right, a key point of this method is that they are doing transfer learning from the general gmail population to each individual user. the reason for this is tha

right, a key point of this method is that they are doing transfer learning from the general gmail population to each individual user. the reason for this is that the data from each user is small and sparse, and has relatively little statistical power. you simply cannot do this on a desktop system unless you pull in data from other users. you might be able to do that but: bandwidth, privacy concerns, and you might find you have to do just as much work to get the same performance.