The researchers used soft-labelled examples to train a kNN algorithm to encode increasingly complex boundary lines, splitting the chart into far more classes than data points.