Random projections as regularizers: learning a linear discriminant from fewer observations than dimensions