Robust spatial memory maps encoded by networks with transient connections

Andrey Babichev, Dmitriy Morozov, Yuri DabaghianPLOS Computational Biology 2018 · OpenReview

TL;DR

This paper uses algebraic topology to model how the brain can form stable spatial memories despite the constantly changing connections between neurons. They demonstrate that even with transient neural networks, a robust topological representation of the environment can emerge, suggesting this is a fundamental property of neural information processing.

What Problem Does This Paper Address?

The brain's ability to form and retain memories, especially spatial memories, is remarkable. However, the underlying neural networks are highly dynamic. Synaptic connections between neurons are constantly being formed, strengthened, weakened, and eliminated through processes like synaptic plasticity. This means the physical structure of neural networks is in perpetual flux. The central question this paper addresses is: how can a stable, reliable representation of space (a cognitive map) be encoded and maintained by a network whose very architecture is transient and constantly changing? This is a fundamental challenge in understanding how the brain achieves robust memory formation.

Methods

The authors employ a computational framework that models the hippocampal network using concepts from algebraic topology, specifically Zigzag persistent homology.

  • Simplicial Complexes: They represent groups of coactive place cells as "simplexes" (geometric objects like points, lines, triangles, tetrahedrons). The collection of these simplexes forms a "simplicial complex" that represents the spatial connectivity of place fields.
  • Transient Connections: The model incorporates the transient nature of neural connections by allowing simplexes (and thus connections) to "flicker"—appear and disappear over time. This is modeled using a probabilistic decay function for connections.
  • Zigzag Persistent Homology: This is a novel mathematical tool that allows tracking topological features (like holes and connected components) in a sequence of evolving complexes where the connections can appear and disappear. Unlike standard persistent homology, which assumes a monotonic growth of features, zigzag homology can handle these dynamic changes, making it suitable for modeling transient networks.
  • Cognitive Graph Model: The implementation is based on a "cognitive graph" model, where place cells are vertices and connections are links. Coactive cell assemblies are represented as cliques in this graph, which directly translate to simplexes in the simplicial complex.

Datasets

Not discussed in the paper. The study is purely computational, using simulated environments and place cell activity.

Key Insights and Ideas

  • Topology as a Stable Representation: The core insight is that the topological structure of the environment can be robustly captured by the neural network, even if the specific connections forming that structure are transient. This is akin to a subway map (topology) being more stable than the exact road network (specific connections).
  • Generic Phenomenon: The emergence of stable topological maps from transient networks is presented as a "generic phenomenon," meaning it's not dependent on highly specific parameters but rather a fundamental consequence of how information is processed over time.
  • "Fluttering" vs. "Survivor" Simplexes: The model reveals that the coactivity complex consists of two populations of simplexes: rapidly recycling ("fluttering") ones and a stable "core" of long-lasting ("survivor") ones. This core is crucial for maintaining the overall topological structure.
  • Compensatory Mechanisms: Increased neuronal activity (higher firing rates or more active cells) can compensate for rapid connection decay, helping to stabilize the topological map. This suggests a potential therapeutic target for conditions affecting memory.
  • Multiple Timescales: The study highlights three principal timescales of spatial information processing: working memory (assembly formation/dissolution), intermediate memory (topological feature unfolding), and long-term memory (averaging out fluctuations for stable representation).

Key Results

  • Stable Topology Despite Flickering: The study demonstrates that despite a significant turnover of individual connections (up to 60% for links and 40% for triples over a certain period), the large-scale topological characteristics of the simulated cognitive map remain stable and accurately reflect the environment's shape.
  • Decay Time Matters: A longer decay time (τ) for connections leads to more stable topological representations, with stabilization achieved around τ ~ 2 minutes for a typical rat's movement speed.
  • Fixed Lifetimes Lead to Instability: In contrast to a probabilistic decay, using fixed lifetimes for connections resulted in significant topological instabilities, highlighting the importance of a distribution of lifetimes.
  • Random Flickering is Similar to Decaying: A network with randomly appearing and disappearing links (stochastic complex) showed similar topological stability to the decaying network, as long as the net probability of link existence was the same.
  • Activity Boost Stabilizes Maps: Increasing the mean firing rate or the number of active place cells significantly reduced topological fluctuations and increased the success rate of capturing the correct environmental topology.

Limitations and Future Work

  • Model Simplifications: The model uses simplified assumptions, such as a unified decay time for all links and a specific environment shape.
  • Physiological Realism: While inspired by biological findings, the model is a schematic representation and doesn't capture all the complexities of the hippocampal network.
  • Future Directions: The authors suggest that this framework can be extended to explore how different types of neurodegenerative diseases affect spatial memory and how interventions like deep brain stimulation might work by influencing neuronal activity to stabilize cognitive maps. They also propose that the observed timescales of information processing might involve interactions between the hippocampus and cortical networks.

Summary updated

Abstract

Author summary The reliability of our memories is nothing short of remarkable. Synaptic connections between neurons appear and disappear at a rapid rate, and the resulting networks constantly change their architecture due to various forms of neural plasticity. How can the brain develop a reliable representation of the world, learn and retain memories despite, or perhaps due to, such complex dynamics? Below we address these questions by modeling mechanisms of spatial learning in the hippocampal network, using novel algebraic topology methods. We demonstrate that although the functional units of the hippocampal network—the place cell assemblies—are unstable structures that may appear and disappear, the spatial memory map produced by a sufficiently large population of such assemblies robustly captures the topological structure of the environment.
BibTeX
@article{babichev2018robust,
  author = {Andrey Babichev and Dmitriy Morozov and Yuri Dabaghian},
  title = {{Robust spatial memory maps encoded by networks with transient connections}},
  journal = {PLOS Computational Biology},
  year = {2018},
  url = {https://openreview.net/forum?id=LAjelF0UHR},
}
Figure