To understand how a neuron computes, we often ask: "What specific stimulus makes this cell fire?" The traditional answer is the Spike-Triggered Average (STA)—the average of all stimulus snippets that preceded a spike. While intuitive, the STA is limited. It treats the neuron as a simple linear filter, assuming it only cares about one specific pattern.
However, real neurons are far more sophisticated. They can be sensitive to multiple features simultaneously, or they might be "complex" cells that respond to a pattern regardless of its phase. To capture this complexity, researchers use Spike-Triggered Covariance (STC), a method that looks not at the average stimulus, but at the variance of the stimuli that trigger a response.
Beyond the Average: Finding the Subspace
If the STA is a single "template" the neuron looks for, STC identifies the entire "subspace" or "feature space" the neuron operates within. As Blaise Aguera y Arcas, Adrienne Fairhall, and William Bialek demonstrated in What Can a Single Neuron Compute?, even the classic Hodgkin-Huxley model of a neuron is better described as a nonlinear filter operating on a low-dimensional (often 2D) manifold rather than a simple sum-and-threshold gate.
STC works by comparing the covariance of the "spike-triggering ensemble" to the covariance of the raw, unfiltered stimulus (usually white noise). By performing an eigenvalue decomposition on the difference between these two matrices, researchers can identify specific "eigenfunctions"—the axes in stimulus space that the neuron is most sensitive to.
Excitatory and Suppressive Features
The power of STC lies in its ability to distinguish between different types of influence:
- Significant Large Eigenvalues: These represent excitatory features. If a stimulus has a high projection onto these axes, the neuron is more likely to fire.
- Significant Small Eigenvalues: These represent suppressive features. These are directions in stimulus space where the variance of the spike-triggering ensemble is lower than the raw noise.
Odelia Schwartz, E. J. Chichilnisky, and Eero P. Simoncelli used this approach in Characterizing Neural Gain Control using Spike-triggered Covariance to show that these suppressive signals are the primary drivers of neural gain control. By identifying the stimulus components that inhibit firing, they moved beyond linear models to explain how neurons regulate their own sensitivity based on the context of the input.
Bridging Biophysics and Computation
One of the most difficult challenges in neuroscience is connecting the "wetware" (ion channels and voltages) to the "software" (the features being detected). Keisuke Ota and Toru Aonishi have developed a mathematical bridge between these worlds. In Higher-Order Spike Triggered Analysis of Neural Oscillators, they showed that the STC eigenfunctions can be derived directly from a neuron's Phase Response Curve (PRC)—the measure of how a single pulse of current shifts the timing of a neuron's next spike. This allows researchers to predict a neuron's complex feature space using significantly less data than traditional experimental methods.
Adaptation and Information Transmission
The neural feature space isn't static; it adapts to the environment. In the barrel cortex (which processes whisker input), Miguel Maravall and colleagues found that neurons rescale their tuning functions to match the variance of the incoming signal. As described in Shifts in Coding Properties and Maintenance of Information Transmission during Adaptation in Barrel Cortex, this gain control ensures that the rate of information transmission remains constant, whether the environment is "quiet" or "noisy."
Similarly, in the auditory system, Sean J. Slee used STC to show that auditory neurons act as high-precision differentiators, sensitive to the rate-of-change of a signal. This multidimensional sensitivity allows the brain to preserve sub-millisecond timing information, a feat that simple linear averaging would fail to explain.
Why It Works
STC works because it treats the neuron as a dimensionality reduction engine. The world is high-dimensional and noisy, but a neuron only cares about a few specific patterns. By looking at how the variance of the stimulus changes when a spike occurs, STC reveals the "filters" the neuron uses to discard irrelevant information and focus on the features—both excitatory and suppressive—that define its computational role.