The researchers used functional MRI (fMRI) to map areas of brain activation in 24 college students undergoing an associative learning and generalization task. The participants were shown pairs of images—faces and scenes—learning to associate the two. After the learning phase, the students were then asked to link faces and scenes in a test phase. Since some faces and scenes were paired more than once, the researchers were able to test how well the subjects generalize based on overlap. For example, if Mary’s face had been paired with scenes of an oak tree and a sunset, but John’s face had been only paired with the oak tree, then would the subjects generalize at test phase by linking John’s face with the sunset scene as well? That is, in fact, what the researchers found. But it was not so much that the subjects were able to generalize in this manner, but what goes on in the brain when they do, that supports the “integrative encoding” hypothesis.

If generalization is to be explained by the alternative “logical inference” model—where memories are retrieved and analyzed on the spot—then it should correlate with activation of the brain areas involved in the process. However, Shohamy and Wagner found no link between hippocampal activation and performance in the generalization part of the tests. That suggests that there is no additional retrieval process going on during generalization. On the other hand, the researchers did find a correlation between generalization prowess and hippocampal and midbrain activation during the learning phase. “We found that all the action happened essentially while people were experiencing the individual events, what we call the premise event. That is when people who later generalize well showed a lot of hippocampal activity. People who later didn’t generalize well didn’t show this early on,” said Shohamy. The results show that the brain events that predict the behavior were happening not at the time of generalization but earlier on, at the time of learning. “That was really the key thing,” said Shohamy. Generalization was also much more rapid than might be expected if the “logical inference” model held true.

The areas of the brain that were predictive of generalization ability were the hippocampus and the ventral tegmental area (VTA), and substantia nigra in the midbrain. “That is a relatively novel finding,” suggested Shohamy. “Generally the role of dopamine in learning is thought to be separate from what the hippocampus is doing—inputs for procedural learning and habits, exactly the kind of thing that people think is intact in people with Alzheimer disease and hippocampal damage,” said Shohamy. However, more recent research, including this study, suggests dopaminergic involvement is not so simple and that it may modulate what happens in the hippocampus. Alison Adcock at Duke University, Durham, North Carolina, for example, has shown that dopaminergic innervation may link motivation with better encoding in the hippocampus (see Adcock et al., 2006) and John Lisman at Brandeis University, Waltham, Massachusetts, and Anthony Grace at the University of Pittsburgh, Pennsylvania, have theorized that VTA and hippocampal neurons form a functional loop (see Lisman and Grace, 2005).

Exactly how dopaminergic innervation influences hippocampal memory is not clear. One possibility, posited by Dharshan Kumaran, Wellcome Trust Center for Neuroimaging, London, and Emrah Duzel, University College London, in an accompanying Neuron preview, is that dopamine alters neuronal plasticity by inducing synaptic proteins. Since that process would take some time, Kumaran and Duzel suggest that adjusting the interval between presenting the overlapping pairs of visual stimuli might be insightful.

“This is pretty exciting work,” said Stephan Heckers, Vanderbilt University, Tennessee, in an interview with ARF. Heckers also studies learning and memory in humans and previously showed that generalization is related to activation of the hippocampus (Heckers et al., 2004). “The relationship between the hippocampus and the ventral tegmental area is not entirely novel, but what they have shown is that it takes place at the time of encoding. That is novel, and this might be the first study that supports the Lisman and Grace model,” he said. Heckers also found that in cued-recall tests only two areas of the brain predict accuracy, the hippocampus and the VTA, and he has seen other links between the VTA and memory. “Now I’m intrigued, because we have seen something similar not only during encoding but also during the retrieval phase,” he said.

Could this interplay between the dopaminergic system and the hippocampus explain, even partly, cognitive dysfunction in Parkinson disease or even affect cognition in AD? Shohamy said it is not so clear. “It is complicated. Midbrain dopamine modulates several different systems,” she said. The striatum, for example, has received a lot of attention, and there has been some work linking it to cognition in PD, but in this study Shohamy found no correlation between generalizability and the striatum. Also, in collaboration with colleagues at Rutgers, Shohamy previously reported that while PD patients have trouble learning episodes, once they do, they have no trouble generalizing (see Shohamy et al., 2006).

Heckers is also not sure how this work might relate to AD, PD, or other cognitive deficit conditions. “For people who are primarily presenting with cognitive deficits, such as dementia, or cognitive deficits in Parkinson disease, I do not know how much this particular experiment explains it, because they show these nice relationships between behavior and brain activation only for the good learners who make generalizations. For the poor learners there is no role for the hippocampus or the VTA, so it does not really give us clues about what is not working in a patient who has cognitive deficit,” he said.

In fact, Shohamy is interested in studying how good versus poor generalizability may affect daily life. “I think the notion of generalizability is interesting. On one hand you can imagine that it is a powerful thing because you want to be able to create links across different experiences so that you can relate them. But you can also imagine that you might want to do that with a certain degree of caution. You would not want to overgeneralize everything. So there is a certain optimal degree of generalization,” she said.

As for psychiatric disorders where deficits are not as apparent, Heckers sees this study as being quite relevant. He said that in psychiatric disorders you do not have a broken memory, but a system that has lost fidelity and is not as accurate anymore. In this regard it is relevant that those subjects who did generalize well had no idea that they were doing it—they appeared to have a memory that they had seen a given face and scene as a pair when in fact they had not. “If that is not a cognitive neuroscience model for hallucinations, then I don’t know what is,” said Heckers.—Tom Fagan.

References:
Shohamy D, Wagner AD. Integrating memories in the human brain: hippocampal-midbrain encoding of overlapping events. Neuron 2008 October 23; 60:378-389.

Kumaran D, Duzel E. The hippocampus and dopaminergic midbrain: old couple, new insights. Neuron 2008 October 23; 60: 197-200.