A kernel function computes the inner product of two data points in some (potentially high-dimensional) feature space without requiring an explicit mapping.
By carefully choosing the kernel, one can capture complex relationships or nonlinear separations in the data.