Skip to contents

Online recursive independent-component analysis

Source: R/incremental-decomp.R
onlineICA.Rd

The processor combines exponentially weighted whitening with a natural gradient update and symmetric decorrelation. Reported components have a deterministic order and sign, but physiological source identity is not inferred. Only complete update blocks are emitted; an incomplete block is retained with its timestamps and sequence identity until a later update.

Usage

onlineICA(
  n_components = NULL,
  n_features = NULL,
  learning_rate = 0.01,
  forgetting = 0.995,
  nonlinearity = c("tanh", "extended"),
  block_size = 1L,
  orthogonalize_every = 1L,
  seed = 1L
)

Arguments

n_components

Optional number of sources; defaults to all features.

n_features

Optional feature count to bind at construction.

learning_rate

Positive natural-gradient step size.

forgetting

Per-sample whitening forgetting factor in (0, 1].

nonlinearity

Exact tanh or extended score rule.

block_size

Exact positive update block size.

orthogonalize_every

Exact positive number of blocks between symmetric decorrelations.

seed

Exact integer used only for deterministic initialization.

Value

A mutable OnlineICA streaming processor.

Examples

set.seed(1)
sources <- cbind(sin(seq_len(200) / 5), sign(sin(seq_len(200) / 7)))
samples <- sources %*% t(matrix(c(1, 0.4, 0.3, 1), 2, 2))
ica <- onlineICA(n_components = 2L, block_size = 10L)
result <- update(ica, samples)
dim(result$output)
#> [1] 200   2