Positional inference in rhesus macaques

Abstract

Understanding how organisms make transitive inferences is critical to understanding their general ability to learn serial relationships. In this context, transitive inference (TI) can be understood as a specific heuristic that applies broadly to many different serial learning tasks, which have been the focus of hundreds of studies involving dozens of species. In the present study, monkeys learned the order of 7-item lists of photographic stimuli by trial and error, and were then tested on "derived" lists. These derived lists combined stimuli from multiple training lists in ambiguous ways. We found that subjects displayed strong preferences when presented with novel test pairs. These preferences were helpful when test pairs had an ordering congruent with their ranks during training, but yielded consistently below-chance performance when pairs had an incongruent order relative to training. This behavior can be explained by the joint contributions of transitive inference and another heuristic that we refer to as "positional inference." Positional inferences play a complementary role to transitive inferences in facilitating choices between novel pairs of stimuli. The theoretical framework that best explains both transitive and positional inferences is a spatial model that represents both the position and uncertainty of each stimulus. A computational implementation of this framework yields accurate predictions about both correct responses and errors for derived lists.

Superstitious learning of abstract order from random reinforcement

“…Thus, if subjects begin the task believing that each stimulus is pre-assigned to a certain rank, it stands to reason that this a priori representation would be resilient against at least some counterfactual information. This view is consistent with previous evidence that when monkeys and humans were shown the "derived" pairings consisting of each stimulus from distinct trained ordered set, they relied on the known ranks held by each stimulus in their original set to judge the novel combination, suggesting that they spontaneously assume that the two sets use the same ranking scale even when there is no logical necessity for this being the case [31,32,33]. More broadly, this view is consistent with the proposed role of generalization heuristics in guiding exploration in complex contexts under high uncertainty [34,35,36].…”

Superstitious learning of abstract order from random reinforcement

“…This view is consistent with other studies in which monkeys and humans were shown “derived” pairings that mixed stimuli from two different pretrained objectively ordered sets. In such tasks, subjects rely on the known ranks held by each stimulus in their original sets to judge the novel across-set combinations, suggesting that they spontaneously assume that the two sets use the same ranking scale even when there is no logical necessity for this being the case ( 33 – 35 ).…”

“…Popular methods for coding generalization in RL algorithms involve inferring an abstract latent state from one context and applying it to guide learning in a new context ( 39 ). In our case, since each state (each trial) is independent of every other, generalization may be operationalized with the equation , where E ( R | s , a ) is the expected reward given state s and action a and refers to an internal ranking over all stimuli analogous to those proposed in positional inference models in the previous TI literature ( 35 ).…”