Generative replay across hippocampal-neocortical circuits

Hippocampal Generative Replay Propagates to Sensory Neocortex to Build Hierarchical Models

When we sleep, our brains do not merely replay the literal footage of our day. They also "imagine" new connections between things we have learned. This study shows that the hippocampus sends these imagined sequences to the visual cortex. This helps the brain build a more complex, predictive understanding of the world.

Building a Mental Map of the Unseen

Flexible decision-making requires an organism to infer relationships between events. These events may never have been directly experienced together. If a certain sound precedes a light, and that light precedes a reward, an animal can conclude the sound predicts the reward. This ability to "connect the dots" is a hallmark of higher intelligence.

The core question addressed by Shearer and colleagues is whether this process of mental synthesis—termed "generative replay"—is confined to the hippocampus. Or does it propagate to the rest of the brain? The hippocampus is widely recognized as the engine of memory. However, it remains unclear if it communicates these newly synthesized ideas to the sensory neocortex (the outer layer of the brain responsible for processing stimuli like sight and sound).

The Mechanics of Offline Consolidation

The theoretical framework for this research rests on the "two-stage" model of memory consolidation. In this model, the hippocampus acts as a rapid encoder of new experiences. These are then redistributed to the neocortex for long-term storage. This redistribution is thought to occur during "offline" periods, such as rest or sleep. This process is facilitated by high-frequency oscillations called Sharp-Wave Ripples (SWRs; fast bursts of electrical activity in the hippocampus).

During these SWR events, the hippocampus undergoes "replay." Neural spiking sequences recapitulate past experiences on a temporally compressed scale. Crucially, this can manifest as generative replay. The hippocampus co-activates memories of discrete cues that were never paired in reality. This effectively simulates potential future scenarios. The researchers hypothesized that if the hippocampus is generating these new, inferred relationships, it might provide a "teaching signal" to the neocortex. To test this, they utilized a three-stage inference task in mice [Figure 1A]. The mice had to learn associations between auditory cues ($X$), visual cues ($Y$), and outcomes ($Z$). The goal was to see if the mice could infer the $X \rightarrow Z$ relationship despite only experiencing $X \rightarrow Y$ and $Y \rightarrow Z$.

Mapping the Flow of Information

To investigate how these signals travel, the authors first established the physical infrastructure. They used a trans-synaptic intersectional strategy (a genetic technique that labels only specific neurons connecting two regions). They confirmed a disynaptic anatomical pathway from the dorsal CA1 (dCA1; a specific part of the hippocampus), through the retrosplenial cortex (RSCg; a middle-layer brain region), and into the primary visual cortex (V1) [Figure 2A].

Once the pathway was mapped, the researchers used a combination of multi-unit electrophysiology (recording individual neuron spikes) and calcium imaging (tracking population activity via fluorescent signals). Their findings were twofold. First, during the awake inference task, V1 exhibited a "prospective code." When mice heard the auditory cue ($X$), V1 neurons began representing the expected visual cue ($Y$) [Figure 3H].

Second, the researchers looked at what happens during sleep. They found that during dCA1 ripples, the information content in the hippocampus relates to upcoming information in V1 [Figure 4G]. Specifically, they discovered a temporal asymmetry. The activity content in dCA1 best predicted the activity content in V1 approximately 80ms before the ripple peak [Figure 4I]. This suggests a directional relationship where information is represented in dCA1 before it appears in V1.

The researchers further demonstrated that this replay in V1 is not just a random echo. Using Temporal Delayed Linear Modelling (TDLM; a mathematical framework used to detect structured sequences), they showed that V1 engages in "forward sequenceness" during sleep [Figure 6G]. This means V1 neurons fire in the specific order of the learned task structure ($X \rightarrow Y \rightarrow Z$). Most strikingly, they found that V1 neurons also participate in generative replay. They co-activated cells representing the unobserved relationship ($X \rightarrow Z$) during ripples [Figure 7D].

From Echoes to Internal Models

These results redefine our understanding of memory consolidation. Instead of the neocortex being a passive warehouse, it appears to be an active participant. It seems to help build a hierarchical generative model. In this view, the hippocampus provides a signal to organize the content of memories across the cortical hierarchy.

By receiving these generative signals, the visual cortex (V1) gains the capacity to represent abstract, inferred relationships. This explains how a sensory region can participate in complex, model-based planning. As the task becomes familiar, the researchers noted that V1's sequential activity becomes increasingly decoupled from the hippocampus [Figure 6M]. This suggests the neocortex eventually develops its own internal model of the world.

Limits of the Circuit

While the evidence for this hippocampal-neocortical dialogue is compelling, the study has defined boundaries. The findings are specifically tied to the dCA1-RSCg-V1 pathway. The researchers note that this study may not capture all possible interactions across the cortical hierarchy. Furthermore, the TDLM method relies on pre-specified transition sets. This means the researchers could only definitively prove the existence of the specific unobserved relationships they were looking for. They could not discover every possible latent connection the brain might be simulating. Future work will need to determine if this signal operates similarly across other sensory modalities.

Figures from the paper