The hippocampus and orbitofrontal cortex jointly represent task structure during memory-guided decision making

Abstract

The hippocampus, well known for its role in episodic memory, might also be an important brain region for extracting structure from our experiences in order to guide future decisions. Recent evidence in rodents suggests that the hippocampus supports decision making by representing task structure in cooperation with the orbitofrontal cortex (OFC). Here, we examine how the human hippocampus and OFC represent task structure during an associative learning task that required learning of both context-determined and context-invariant probabilistic associations. We find that after learning, hippocampal and lateral OFC representations differentiated between context-determined and context-invariant task structures. The degree of this differentiation within the hippocampus and lateral OFC is highly correlated. These results advance our understanding of the hippocampus and suggest that the hippocampus and OFC support goal-directed behavior by representing information that guides the selection of appropriate decision strategies.

Keywords: cognitive map; decision making; episodic memory; hippocampus; orbitofrontal cortex; representational similarity analysis; task structure.

Figure 1.. Experimental paradigm and behavioral results

(A) Learning phase. Illustration of a trial from the learning phase. On day 1, participants learned the customer preferences for each food in each store. For the first 8 blocks, subjects are in one store for the entire block and loop through all eight foods four times. In blocks 9–16, subjects loop through all four stores within a single block, looping through all eight foods once in each of the four stores. For each trial, one food item is shown on the store context, and the participant is asked to predict if customers like this food item in this store context or not. After making their predictions, participants were given feedback and were told whether they were correct, along with a picture of a customer’s reaction to that food. At the end, participants would have seen each food item 16 times in each store. (B) Task structure. During the learning phase, participants made predictions about food preferences, and then they received feedback about the actual outcome: customer preference for that food in that store. The matrix illustrates the distribution of outcomes for eight different foods (columns) in each of the four stores (rows). Each cell corresponds to the probability of a like outcome for the given food within the given store; 1.0 means 100% liked, and 0 means 0% liked or 100% disliked. Context-determined foods are defined as those with outcomes that are determined by the store context (left side of the matrix), and context-invariant foods are those with probabilities that are the same across stores (right side of the matrix). The task was designed so that there two task sub-structures: context determined and context invariant. Additionally, there were pairs of similar contexts for which the same foods were liked and disliked in both stores. (C) Learning performance. Performance across all 16 blocks of learning. Foods were separated into four categories on the basis of their like outcome probabilities: 1, 0, 0.75, and 0.25. A repetition consisted of four loops of learning of each food in each store. Therefore, the first repetition was the first time a food category was learned at each store. Learning performance was calculated for each food category by averaging across trials of the foods at each store. Performance corresponded to the proportion of responses of the most likely outcome for a food in a store; like responses for foods with like probabilities of 0.75 and 1 and dislike responses foods with like probabilities of 0.25 and 0. Performance increased with repetitions, and food preferences for each food category were learned with above 60% performance at the end of the learning blocks. Error bars denote within-subject 95% confidence intervals.

Figure 2.. Hippocampus and lateral OFC both represent task structure

(A) Decision phase. The trial procedure was like learning blocks 9–16, with the only difference being that no feedback was given after decisions. The same eight foods were shown in each mini-block in different contexts but in a randomized order. Each run consisted of eight mini-blocks. (B) Decision performance. Foods were categorized as context dependent (CD) or context invariant (CI), and the proportion of optimal decisions was calculated for CD and CI foods separately and averaged across stores for each run. Decision performance showed that participants learned the task structure well enough to make optimal decisions for both CD and CI foods. CD foods were slightly better than CI foods (p = 0.049). (C) Task structure. We calculated pattern similarity (PS) for the two same sub-structure trial pairs (CD-CD or CI-CI) and different sub-structure trial pairs (CD-CI). (D) Task structure representation in hippocampus. Behavioral performance indicated that participants learned the difference between CD and CI task structures. Pattern similarity for same task structure trial pairs was higher than different task structure. (E) Task structure representation in OFC. Lateral OFC pattern similarity was higher for same task structure compared with different task structure. (F) Learning and task structure in hippocampus. We calculated a neural index for task structure representation in hippocampus, PS same-different task structure, and an index for learning differentiation, performance CD-CI for each participant. The learning index (i.e., difference in learning performance between CD and CI trials) was positively correlated with the neural index (i.e., differences in neural representation of same versus different sub-structure). (G) Learning and task structure in lateral OFC. There was a positive correlation between neural index of task structure representation in lateral OFC and learning index. (H) Hippocampus and lateral OFC represent task structure similarly. There was a strong correlation between the hippocampus neural index and lateral OFC neural index. Error bars denote within-subject 95% confidence intervals.