The standard textbook model of a neuron is often a "sum-and-threshold" device: it adds up incoming signals and fires if they cross a limit. In statistical neuroscience, this is frequently formalized as a Poisson process, where the probability of a spike depends only on the current stimulus. However, biological neurons are not memoryless. A neuron that just fired is physically incapable of firing again immediately (refractoriness), and its past activity can trigger long-term changes in its sensitivity (adaptation) or cause it to fire in rapid successions (bursting).
Modern research has shifted toward models that treat a neuron’s own firing history as a fundamental input, as critical to its output as the external stimulus itself.
The Failure of the Memoryless Assumption
Most early neural models assumed that spikes in separate time intervals are independent. As Maneesh Sahani and Jonathan W. Pillow argue in The Equivalence of Information-Theoretic and Likelihood-Based Methods for Neural Dimensionality Reduction, this assumption is a mathematical convenience that fails to capture biological reality. If we treat a neuron as a simple Poisson processor, we miss the "temporal autocorrelation" inherent in spike trains—the fact that a spike at time $t$ strongly influences the probability of a spike at $t+1$.
Blaise Aguera y Arcas and William Bialek demonstrated in What Can a Single Neuron Compute? that even the classic Hodgkin-Huxley model—the gold standard for biophysical realism—is best understood as a nonlinear filter operating on a low-dimensional subspace of inputs. Crucially, they found that accounting for history-dependent effects like bursting is essential for bridging the gap between simple models and complex biology.
The Spike-History Filter
To fix the memoryless model, researchers introduced the Generalized Linear Model (GLM) with a feedback loop. In this framework, a neuron’s firing rate is determined by two main components:
- A Stimulus Filter: What the neuron is "looking for" in the outside world.
- A Post-Spike Filter: A kernel that is "added" to the neuron's internal state every time it fires.
Alison I. Weber shows in Capturing the Dynamical Repertoire of Single Neurons with Generalized Linear Models that this simple addition allows GLMs to replicate complex behaviors like adaptation and bistability. By adding a negative post-spike filter, the model simulates a refractory period; by adding a filter that briefly increases the firing probability, it simulates bursting.
This isn't just a theoretical improvement; it's a statistical one. Jonathan W. Pillow and colleagues proved in Maximum Likelihood Estimation of a Stochastic Integrate-and-Fire Neural Encoding Model that these models have a "log-concave" likelihood. This means there is a single "best" set of parameters that researchers can find reliably using standard optimization, allowing them to fit biophysically realistic models to real-world data with high precision.
Beyond the Last Spike: Non-Renewal Processes
A common simplification in neuroscience is the renewal process, which assumes that when a neuron fires, its internal state "resets" completely. In this view, the timing of the next spike depends only on the time elapsed since the most recent spike.
However, David Liu and Máté Lengyel challenge this in Bayesian nonparametric (non-)renewal processes for analyzing neural spike train variability. They argue that neurons have a longer memory. The probability of firing might depend on the last five spikes, or the average rate over the last second. Their NPNR framework uses Gaussian Processes to capture these complex, multi-spike dependencies, revealing that what looks like "noise" in neural data is often a structured signal driven by the neuron's long-term history.
Similarly, Eilif B. Muller developed Non-renewal Markov models to account for correlations between consecutive spike intervals, providing a more accurate way to predict how ensembles of neurons adapt to stimuli over time.
Why History Matters for Higher-Level Theory
Ignoring spike history doesn't just make models less accurate; it can lead to incorrect conclusions about how the brain works. A prominent example is the debate over how neurons in the lateral intraparietal (LIP) cortex represent decisions. Do they "ramp" up their activity smoothly as evidence accumulates, or do they "step" up in discrete jumps?
Xuelong Zhao and David M. Zoltowski found that once you account for a neuron's self-history and trial-to-trial fluctuations, the evidence for smooth ramping becomes much stronger. In Discrete stepping and nonlinear ramping dynamics underlie spiking responses of LIP neurons during decision-making, they show that failing to model refractory periods and bursting can cause a smooth process to look like a series of discrete steps, potentially misleading researchers about the fundamental nature of neural computation.
Scaling Up to Populations
The challenge of history dependency extends from single neurons to entire populations. Tilo Schwalger and Moritz Deger have worked to bridge this gap in Bridging spiking neuron models and mesoscopic population models. They derived "mesoscopic" equations that describe the collective behavior of thousands of neurons while still respecting the microscopic history-dependent properties—like refractoriness—of the individual cells.
By incorporating these "quasi-renewal" approximations, researchers can now simulate large-scale brain activity that remains grounded in the temporal realities of biological signaling. Whether in generative models like Latent Diffusion for Neural Spiking Data by Jaivardhan Kapoor or in the training of Spiking Neural Networks by Peng Li, the message is clear: to understand where a neuron is going, you must precisely model where it has been.