Poisson reduced-rank regression (p-RRR) builds on a standard linear regression in three ways.

  1. Counts rather than continuous values. Spike counts are non-negative integers, so the noise model is Poisson rather than Gaussian.
  2. A non-negative rate. The linear term is passed through softplus, which is smooth and never negative.
  3. Neurons fitted jointly rather than independently. Neural populations often have low-dimensional shared structure, which is modeled with a rank constraint between predictors and targets.

Rank constraint

When the middle dimension equals the smaller of the predictor and target counts, the model is at full rank: a set of independent regressions, one per target.

A lower rank requires the targets' weights to be built from the same r patterns.

Rank selection demo Explore p-RRR on ground-truth simulated data

Rank selection in the data

Fitted models on the recorded data

Every neuron and every model is browsable in the V1 neuron explorer.