TL;DR
This study reveals that the human hippocampus rapidly integrates newly learned connections between previously separate sequences, forming a unified representation that supports flexible predictions. This integration happens quickly after learning and is linked to improved behavioral performance, suggesting the hippocampus dynamically updates its internal models of the world.
What Problem Does This Paper Address?
Our ability to navigate the world and make informed decisions relies heavily on predicting future events. This predictive power is often built upon memories of temporally extended sequences of events. The hippocampus is known to be crucial for forming these memories and supporting predictions. However, a key challenge is that environments are not static; they change, and we need to update our internal models accordingly to make accurate predictions in novel situations. For example, if two familiar routes in a city become connected by a new street, we can plan journeys that combine parts of both routes. This paper investigates how and when the hippocampus updates its representations of these extended sequences when new connections are learned, and how this neural updating translates into flexible behavioral predictions.
Methods
The researchers employed a combination of behavioral tasks and functional Magnetic Resonance Imaging (fMRI). Participants first learned sequences of virtual environments. Then, they were taught new "bridge" connections that linked previously separate sequences into a single, continuous one. During fMRI, participants performed a prediction task where they had to anticipate future environments in both the newly connected sequences and control sequences that remained separate.
Key analytical techniques included:
- Multivoxel Pattern Analysis (MVPA): This was used to examine patterns of brain activity in the hippocampus and visual cortex. Specifically, researchers looked at the similarity of activity patterns between different maps (Map A and Map B) within the connected and unconnected sequence conditions. This helped determine if the hippocampus was integrating representations of the connected sequences.
- Representational Similarity Analysis (RSA): This was used to quantify the similarity between neural representations over time and across conditions.
- Behavioral Analysis: Statistical models (t-tests, GLMMs, LMMs) were used to analyze participants' accuracy and performance changes across different conditions and experimental runs.
Datasets
The study used data collected from 32 human participants. No external datasets were used for training or evaluation. The data was collected specifically for this experiment, involving participants learning sequences in virtual reality and then undergoing fMRI scanning.
Key Insights and Ideas
- Rapid Integration: The hippocampus doesn't wait to integrate new information; it starts doing so almost immediately after learning a new connection between sequences. This suggests a highly dynamic and responsive neural system.
- Unified Representations: When sequences are connected, the hippocampus represents them more similarly than when they remain separate. This indicates the formation of a more unified, integrated representation.
- Stability with Novelty: The updated hippocampal representations aren't entirely new; they retain features of the original sequences while also incorporating novel activity patterns. This suggests a balance between preserving existing knowledge and adapting to new information.
- Behavioral Relevance: The degree of hippocampal integration directly correlated with participants' ability to make better predictions in the newly connected sequences, highlighting the functional importance of this neural process.
- Hippocampus vs. Visual Cortex: While the hippocampus showed clear signs of integration, the visual cortex did not, underscoring the hippocampus's specialized role in sequence integration and spatial/temporal memory.
Key Results
- Behavioral Performance: Participants performed well above chance on the prediction task in both connected and unconnected conditions. Performance improved over time, particularly in the connected sequences, suggesting participants adapted to the new structure.
- Hippocampal Integration: Across-map pattern similarity in the hippocampus was significantly higher for connected sequences compared to unconnected ones, indicating integration. This effect was present even in the early runs after learning.
- Temporal Dynamics: Hippocampal integration emerged rapidly in early runs and was significantly higher in the connected condition compared to the unconnected condition. While pattern similarity in the connected sequences showed a trend to decrease in late runs, the overall integration effect persisted.
- Stability: Hippocampal representations showed stability, retaining features of their pre-updating state even after integration.
- No Blending: There was no evidence of "blending" where representations of one sequence inherited features from the other. Instead, the increased similarity was attributed to the incorporation of novel patterns.
- Behavioral Link: Greater hippocampal integration (higher across-map similarity) was associated with reduced behavioral costs (or even benefits) in the connected sequence condition, suggesting that integrated representations facilitate better prediction performance.
Limitations and Future Work
- Spatial Resolution: The study's analyses were across the hippocampus as a whole, lacking the spatial resolution to examine differences between hippocampal subfields or along the long axis. Future studies with high-resolution fMRI could investigate if integration is localized to specific hippocampal regions.
- Trial Number: The authors note that the unconnected sequence condition had fewer trials, which might explain some of the variance in pattern similarity.
- Distant Environments: It remains unclear whether hippocampal integration extends to temporally distant environments within a sequence, or if it's primarily localized to environments adjacent to the newly learned connections. Future research could address this by increasing trial numbers to gain sufficient statistical power.
- Mechanisms of Updating: The study provides evidence for a link-based strategy supporting multistep predictions. However, future work could disentangle the interplay between successor representations, model-based computations, replay, consolidation, and task demands in shaping hippocampal sequence representations.
Summary updated