Visualization of Hyperparameter Effects in Gaussian Process Regression
Gaussian process regression (GPR), it's all the rage. However, many people may feel that it is difficult to understand what role the hyperparameters of GPR play. Therefore, I created a tool to visualize in real time how the prediction model by GPR changes when the hyperparameters are changed. GPy (https://github.com/SheffieldML/GPy) was used for GPR modeling and Bokeh (https://docs.bokeh.org/en/latest/) was used for interactive visualization. The code is available on github.
usage:
bokeh serve GPy_Bokeh_interactive.pyand access to http://localhost:5006/GPy_Bokeh_interactive
The RBF kernel is used for GPR modeling, and if the input is a single variable, the default hyperparameters are rbf.variance (σ2), rbf.lengthscale (l), Gaussian_noise.variance(σ2 noise). δ is Kronecker delta function.
The demo is here. When you change those values with the slider on the bokeh, the predicted value (mean) and uncertainty will be visualized according to the changed parameters.
It is also possible to train and model by entering arbitrary 1-d input and 1-d output data, so please try it. It may also be interesting to try modeling when the input is multivariate, using this code as a reference.