Beyond Trial Averaging in Neural Population Dynamics

37 concepts33 papers1 proposal

For decades, a foundational assumption in systems neuroscience has been that the brain's true response to a stimulus is deterministic, and that any deviation from this response across repetitions is merely random noise. To eliminate this noise, researchers traditionally collect neural recordings over dozens of identical trials and compute an average response, such as a Peristimulus Time Histogram (PSTH) or an Event-Related Potential (ERP).

While trial-averaging succeeds at smoothing out messy data, it relies on a dangerous mathematical convenience: it assumes the underlying cognitive state is static across trials. In reality, the brain must process information and execute actions on a single-trial basis. By collapsing data into a pooled average, traditional methods systematically destroy trial-to-trial variability, mask internal cognitive dynamics, and can even invent fictional neural states that never actually occurred.

How Averaging Distorts Neural Reality

To understand why trial-averaging can be actively misleading, consider two distinct scenarios where taking the mean of a population alters the apparent mechanism of the brain.

The Illusion of the Hybrid State

Imagine an animal tasked with planning a movement when presented with two potential targets. Does the brain prepare both movements simultaneously, or does it commit to one and switch if necessary?

If the neural population alternates between committing to Target A on some trials and Target B on others, averaging these trials together yields a smooth, intermediate profile. This average suggests a hybrid strategy—preparing both targets at once. However, by analyzing population activity on a single-trial basis, Dekleva et al. demonstrated that the dorsal premotor cortex actually commits to a single movement plan immediately. The apparent hybrid state is an artifact of the average; it never occurs in any individual instance.

Ramping versus Stepping

A similar distortion occurs during decision-making. When an animal accumulates evidence over time, traditional trial-averaged firing rates show a gradual, smooth "ramping" of activity. This led to the long-standing theory that individual neurons continuously drift upward as evidence builds.

However, Zoltowski et al. applied single-trial Bayesian modeling to spike trains and revealed a different reality: many neurons do not ramp smoothly. Instead, they undergo discrete, high-variability "jumps" or steps in activity. When multiple trials with staggered step-times are averaged together, the resulting curve looks like a smooth ramp. The smooth population-level trend incorrectly implies a continuous underlying process, completely obscuring the discrete nature of the single-trial dynamics.

Individual Trials (Discrete Steps)
Trial 1: ____|¯¯¯¯¯¯¯¯¯¯
Trial 2: ______|¯¯¯¯¯¯¯¯
Trial 3: ________|¯¯¯¯¯¯

Trial Average (Illusory Smooth Ramp)
Average: ___/¯¯¯\_______

The Critical Information Lost in the Mean

Beyond creating artifacts, trial-averaging discards the exact fluctuations that explain behavior, individual differences, and temporal precision.

Behavioral Fluctuations and Continuous States

Why does a subject successfully resolve a cognitive conflict on one trial but fail on the next? Averaging erases the answer. Population dynamics on individual trials are highly predictive of trial-to-trial behavioral variability, such as reaction times and decision errors, as shown by Petreska et al. and Wei et al..

Furthermore, internal states like attention are not static across an hour-long experiment. Cohen identified that cognitive states fluctuate continuously from moment to moment, a reality that traditional trial-averaging cannot track. In clinical contexts, Prillinger et al. argued that averaging across trials and participants obscures individual variability and neural noise, making it difficult to isolate the specific factors—such as comorbid impulsivity—that drive varied responses to therapeutic interventions.

Temporal Precision and Coordination

Averaging also flattens the temporal structure of neural communication. Tang et al. measured the timing variability of neuronal populations and discovered that temporal "jitter" increases as information moves up the neural hierarchy—a phenomenon completely hidden when signals are aligned and averaged.

Crucially, trial-to-trial variations are often coordinated across regions. Seidel et al. and Delis et al. demonstrated that simultaneous power increases and correlations between muscle synergies are not random noise; they carry vital task-discriminating information that ensures reliable performance. When Karunasekara et al. analyzed population spike trains without averaging, they found that precise single-trial first-spike latencies encode fine-grained spatial information that simple averaged firing rates completely miss.

Recovering Single-Trial Dynamics

To resolve these limitations, modern computational neuroscience treats trial-to-trial variability as a feature rather than noise. By leveraging high-dimensional recordings, researchers use latent variable models to extract the "hidden" state of the brain in real time.

Hidden Markov Models and Switching Systems

Early work by Gat et al. pioneered the use of multivariate Poisson Hidden Markov Models (HMMs) to segment collective cortical firing patterns into discrete behavioral states on a single-trial basis. This concept was later expanded by Petreska et al. into Hidden Switching Linear Dynamical Systems (HSLDS). Instead of averaging across time and trials, these models segment individual trials into distinct dynamical stages, allowing researchers to map out how the brain transitions between cognitive states in real time.

Deep Learning and Latent Factor Analysis

With the advent of deep learning, frameworks like Latent Factor Analysis via Dynamical Systems (LFADS), introduced by Pandarinath et al., use sequential autoencoders to uncover the underlying logic of neural populations. By training on single-trial data, these networks learn to de-noise population activity while preserving the unique, instance-by-instance fluctuations that correlate with behavior.

Even when data is missing or constrained by recording technology, modern architectures can recover single-trial trajectories. For example, the transformer-based framework developed by Xia et al. "in-paints" unrecorded brain dynamics across multiple animals while explicitly preserving the trial-by-trial variability essential for understanding behavior.

By shifting the analytical paradigm from trial-averaging to single-trial modeling, contemporary neuroscience moves closer to how the brain itself operates: processing information dynamically, adapting to real-time fluctuations, and generating behavior from one moment to the next.

Key Foundational Work

Statistical Modeling of Cell Assemblies Activities in Associative Cortex of Behaving Monkeys (1992) — contributed 1 concept elements to this direction.
Analyzing and Visualizing Single-Trial Event-Related Potentials (1998) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_trial_averaging_bias_693b8f6d,
  title = {trial-averaging bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/trial-averaging-bias/rdr_693b8f6dffba4557974c19fe12702c0f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_trial_averaging_bias_693b8f6d_v1,
  title = {trial-averaging bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/trial-averaging-bias/rdr_693b8f6dffba4557974c19fe12702c0f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake