Midazolam Enhances Implicit Inference in Humans

People often make logically sound decisions using explicit reasoning strategies, but sometimes it pays to rely on more implicit ‘‘gut-level’’ intuition. The transitive inference paradigm has been widely used as a test of explicit logical reasoning in animals and humans, but it can also be solved in a more implicit manner. Some researchers have argued that the hippocampus supports relational memories required for making logical inferences. Here we show that the benzodiazepene midazolam, which inactivates the hippocampus, causes profound explicit memory deficits in healthy participants, but enhances their ability in making implicit transitive inferences. These results are consistent with neurocomputational models of the basal ganglia–dopamine system that learn to make decisions through positive and negative reinforcement. We suggest that disengaging the hippocampal explicit memory system can be advantageous for this more implicit form of learning. When told that Zoey is older than Jillian, who is older than Allison, one can infer that Zoey is older than Allison. Some researchers have argued that this ability to flexibly draw novel conclusions based on prior premises—to make transitive inferences in this case—depends on specialized neural properties of the hippocampus (Dusek & Eichenbaum, 1997; Eichenbaum, 2004). These authors have shown that even rats can make transitive judgments, but only if their hippocampal system is intact. Other researchers have suggested that simple associative mechanisms can explain transitive responding in animals, and that these mechanisms are independent of (but interact with) the hippocampus (Frank, Rudy, Levy, & O’Reilly, 2005; Frank, Rudy, & O’Reilly, 2003; Frank, Seeberger, & O’Reilly, 2004). This explains why pigeons with hippocampal damage continue to respond transitively (Strasser, Ehrlinger, & Bingman, 2004), and why humans can respond transitively even when they are prevented from becoming explicitly aware of hierarchical relationships (and therefore from employing logical reasoning; Frank, Rudy, et al., 2005). In the study reported here, we investigated in humans the effects of the drug midazolam, which has potent amnestic properties and transiently deactivates the hippocampus (Hirshman, Passannante, & Arndt, 2001; Kobayashi, Fujito, Matsuyama, & Aoki, 2004; Kristiansen & Lambert, 1996; Poncer, Durr, Gahwiler, & Thompson, 1996; Rovira & Ben-Ari, 1993; Thomas-Anterion, Koenig, Navez, & Laurent, 1999). We show that midazolam-induced amnesia is accompanied by a marked enhancement in implicit transitive inference (TI) performance. Thus, our results strongly challenge the notion that the hippocampal explicit memory system is necessary for making relational judgments, and instead suggest that hippocampal disengagement allows the implicit system to have full reign on behavior. Transitive inference is typically evaluated in the laboratory by first asking participants to select one of two stimuli in each of a series of ‘‘premise’’ pairs. The correct choices are learned via error feedback. More concretely, four pairs are presented: A1B , B1C , C1D , and D1E (the plus and minus signs indicate the reinforced and nonreinforced choice, respectively). After learning the correct choices, participants are presented with the novel test pairs AE and BD. Successful AE performance is trivial, because A was always reinforced during training, and E was never reinforced. In contrast, because B and D were reinforced equally often during training, the selection of B over D is taken to indicate that an inference has been made. The question of interest is, which neural mechanisms support such inference-like behavior? Address correspondence to Michael J. Frank, University of Arizona, 1503 E. University Blvd., Building 68, Tucson, AZ 85721, e-mail: [email protected]. PSYCHOLOGICAL SCIENCE 700 Volume 17—Number 8 Copyright r 2006 Association for Psychological Science Insight into this question comes from analysis of exactly what participants learn during the training procedure. In choosing between the members of a premise pair, participants can either explicitly memorize the correct choice or implicitly assign reinforcement value to each of the stimulus elements and choose the one with the higher value. Computational models of transitive responding suggest that these two processes occur in parallel and describe them formally by the use of conjunctive and elemental representations (Frank et al., 2003; Siemann & Delius, 1998), which may be encoded by complementary neural systems. The hippocampus supports explicit memorization of the conjunction of each stimulus pair (AB, BC, etc.) by automatically and rapidly binding together individual elements of the event (Atallah, Frank, & O’Reilly, 2004; Davachi & Wagner, 2002; Frank et al., 2003; O’Reilly & Rudy, 2001). In parallel, the basal ganglia–dopamine system learns an implicit elemental reinforcement value for each stimulus depending on how often its selection is associated with positive versus negative reinforcement (Frank, 2005; Frank, Rudy, et al., 2005; Frank et al., 2004). Correct performance on a novel test pair such as BD depends on the elemental learning system, because the conjunction is novel, but the elements are not. That is, although the B and D stimuli are associated with positive and negative reinforcement equally often during training, they may nevertheless develop asymmetrical associative strengths such that B has a net positive value and D has a negative value (Frank, Rudy, et al., 2005; Frank et al., 2003; Siemann & Delius, 1998; von Fersen, Wynne, Delius, & Staddon, 1991). In brief, because the anchor pairs AB and DE can be solved by simply learning that A is always correct and E is never correct, one does not have to learn anything about the companion stimuli B and D. As a result, B can take on a positive association to support BC performance, and D becomes negative to support CD performance. Therefore, participants may choose B over D simply because B has a higher dopamine reinforcement value; they do not have to explicitly perform any logical reasoning. This mechanism explains why participants can respond transitively even when they are prevented from becoming explicitly aware of logical structure (Frank, Rudy, et al., 2005; Greene, Spellman, Dusek, Eichenbaum, & Levy, 2001) and is consistent with the effects of dopaminergic manipulation on learning in a TI task (Frank et al., 2004). In contrast, a conjunctive strategy that treats each stimulus pair as a distinct event may rapidly produce correct training performance, but would not by itself lead to transitive responding: Upon presentation of the BD test pair, the participant does not have a stored memory representation from which to retrieve the correct response. Thus, this account suggests that successful choice of B over D in nonexplicit TI tasks depends not on the hippocampus, but on reinforcement learning systems that assign differential associative strengths to these stimuli (Frank, Rudy, et al., 2005). This computational framework led us to the counterintuitive prediction that inactivating the hippocampal explicit memory system would enhance implicit TI performance. This prediction stems from various observations that the basal ganglia and hippocampal memory systems interact competitively (Atallah et al., 2004; Packard & McGaugh, 1996; Poldrack et al., 2001; Poldrack & Packard, 2003; Seger & Cincotta, 2005), such that deactivating the hippocampus should lead to greater recruitment of the implicit basal ganglia system. In other words, the more participants memorize stimulus conjunctions during training, the less they learn about individual stimulus reinforcement values, and therefore the worse they will perform on novel test pairs in TI tasks (Siemann & Delius, 1998). This prediction has been formalized in our models, in which learning about conjunctive hippocampal representations can actually block learning of elemental associations (Frank et al., 2003). The present experiments were motivated by the observation that midazolam, by acting on gamma-aminobutyric acid-A (GABA-A) receptors densely expressed in the hippocampus (Montpied et al., 1988), transiently but profoundly impairs explicit memory processes while leaving implicit memory intact (Thomas-Anterion et al., 1999). We tested 23 participants in a double-blind within-subjects design. Each participant was tested once on midazolam and once on saline (both administered by intravenous injection; order counterbalanced). To minimize potential learning effects across sessions, we used a different cognitive learning task in each session, with the Session 1 task selected at random. Participants performed the TI task in one session and a probabilistic selection (PS) task in the other session (Frank et al., 2004). This task was used as a control: Because probabilistic learning recruits the striatum, and actually disengages the hippocampus (Poldrack et al., 2001), we hypothesized that midazolam would have minimal effects on PS performance. In contrast, midazolam should improve TI test performance by preventing memorization of the stimulus pairs and encouraging greater implicit learning of reinforcement value. Finally, to verify that midazolam was effective in inducing amnesia, after each task we gave participants a list of names and told them to remember the names for a subsequent recall test.