Latent representations in hippocampal network model co-evolve with behavioral exploration of task structure

Ian Cone, Claudia Clopath

This paper, published in Nature Communications, offers a biologically grounded model for how the brain constructs "cognitive maps"—internal representations of the world that allow us to navigate complex tasks. By combining single-cell biophysics with reinforcement learning, the authors demonstrate that our internal maps and our outward behaviors don't just happen at the same time; they actively shape one another in a "co-evolutionary" loop.

TL;DR

The researchers developed a hippocampal network model where single-cell "plateau potentials" and agent-level rewards drive the emergence of task-specific neural representations. They show that complex "splitter cells"—neurons that fire based on past context—only emerge when they are necessary to solve a task, evolving in a "zipper-like" fashion alongside the agent's behavioral proficiency.

What Problem Does This Paper Address?

In simple environments, a "place cell" that fires whenever you are at a specific coordinate is enough to navigate. However, real-world tasks are often non-Markovian: your next move depends on where you’ve been, not just where you are now. To solve these, the hippocampus develops "splitter cells" that represent the same physical location differently depending on the preceding context (e.g., "I am at the intersection, and I previously saw a left-turn cue").

The "chicken and egg" problem in neuroscience is this: to learn a complex task, an animal needs an appropriate internal map. But to build that map, the animal needs to know which parts of its history are relevant to the task. This paper investigates the biophysical mechanisms that allow a population of neurons to "discover" these latent task structures without a pre-existing template.

Methods

The authors propose a three-stage closed-loop model:

  1. The Environment: A simulated Y-maze where an agent receives a cue (A or B) and must choose a path (Left or Right) to get a reward.
  2. The HPC Network: A model of the hippocampus using three-compartment neurons (soma, basal dendrites, and apical dendrites).
  3. Basal dendrites receive sensory/location input.
  4. Apical dendrites receive recurrent feedback from other neurons.
  5. A three-factor learning rule governs the weights: it requires pre-synaptic activity, a post-synaptic "plateau potential" (a long-lasting voltage jump), and a reward signal.
  6. The Action Network: A simple reinforcement learning agent that translates hippocampal activity into movement decisions.

The "secret sauce" is the plateau-based learning. In the model, plateaus act as a filter, allowing the network to "stitch together" sequences of activity that lead to rewards.

Datasets

The study primarily uses simulated data generated within their custom Y-maze environment. To validate the model, the authors compared their results against experimental calcium imaging and electrophysiology data from Zhao et al. (2021), which documented the emergence of splitter cells in mice performing a similar cue-dependent navigation task.

Key Insights and Ideas

  • The "Zipper" Effect: One of the most striking findings is how representations emerge. Learning doesn't happen all at once. Instead, "split" representations (differentiating between Cue A and Cue B) appear first near the cue and then propagate forward toward the goal over hundreds of trials. This "zippering" creates a chain of memory that bridges the time gap between a cue and a reward.
  • Task-Relevance as a Pruning Tool: When the researchers ran a "Random Reward" task (where cues were irrelevant), the network failed to develop splitter cells, even when the same induction protocols were used. This suggests the hippocampus is efficient: it only maintains complex, "split" maps if they provide a statistical advantage for gaining rewards.
  • Co-dependence of Map and Policy: The model shows that if you reset the agent's behavior to random mid-way through, the internal map "collapses" back into simple place cells. The map requires the behavior to remain stable, and the behavior requires the map to be accurate.
  • Scaffolding for Few-Shot Learning: While the initial population takes hundreds of trials to learn the task structure, once that "scaffold" is built, new individual neurons can be integrated into the map in as few as 10 trials. This mirrors the "Behavioral Long-Term Potentiation" (BTSP) observed in real mice.

Key Results

  • Successful Replication: The model successfully generated "splitter" populations that matched the firing patterns and ramp voltages seen in biological experiments.
  • Behavioral Correlation: The "splitness" of the neural population (how well it distinguished between contexts) evolved in lockstep with the agent's "Correct Turn %."
  • Ablation Effects: Artificially "silencing" context-specific neurons led to specific behavioral deficits (e.g., the agent could solve Cue A trials but performed at chance on Cue B trials), proving the latent representations were functionally necessary for the task.

Limitations and Future Work

The primary limitation is the Plateau Induction Protocol. In this model, the researchers manually "induce" plateaus in specific neurons to mimic experimental conditions. In a truly autonomous biological system, the brain must decide for itself when to trigger these plateaus.

Future work could explore "meta-learning" rules where the network autonomously generates plateaus based on novelty or uncertainty. Additionally, the authors suggest exploring how the entorhinal cortex—the primary input to the hippocampus—might provide the initial "noisy" signals that kickstart this co-evolutionary process.

Summary updated

Abstract

To successfully learn real-life behavioral tasks, animals must pair actions or decisions to the task's complex structure, which can depend on abstract combinations of sensory stimuli and internal logic. The hippocampus is known to develop representations of this complex structure, forming a so-called "cognitive map". However, the precise biophysical mechanisms driving the emergence of task-relevant maps at the population level remain unclear. We propose a model in which plateau-based learning at the single cell level, combined with reinforcement learning in an agent, leads to latent representational structures codependently evolving with behavior in a task-specific manner. In agreement with recent experimental data, we show that the model successfully develops latent structures essential for task-solving (cue-dependent "splitters") while excluding irrelevant ones. Finally, our model makes testable predictions concerning the co-dependent interactions between split representations and split behavioral policy during their evolution.
BibTeX
@inproceedings{cone2024latent,
  author = {Ian Cone and Claudia Clopath},
  title = {{Latent representations in hippocampal network model co-evolve with behavioral exploration of task structure}},
  booktitle = {Nature Communications},
  year = {2024},
  doi = {10.1038/s41467-024-44871-6},
}
Figure