Abstract
Information transmission is a key element for information processing in the brain. A number of mechanisms have been proposed for transferring volleys of spikes between layers of a feedforward neural circuit. Many of these mechanisms use synchronous activity to provide windows in time when spikes may be transferred more easily from layer to layer. Recently, we have demonstrated that a pulse-gating mechanism can transfer graded information between layers in a feedforward neuronal network. Our transfer mechanism resulted in a time-translationally invariant firing rate and synaptic current waveforms of arbitrary amplitude, thus providing exact, graded information transfer between layers. In this paper, we apply a Fokker-Planck approach to understand how this translational invariance manifests itself in a high-dimensional, non-linear feedforward integrate-and-fire network. We show that there is good correspondence in spiking probabilities between the Fokker-Planck solutions and our previous mean-field solutions. We identify an approximate line attractor in state space as the essential structure underlying the time-translational invariance of our solutions. This approach has enabled us to better understand the role played by the synaptic coupling, gating pulse and the membrane potential probability density function in information transfer.
Introduction
A fundamental question in neuroscience is how the brain transmits "graded" information—signals with continuous, varying amplitudes—across feedforward layers without significant degradation. While traditional models like synfire chains often focus on binary "all-or-nothing" transmission, biological systems must communicate stimulus intensity (e.g., the brightness of a light or the volume of a sound) through continuous signal levels.
Previously, a pulse-gating mechanism was proposed to achieve this graded transfer in simplified mean-field models. This study extends that work by applying a rigorous Fokker-Planck (FP) framework to a high-dimensional network of Integrate-and-Fire (I&F) neurons to determine if these dynamics remain stable in more complex, non-linear environments.
The Pulse-Gating Mechanism
The core of the proposed architecture is a feedforward network where each layer receives a timed "gating pulse." This pulse shifts the population's membrane potential closer to the firing threshold, creating a window for information transfer.
By using FP equations, the researchers track the evolution of the probability density function (PDF) of membrane potentials across the population rather than individual spikes. This approach reveals how the shape and amplitude of the signal remain "time-translationally invariant"—meaning the signal's form is preserved as it propagates from one layer to the next.
The Approximate Line Attractor
The most significant theoretical contribution of this work is the identification of an approximate line attractor in the system's state space. By analyzing the first moments (means) of the probability distributions, the authors found that network activity trajectories rapidly converge toward a diagonal line in state space.
Once the system reaches this line, it can support a continuum of stable signal amplitudes. This geometric structure prevents the signal from collapsing into a single "on" or "off" state, thereby enabling the transmission of graded information. The dynamics behave similarly to a saddle-node with an unstable manifold along the diagonal, creating a "slow drift" that is robust enough for biological processing over multiple layers.
Key Results and Implications
The Fokker-Planck model demonstrated remarkable stability and fidelity:
- High Fidelity Propagation: Graded signals successfully propagated across 12 layers with a maximum fractional change in amplitude of only 0.2% per transfer.
- Stereotypical Waveforms: Regardless of the initial input amplitude, firing rates and synaptic currents quickly converged to stereotypical shapes.
- Parameter Sensitivity: The study found that successful transfer requires precise tuning. For instance, the synaptic coupling strength ($S$) in the FP model needed to be roughly 7% higher than predicted by simpler mean-field theories to maintain stability.
- Role of Noise: The FP approach explicitly accounts for effective diffusivity (noise), showing that the specific distribution of membrane potentials is critical for the timing and success of information transfer.
Conclusion
By identifying the approximate line attractor, this research provides a structural explanation for how feedforward networks can maintain high-fidelity graded signals. While the attractor is "approximate" and subject to slow drift, its stability suggests a robust mechanism for neural communication. Future work may explore how biological processes like spike-timing-dependent plasticity (STDP) help the brain "tune" these parameters to maintain such stable transmission pathways.
Summary updated