We start our discussion of linear models for regression by specifying what we mean by a linear model. We emphasize that the linearity is with respect to the model parameters (the weights) rather than the data. This means that instead of using the raw data in our models and searching for a linear combination that best matches the target output, we can instead look for a linear combination of features, or basis functions, computed on our data. This greatly expands the types of data we can capture while maintain linear dependence on the parameters. The important caveat is that we must specify these features or basis functions before we perform our analysis - they are fixed, not learned from the data.
(Sorry about the echo in the audio track.)
