Abstract: Continuous "bump" attractors are an established model of cortical working memory for continuous variables and can be implemented using various neuron and network models. Here, we develop a generalizable approach for the approximation of bump states of continuous attractor networks implemented in networks of both rate-based and spiking neurons. The method relies on a low-dimensional parametrization of the spatial shape of firing rates, allowing to apply efficient numerical optimization methods. Using our theory, we can establish a mapping between network structure and attractor properties that allows the prediction of the effects of network parameters on the steady state firing rate profile and the existence of bumps, and vice-versa, to fine-tune a network to produce bumps of a given shape.
The Challenge of Continuous Attractor Networks
In neuroscience, continuous attractor networks (CANs) are the standard model for how the brain maintains "working memory" for continuous variables, such as the direction of a visual stimulus or the location of an object. In these networks, activity forms a "bump" that can persist and move smoothly across a population of neurons arranged in a specific topology, often a ring.
However, analyzing these networks is notoriously difficult. Real neurons are "spiking" (sending discrete pulses) and have complex, nonlinear synaptic properties. Traditionally, to understand how changing a single biological parameter affects memory, researchers had to run thousands of computationally expensive simulations. There was no easy way to "solve" the network's state or to find the exact parameters needed to create a specific bump shape without tedious trial and error.
A Low-Dimensional Shortcut
The core innovation of this paper is a low-dimensional parametrization of the network's steady state. Instead of trying to calculate the firing rate of every individual neuron, the authors assume the "bump" of activity follows a specific mathematical form: a generalized Gaussian function. This function is defined by only four parameters: baseline activity, peak height, width, and "squareness" (steepness).
To find the steady state, the authors developed a method based on:
- Self-Consistency Equations: They derive equations where the network's output must match its input. If the assumed bump shape is a true steady state, the "error" in these equations should be zero.
- Strategic Sampling: Rather than checking the error for every neuron, they evaluate the equations at just a few key points along the bump.
- Numerical Optimization: They use standard optimization algorithms to quickly find the four parameters that minimize the error at these sampling points.
This turns a high-dimensional simulation problem into a low-dimensional curve-fitting problem, allowing researchers to "reverse-engineer" networks to produce specific memory activity patterns much faster than traditional simulations.
Key Results and Insights
- Significant Speedup: For spiking networks, the optimization method reached a solution in about 2 seconds, compared to roughly 30 seconds for traditional methods—a 15x speedup that enables massive parameter sweeps that were previously unfeasible.
- High Accuracy: The predicted bump shapes closely matched the results of full-scale spiking simulations (validated using the NEST simulator).
- Biological Realism: Unlike previous simplifications that only worked for simple "rate models," this method holds up even in complex "conductance-based" spiking models with nonlinear NMDA synapses.
- Mapping Disinhibition: The authors used the tool to show how "disinhibition" (reducing inhibitory neuron activity) causes memory bumps to widen and lose precision, effectively mapping out the boundaries of stable memory.
Limitations and Future Work
While powerful, the current method focuses on unimodal (single-bump) symmetric states. It may struggle with multi-modal or highly asymmetric activity patterns. Additionally, while the method finds theoretical steady states, it does not inherently prove their dynamical stability against noise. Future work could extend this approach to 2D "sheets" of neurons and incorporate more complex dynamical analysis.
Summary updated