The quasi systematic nature of splitter cells

Technical Summary: The Quasi-Systematic Nature of Splitter Cells

Problem Statement
Recent investigations into the hippocampal formation have identified a diverse array of cells with distinct functional properties. While existing models, often supported by carefully crafted architectures, hypothesize that temporal sequences underlie observed spatial properties, a fundamental question remains: Is a specific, structured architecture or learning rule required to generate these phenomena? Specifically, the paper questions whether a random recurrent structure is sufficient to allow latent sequences to emerge and whether the resulting "splitter cells" (neurons that fire differently depending on the animal's future path at a choice point) are functionally necessary for task resolution.

Methodology
To address these questions, the authors simulated an agent equipped with egocentric sensory inputs tasked with navigating an environment and alternating choices at intersections. The core of the investigation utilized a random recurrent network, avoiding specific architectural constraints or specialized learning rules. The researchers employed a systematic lesioning protocol:

1. Identification: They first identified splitter cells within the trained model.
2. Lesioning: They systematically removed (lesioned) these identified cells.
3. Re-evaluation: They monitored behavioral performance and network reorganization following each lesion.
4. Decoding and Analysis: They decoded position, orientation, and decision representations from the reservoir activity. Furthermore, they utilized subspace alignment analysis to examine how the population geometry and trajectory-encoding dimensions changed across lesions.

Key Results
The study yielded several critical findings regarding the nature and necessity of splitter cells:

• Emergence in Random Networks: Splitter cells were successfully identified within the random recurrent network, demonstrating that such specific functional properties can arise generically in networks that successfully solve the navigation task.
• Robustness to Lesioning: In the vast majority of cases, the model's behavioral performance remained intact after the systematic lesioning of splitter cells. New splitter cells re-emerged through network reorganization. In the remaining cases, the task was solved without any detectable splitter cells, indicating that these specific cells are not necessary for task resolution.
• Preservation of Task-Relevant Geometry: Despite the reorganization, position, orientation, and decision representations could be successfully decoded from the reservoir activity. Subspace alignment analysis revealed that the network preserves the task-relevant population geometry while redistributing activity within the null subspace. Notably, the trajectory-encoding dimension was observed to rotate within neuron space across different lesioning events.
• Parameter Independence: The emergence of splitter cells was found to be robust across a broad range of dynamical parameters.

Significance and Claims
The paper challenges the prevailing notion of the functional necessity of specific neural populations, such as splitter cells, for task performance. The authors conclude that splitter cell activity is primarily task-driven rather than a product of a specific architecture or learning rule. Instead, these cells emerge generically across random recurrent networks that successfully solve the task. Consequently, the presence of splitter cells is not a prerequisite for the behavior; rather, they are a byproduct of the network's solution to the task, which can be achieved through various internal configurations and reorganizations. This suggests that the functional necessity of specific neural populations should be reconsidered in favor of understanding the broader, flexible population dynamics that support behavior.