There are many classification problems that are not linearly separable. Which trick often allows us to still separate the classes with a linear classifier? Adding a regularizer. Deriving new features from the existing ones. Using least-squares-loss instead of log-loss. Reducing the dimensionality with PCA