Model-observer similarity, error modeling and social learning in rhesus macaques

Abstract

Monkeys readily learn to discriminate between rewarded and unrewarded items or actions by observing their conspecifics. However, they do not systematically learn from humans. Understanding what makes human-to-monkey transmission of knowledge work or fail could help identify mediators and moderators of social learning that operate regardless of language or culture, and transcend inter-species differences. Do monkeys fail to learn when human models show a behavior too dissimilar from the animals' own, or when they show a faultless performance devoid of error? To address this question, six rhesus macaques trained to find which object within a pair concealed a food reward were successively tested with three models: a familiar conspecific, a 'stimulus-enhancing' human actively drawing the animal's attention to one object of the pair without actually performing the task, and a 'monkey-like' human performing the task in the same way as the monkey model did. Reward was manipulated to ensure that all models showed equal proportions of errors and successes. The 'monkey-like' human model improved the animals' subsequent object discrimination learning as much as a conspecific did, whereas the 'stimulus-enhancing' human model tended on the contrary to retard learning. Modeling errors rather than successes optimized learning from the monkey and 'monkey-like' models, while exacerbating the adverse effect of the 'stimulus-enhancing' model. These findings identify error modeling as a moderator of social learning in monkeys that amplifies the models' influence, whether beneficial or detrimental. By contrast, model-observer similarity in behavior emerged as a mediator of social learning, that is, a prerequisite for a model to work in the first place. The latter finding suggests that, as preverbal infants, macaques need to perceive the model as 'like-me' and that, once this condition is fulfilled, any agent can become an effective model.

Figure 1. Schematic representation of a learning session.

One of the three models first showed six pairs, the so-called ‘social pairs’, modeling the correct response (success) for three of them, and the incorrect response (error) for the other three. Then, the observer monkey was tested. Three additional pairs were inserted in the list (the ‘individual pairs’) and the now complete 9-pair list was presented 10 times to the monkey. The observer’s performance was thus evaluated over 10 hands-on trials for all pairs, whether ‘individual’ or ‘social’.

Figure 2. Effectiveness of the monkey model and of the two human models (‘stimulus-enhancing’ vs. ‘monkey-like’).

A positive learning Δ denotes fewer errors for ‘social’ pairs than for the ‘individual’ pairs tested during the very same sessions, i.e. a beneficial model. A negative learning Δ denotes more errors for ‘social’ than for ‘individual’ pairs, i.e. a detrimental model. Results are illustrated for each monkey and for the group. Monkeys are grouped per trio of housemates according to their rank in the group hierarchy. For the detrimental ‘stimulus-enhancing’ model, two bars were truncated to keep the figure balanced. The actual scores were −103 for the top-ranking male (♂ 1) and −63 for the bottom-ranking female (♀ 3). Note that although 5/6 monkeys benefited slightly more from the monkey than from the ‘monkey-like’ human, the reverse pattern did occur too (♂ 2), hence, the indistinguishable group means yielded by the two beneficial models.

Figure 3. Learning from the three models’ successes vs. errors.

Learning Δs were calculated separately for the ‘social’ pairs for which the model’s demonstrated the correct response and for the ‘social’ pairs for which the model’s demonstrated the incorrect response relative in both cases to scores for the ‘individual’ pairs tested during the same sessions. Group scores (mean + or – sem) are shown for each model. Note that errors widened the gap between the two effective and the ineffective models.