The demand for adaptive learning methods, e.g. for use in brain computer interfaces (BCIs)
[15] has recently triggered a considerable interest in such algorithms. We may approach
adaptive learning with algorithms that were designed for stationary environments and use
learning rates to make these methods adaptive. These approaches can be traced back to
early work on learning algorithms (e.g. [1]). A more recent account to this approach is
[17], who combines the probabilistic method of sequential variational inference ([9]) and a
forgetting factor to obtain an adaptive learning method. Probabilistic or Bayesian methods
allow also for a completely different interpretation of adaptive learning. We may regard the
model coefficients as latent (i.e. unobserved) states of a first order Markov process.