Latent factor analysis refers to statistical and machine learning methods that model observed neural population activity as arising from a smaller set of unobserved (latent) factors. These latent factors represent the coordinated patterns of neural activity — the "neural population dynamics" — that underlie behavior. By inferring these latent factors from noisy electrode recordings, researchers can extract cleaner, more informative signals for BCI decoding.

Concept

Consider recording from 96 neurons during a reaching movement. Each neuron has a noisy firing rate that reflects both the shared motor command (signal) and independent noise (thermal noise, synaptic noise, unrelated neural activity). Latent factor analysis separates the shared signal from the per-neuron noise, recovering the low-dimensional trajectory that all 96 neurons collectively trace through "neural state space" as the reach unfolds.

LFADS

The most influential latent factor analysis method in BCI neuroscience is LFADS (Latent Factor Analysis via Dynamical Systems), introduced by Pandarinath et al. in 2018. LFADS uses a variational autoencoder with recurrent neural network dynamics to:

  1. Infer smooth, denoised latent trajectories from single-trial neural population recordings
  2. Model the temporal dynamics of how latent states evolve over time
  3. Produce per-neuron firing rate estimates that are less noisy than raw spike counts

LFADS has been shown to dramatically improve the signal-to-noise ratio of neural population recordings, enabling better decoding of movements, reaching trajectories, and other behaviors from fewer trials.

BCI Applications

Latent factor models improve BCI systems by:

  • Denoising: Extracting cleaner control signals from noisy electrodes, particularly useful as electrode quality degrades over time
  • Robustness: If individual electrodes fail or drift, the latent factors can still be estimated from the remaining electrodes, as the information is distributed across the population
  • Calibration efficiency: Latent factor models may require fewer calibration trials because they pool information across electrodes
  • Cross-session stability: If the latent dynamics are stable even as individual neuron tuning changes (due to electrode drift), latent-space decoders may maintain performance without daily recalibration

Neural Population Geometry

Recent work by Card et al. (2024) and others has explored the geometric structure of neural population activity in latent space — how the "neural manifold" is shaped and organized. Understanding this geometry provides insights into how the brain represents and transforms information, with direct implications for designing more effective BCI decoders.