Neural Population Analysis Guide

Choosing a dimensionality reduction method for neural population data depends on your structure: use dPCA to demix stimulus and decision parameters, GPFA for single-trial trajectories, jPCA for rotational dynamics, and PCA for general variance extraction.

The best way to normalize neural population activity for dimensionality reduction is applying soft normalization or variance-stabilizing transforms, which ensure reliable dimensionality estimates and prevent highly active neurons from dominating the latent structure.

dPCA demixes neural variance by task parameters by separating simultaneous multi-neuron recordings into latent components representing distinct task variables like time, stimuli, and decisions, enabling condition-averaged analysis without requiring trial alignment.

Yes, you need cross-validation to determine neuron and trial counts for latent trajectory analysis. Applying cross-validation alongside data-quality criteria ensures sufficient sampling to robustly identify the dimensionality of neural population activity.

PCA extracts condition-averaged latent trajectories from multi-neuron recordings, but extracting single-trial latent trajectories requires GPFA, which applies dimensionality reduction across time bins to capture trial-specific neural population dynamics.

Use jPCA instead of GPFA when your goal is identifying rotational dynamics in neural population activity, as jPCA extracts latent trajectories constrained to rotational structures, whereas GPFA captures general single-trial temporal dynamics without structural constraints.