social partners. One signal that humans track is the emotional state of other people. However, emotion signals vary based on social partners, situational context, and cultural norms (Barrett et al., 2019), creating a learning challenge. For example, if a social partner is smiling and offers to buy you coffee, you might generalize that positive association as conveying information about the partner (she does nice things), the emotion signal (smiling people do nice things), neither, or both. How do people explore the complex relationship between specific social partners and emotion signals to make predictions? We asked two primary questions:
(1) Do participants learn reward-based associations more slowly when the underlying reward structure tracks with emotion signals versus individual models?
(2) Does prior learning (whether rewards tracked with emotion or model) influence generalization to novel emotions and models?
Method. Adult participants (N=100) completed two phases (“learning phase” followed by “generalization phase”; 48 trials each) of a 4-arm-bandit task. On each trial, participants saw four images on the screen with each image being an individual model displaying an emotional face configuration (RADIATE stimuli; Conley et al., 2018; Figure 1). Participants clicked on an image in search of stars, aiming to obtain as many stars as possible throughout the game. Each image was rewarded at a different (noisy) rate. Reward was assigned to images based on one of two structure conditions: in the emotion structure condition, rewards were assigned to images based on the emotion cued by the image. In the model structure condition, rewards were assigned to images based on the identity of the models. Next, in the “match” condition, reward structure in the generalization phase was consistent with the learning phase (if rewards were predictable based on emotion, they were again predictable based on emotion). In the “mismatch” condition, reward structure in the generalization phase was inconsistent with the learning phase (if rewards were predictable based on emotion, they were then predictable based on model).
Results. Following a pre-registered analytic approach (https://aspredicted.org/H6R_T5N), an effect of trial number indicated that participants learned the reward structure overall (b=0.07, z=10.07, p<.001). An interaction with structure indicated that participants learned the reward structure slower in the emotion versus model condition (b=-0.04, z=-3.25, p=.001). In the generalization phase, there was an interaction between trial number and match (indicating faster learning in the match condition, b=0.05, z=2.61, p=.009). But this relationship was not moderated by structure (b=-0.07, z=-1.86, p=.06; Figure 2).
Conclusion. Participants learned reward associations using social cues including individual identity and emotionality, but learning slowed when there was a mismatch with a prior structure. Additionally, learning occurred more slowly for reward associations predicted by emotional signals. Therefore, while emotion signals are potent sources of information in our social worlds, variability across individuals may present an acute challenge for making inferences about others. We have planned data collection to investigate how children explore this challenging learning problem across development.

Authors

• Rista Plate, University of Pennsylvania
  Presenting Author

• Martin Zettersten, Princeton University