Linking Neural Manifolds to Circuit Structure in Recurrent Networks

Louis Pezon, Valentin Schmutz, Wulfram GerstnerbioRxiv 2024 · DOI

TL;DR

This paper bridges the gap between two views of neural computation: the "circuit structure" perspective focusing on individual neuron properties and connections, and the "neural manifold" perspective describing population activity as low-dimensional dynamics. They introduce a unifying framework showing that different circuit structures can produce the same neural dynamics, but crucially, they also develop a method to extract circuit structure from neural activity, opening doors for understanding brain computation from experimental data.

What Problem Does This Paper Address?

Neuroscience currently grapples with two main ways of understanding how the brain works. One is the "circuit structure" view, which emphasizes the specific connections between neurons and their individual tuning properties (e.g., a neuron firing maximally for a specific stimulus orientation). The other is the "neural manifold" view, which, fueled by advances in recording many neurons simultaneously, observes that the collective activity of large neural populations often lies on a low-dimensional "manifold" within the high-dimensional space of all possible neural states. This manifold represents the computations the network is performing.

The challenge is to connect these two perspectives: how does the underlying circuit structure give rise to these observed low-dimensional dynamics, and can we infer the circuit structure from the dynamics? Existing work has struggled to establish a clear mathematical link, and it's unclear if the neural manifold uniquely determines the circuit structure. This paper aims to bridge this gap by developing a theoretical framework and a practical method for this connection.

Methods

The authors propose a unifying mathematical framework based on field models of low-rank recurrent neural networks (RNNs). They model networks of spiking neurons where the connectivity matrix is constrained to have a low rank (D). This low-rank structure is key, as it implies that the population activity will lie on a D-dimensional neural manifold.

Their framework connects the "circuit structure" (represented by neuronal locations in an abstract "similarity space" M, and functions f and g that map these locations to connectivity vectors) to the "neural manifold" (the low-dimensional dynamics of the population activity). They derive equations showing how the collective dynamics of the network (represented by latent variables κ) emerge from the underlying circuit structure.

A crucial part of their work is developing a method for extracting the similarity space from neuronal activity. This method is "task-agnostic," meaning it doesn't rely on knowing the specific task the network is performing or the stimuli it's receiving. Instead, it analyzes statistical properties of individual neurons' activity (like their tuning to principal components of the population activity) to infer their relative positions in the similarity space, effectively reconstructing the circuit's underlying organization.

Datasets

The paper primarily uses simulated data from their own recurrent neural network models. They do not introduce new benchmark datasets but rather test their methods on simulated networks designed to mimic different circuit structures and collective dynamics.

Key Insights and Ideas

  • Mismatch is the norm: The paper's most striking insight is that the dimension of the "similarity space" (representing circuit structure) and the dimension of the "neural manifold" (representing dynamics) do not necessarily match. Different circuit structures can lead to identical low-dimensional dynamics. This challenges the idea that observing the dynamics directly reveals the circuit.
  • Symmetries are key: While circuit structure and dynamics aren't uniquely linked, symmetries in the circuit structure (e.g., rotational symmetry in a ring model) do translate into symmetries in the collective dynamics. This provides a crucial, albeit indirect, link.
  • Task-agnostic circuit extraction: The development of a method to infer circuit structure from neural activity without relying on task-specific information is a major contribution. This allows for the discovery of circuit features that might be irrelevant to a specific task but still important for understanding brain organization.
  • Unifying framework: The paper provides a powerful theoretical lens that unifies existing models of neural fields and low-rank RNNs, showing how they can be viewed as instances of a more general circuit model.

Key Results

  • The authors demonstrate through simulations that distinct circuit models (e.g., a 1D ring model vs. a 2D planar model) can produce identical limit-cycle dynamics in a 2D neural manifold.
  • They successfully apply their task-agnostic method to simulated data, showing it can distinguish between different circuit structures (e.g., a ring-like similarity space vs. a 2D surface-like space) even when the networks exhibit similar population-level dynamics and solve the same task.
  • They show that their method can even extract "task-irrelevant" circuit structure components, such as additional features assigned to neurons that don't directly influence task performance but are reflected in the inferred similarity space.

Limitations and Future Work

The authors acknowledge several limitations:

  • Homogeneous structural principle: Their model assumes a homogeneous structural principle for connectivity and neuronal properties, whereas biological neural networks exhibit more trial-to-trial variability and potentially chaotic dynamics.
  • Neuron model: They use linear-nonlinear-Poisson (LNP) neurons, which don't account for neuronal refractoriness. While they suggest their theory can be extended to renewal neurons, it requires more complex formalism.
  • Balanced networks: Their framework doesn't explicitly incorporate the balance of excitation and inhibition observed in cortical networks, and its extension to spatially structured balanced networks is an open problem.
  • Experimental validation: The proposed method for extracting circuit structure has only been tested on simulated data. Future work should involve applying it to real experimental recordings.

They suggest future work should focus on extending the theory to more realistic neuron models and network structures, and crucially, on applying their methods to experimental data to uncover the intricate features of cortical circuits.

Summary updated

Abstract

Abstract The classic view of cortical circuits composed of precisely tuned neurons hardly accounts for large-scale recordings indicating that neuronal populations are heterogeneous and exhibit activity patterns evolving on low-dimensional manifolds. Using a modelling approach, we connect these two contrasting views. Our recurrent spiking network models explicitly link the circuit structure with the low-dimensional dynamics of the population activity. Importantly, we show that different circuit models can lead to equivalent low-dimensional dynamics. Nevertheless, we design a method for retrieving the circuit structure from large-scale recordings and test it on simulated data. Our approach not only unifies cortical circuit models with established models of collective neuronal dynamics, but also paves the way for identifying elements of circuit structure from large-scale experimental recordings.
BibTeX
@misc{pezon2024linking,
  author = {Louis Pezon and Valentin Schmutz and Wulfram Gerstner},
  title = {{Linking Neural Manifolds to Circuit Structure in Recurrent Networks}},
  year = {2024},
  doi = {10.1101/2024.02.28.582565},
  howpublished = {bioRxiv},
}
Figure