It isn't a great description at all.
In supervised learning you have a training data set in which you are able to give the learning algorithm a label or category or other form of result for each record in the set. The algorithm then maps the input values of each record in the training data to the output labels/categories/results (according to some type of distance measurement) and creates decision boundaries that are used to determine which category any future record ought to belong to.
In unsupervised learning there is no a priori classification or labelling of the training data and the learning algorithm partitions the output data according to optimality that is derived from values of the output data (ie. is data-driven).
Example one (supervised): we have measured the height and weight (input variables) of a class of kids and calculated their BMI value. The BMI value can, of course, be classified as obese, overweight, etc. We do this for the each record inm the training data (using overweight/obese etc as output variable... in reality they will be integer numbers not words) and run our learning algorithm. We should end up with decision boundaries (inside the model) that delineate the different caegories we used. Take the model and present it the same data from other kids (ie. test data) and it should be able to tell which category they are in. In essence it is learning the BMI equation and the boundaries on a BMI chart. How good it is depends on using the right distance measure, enough (but not too much) training data, and using an appropriate learning algorithm with suitable parameters.
Example two (unsupervised): Using the same training data as example one we would skip the labelling and use the BMI value as the output value. In training the model then attempts to discover structure in the input and output data. For our BMI data, the number of output classes the model finds will depend on the method (and parameters) used and the data itself (it is tempting to imagine that if there were no clinically obese kids the model might find only four classes but with unsupervised learning it is possible that only two are found or even ten... ). When presented with the data of the other kids (test data) the model will separate them into the categories it found according to the decision boundaries it learned.
Supervised learning has known outputs such that, for any given predicted output, a distance measure can be used to decide if the prediction is right or wrong (and often, how wrong). Unsupervised learning has no a priori knowledge of structure within the data.
A great book is Pattern Classification by Duda, Stork, and Hart.
