Kernel trick
Suppose we are in the modelling framework with training data
. When using SVMs we want to find a hyperplane that linearly separates the data - though this might not be possible for the current embedding of in . Though it might be possible for a map The kernel trick is to define a kernel of similarity Whilst the form of may look complicated it usually is simpler than the embedding . Normally you do not find the function instead you define and check the Mercer’s condition which guarantees the existence of .