Learning Involves Flexibly Switching Between Imitation and Emulation Strategies

Post by Shireen Parimoo

What's the science?

Much of learning occurs through observation. Children learn how to act and interact with objects by watching their parents, while adults observe others to figure out how to act in an unfamiliar situation. Two observational learning strategies include imitation and emulation, which are beneficial under different circumstances. In choice imitation, an individual’s choices are based on simply copying someone else's previous actions, whereas emulation involved inferring the goals and intentions of others.  How one learning strategy is selected over the other, and the mechanism that supports this selection process remains unknown. This week in Neuron, Charpentier, and colleagues used fMRI and computational modeling to determine when and how the imitation and emulation strategies are selected during observational learning, as well as the neural signatures associated with each strategy.

How did they do it?

Across two studies, participants completed an observational learning task with the goal of choosing a slot machine to obtain a valuable token. There were three slot machines and three tokens (red, blue, green), and the slot machines dispensed tokens with certain probabilities (e.g. 75% red, 20% green, 5% blue). Participants were told that the valuable token would switch throughout the task but were not told which token was valuable. Instead, they had to learn this by watching a partner who was aware of the valuable token play (“observe” trials) and use this knowledge to choose the slot machine on their turn (“play” trials). The authors manipulated how frequently the valuable token changed (1 = low volatility, 5 = high volatility), as well as the probability distribution of getting the tokens from each slot machine (high or low certainty). The authors assessed when participants’ choices were guided by their partner’s past actions (suggesting choice imitation strategy) or by inferring which token was valuable from the partner’s choice (suggesting emulation strategy).

To determine how learning strategies were selected, they tested a series of computational models, including emulation and imitation models, as well as an arbitration model - a model that involves comparing both learning strategies and selecting the best one at a given point in time. The single-strategy models were based on choosing the action that was most recently selected by the partner (imitation model) or updating the probability that each token is valuable given the partner’s choice (emulation model). The arbitration model computed the reliability of emulation; when high, emulation was assigned a higher weight, and when low, imitation became more likely. The models were then compared with each other to identify which one predicted behavior. Finally, the authors examined neural activity associated with emulation and imitation when observing the partner's choice (this is when learning occurs), and the arbitration signal. 

What did they find?

Participants used both the choice imitation and emulation strategies to make decisions. The emulation model only predicted choices that were driven by token value, which was common when uncertainty about the slot machine’s token distribution probability was low. On the other hand, the imitation model only predicted imitation-guided choices, and this strategy was preferred under uncertain but low volatility conditions (i.e. the valuable token was not likely to change). The arbitration model was most successful in predicting behavior associated with both learning strategies. Specifically, emulation was selected under volatile and certain conditions, while imitation of the partner’s most recent choice was preferred when emulation was not reliable. These results show that people adaptively use both imitation and emulation strategies under different conditions during observational learning.

Emulation activated regions of the mentalizing network, including the dorsomedial prefrontal cortex, insula, and the temporoparietal junction, while activity in motor regions like the pre-supplementary motor area and the motor cortex was observed during imitation. These neural signals likely reflect updating of token values (emulation) or of the preferred slot machine choice (imitation) based on the partner’s actions. Finally, activation of cognitive control areas, such as the ventrolateral prefrontal cortex and the anterior cingulate cortex, was associated with arbitration between the two learning strategies.

What's the impact?

This is the first study to demonstrate that flexible arbitration between the emulation of others’ goals and imitation of others’ actions occurs during observational learning. The authors further showed that distinct brain networks are recruited for imitation and emulation, while regions previously implicated in cognitive control support arbitration between the two strategies. These findings pave the way for future research on how arbitration between these learning strategies changes in development.

Charpentier et al. A neuro-computational account of arbitration between choice imitation and goal emulation during human observational learning. Neuron (2020). Access the original scientific publication here.