Continual Gaussian Processes

Continual learning of sparse Gaussian process (GP) models based on inducing inputs.

The data consists of a signal of 1K input-output observations. Each sample is the monthly average of sunspot counting numbers. (Royal Greenwich Observatory). We use transform the data via the non-linear mapping log(1+x) to use a Gaussian likelihood distribution.

At each time-step, the model recomputes the entire predictive probability over the input domain seen so far. It is based on the variational inference method. However, data samples are only processed once and never revisited again, that is, only one-sample is used at each optimization run.

machine learning,statistics,gaussian process,online learning,continual learning,bayesian inference,