Hippocampal and ventral medial prefrontal activation during retrieval-mediated learning supports novel inference

Abstract

Memory enables flexible use of past experience to inform new behaviors. Although leading theories hypothesize that this fundamental flexibility results from the formation of integrated memory networks relating multiple experiences, the neural mechanisms that support memory integration are not well understood. Here, we demonstrate that retrieval-mediated learning, whereby prior event details are reinstated during encoding of related experiences, supports participants' ability to infer relationships between distinct events that share content. Furthermore, we show that activation changes in a functionally coupled hippocampal and ventral medial prefrontal cortical circuit track the formation of integrated memories and successful inferential memory performance. These findings characterize the respective roles of these regions in retrieval-mediated learning processes that support relational memory network formation and inferential memory in the human brain. More broadly, these data reveal fundamental mechanisms through which memory representations are constructed into prospectively useful formats.

Figure 1

Experimental design. (A) Color photographs of object (O) and scene (S) stimuli were organized into groups of three stimuli (triads) presented as two overlapping associations (AB, e.g., “zucchini–pail”, and BC, e.g., “pail–truck”). Triads consisted of one of four types: three objects (OOO), two objects and a scene (OOS), three scenes (SSS), and two scenes and an object (SSO). (B) Participants learned the overlapping associations from each triad during blocked-design encoding runs (see Experimental Procedures). The AB and BC associations of all triad types were repeated three times within a functional run in an interleaved manner (AB, BC, AB, BC, AB, BC). (C) After each encoding run, participants were tested on directly learned associations (AB, BC) as well as inferential relationships (AC), using a two-alternative forced-choice judgment. See also Fig. S1.

Figure 2

Multivoxel pattern analysis (MVPA) strategy. MVPA classifiers trained to differentiate brain patterns associated with object and scene processing (see Experimental Procedures) indexed content-specific activation during each encoding condition of the associative inference task. Classifier outputs were compared across AB repetitions when presented information was from the same content class (e.g., two objects for OOO and OOS triads), but the content class of the third, unseen triad member differed (object vs. scene). AB Repetition 1. On the first AB repetition, classifier output is predicted to reflect the content of presented information and not differ for associations comprised of the same content class. AB Repetition 2 & 3. On the second and third AB repetitions, classifier output is predicted to reflect not only presented content, but also reactivated, overlapping BC associations. In this example, two objects are presented, but scene classifier output is predicted to be greater for OOS triads relative to OOO triads, reflecting the reactivation of the associated scene for OOS triads (e.g., “lake scene”), but a third object for OOO triads (e.g., “truck”). The difference in scene classifier outputs across AB repetitions of these triad types (OOO vs. OOS) serves as a critical reactivation measure. A similar analysis (not depicted) compared classifier output across AB repetitions for SSS and SSO triads.

Figure 3

Reactivation of prior event content during encoding of related associations. (A) Difference in scene classifier output across repetitions of AB associations for OOS relative to OOO triads. (B) Difference in object classifier output across repetition of AB associations for SSO relative to SSS triads. For both (a) and (b), error bars denote standard error of the mean; asterisk denotes significant difference between compared classifier outputs at p < 0.05. See also Fig. S2. (C) Across-subject correlation between reactivation index (collapsed across object and scene reactivation measures) and inference (AC) performance. Greater reactivation index was associated with superior AC accuracy.

Figure 4

Across-participant correlation between activation decreases (first – last parameter estimate) in bilateral anterior MTL cortex and the reactivation index. Greater learning-related decreases in anterior MTL cortex were associated with greater reactivation of unseen, related stimulus content. See also Fig. S3.

Figure 5

Across-participant correlation between learning-related changes in hippocampus and VMPFC and subsequent inference performance. (A) Greater learning-related hippocampal decreases (first – last parameter estimate) across encoding repetitions were associated with greater AC performance at test. (B) Greater activation increases in VMPFC (last – first parameter estimate) across encoding repetitions were associated with greater AC performance at test.

Figure 6

Functional connectivity between hippocampus and VMPFC across encoding repetitions displayed separately for each run. A significant increase in hippocampal-VMPFC connectivity was observed across encoding repetitions, but connectivity between these regions did not change as a factor of functional run. Asterisk denotes significant increase in connectivity within an individual run. See also Fig. S4.