Prediction of the Future Can Block Memory Formation

Post by Deborah Joye

What's the science?

Memory is a multifaceted system that helps us recall past events (episodic memory). This system can also use those memories to make predictions about the future (i.e. statistical learning). Remembering every unique detail about a past event may not be necessary, since those precise details are not likely to occur again in the exact same combination. Instead, it may be more efficient for our memory to generalize across common features within past events that can help us make predictions about the future. But these two functions of memory require different processes – episodic memory requires recall of specific memories linked to a particular time and space, while prediction requires synthesis of generalized features across many memories. Although these are very different processes, both have been linked to the hippocampus. How does the hippocampus mediate between these two potentially conflicting processes? This week in PNAS, Sherman and Turk-Browne demonstrate that when our memory uses statistical learning to predict what comes next, it prevents the current experience from being encoded into memory.

How did they do it?

For each experiment, the authors trained participants on a stream of images of outdoor scenes depicting distinct categories, such as beaches, mountains, fields, etc., then later tested their memory. All images were of 3 different overall types – category A images which were always followed by a category B image, and category X images which were randomly inserted and not preceded or followed by any particular type of image. Participants were not informed that images from category A would always be followed by category B, instead they learned about this association as they completed the task. This experimental design allowed the authors to assess how well participants remembered images they saw depending on whether the images were predictive (category A vs. category X) or predictable (category B vs. category X).

In the first experiment, the authors exposed participants to the image stream, then later had them view images and asked them to indicate whether they had seen that image before. To better evaluate the role of the hippocampus in this process, the authors next exposed participants to the image stream, then later asked them to identify whether they had seen that image before, and if so, what time they remembered seeing it. This additional timestamp question is a more detailed form of episodic memory recall that critically depends on the hippocampus. Finally, to see how brain activity changes during the memory task, the authors repeated their image stream test experiment while participants also underwent a functional magnetic resonance imaging (fMRI) scan.

What did they find?

The authors first found that participants were worse at remembering images that predicted another image (category A) than they were at remembering control images (category X). Interestingly, participants remembered seeing predictable images (category B) about the same as control images, suggesting that the memory deficit was specific to images that allowed prediction of what came next. When the authors added the more difficult timestamp question, they found a similar result: participants were much worse at remembering when they had seen predictive images (category A) relative to control images. These findings suggest that when participants saw an image from category A, it may have triggered the hippocampus to retrieve a representation of the category B image that was to immediately follow and not the category A image itself. This prediction of the upcoming image may block the hippocampus from creating a new memory of the specific details of the current image the participant was seeing. Finally, when the authors had participants perform the original image stream task (without the timestamp question) while undergoing fMRI they found that hippocampal activity when viewing category A images looked as if it was viewing category B images (the predicted image). They also found that the better the hippocampus was at predicting category B images, the worse the participant was at recalling category A images (relative to category X images).

What's the impact?

These findings demonstrate that when the hippocampus makes predictions from statistical learning, memory formation of the present is impaired. If statistical learning and memory encoding oppose one another, why does the hippocampus do both? One proposed idea is that hippocampal control of these processes allows them to regulate another. In an excellent analogy from the authors, “using one’s right foot to operate both the brake and gas pedals in a car serves as an anatomical constraint that forces one to either accelerate or decelerate, but not both at the same time.” Similarly, the hippocampus might likewise mediate between using already-formed predictive models to predict what’s next (“we already know what’s gonna happen, let’s GO”), and formation of new memories (“let’s pay attention closer here so we can remember this new situation”).

Sherman & Turk-Browne, Statistical prediction of the future impairs episodic encoding of the present, PNAS (2020). Access the original scientific publication here.