14 ms·

The implicit assumption here is that 100% of the training data comes from regular Gmail users and it didn't start off with any seed data provided by Google empl

• IceMetalPunk 5y ago
  Not at all. It doesn't have to be 100% of the data, it just has to be enough of the data that it influences the filter's decisions toward one direction over another. And since GMail has been in existence for 18 years and has a massive user base, I would be extraordinarily shocked if Google's initial seed training set isn't totally overshadowed by user-supplied data by now.
• labrador 5y ago
  It seems fairly useless if we get different results depending on where we start from. That's not what I call learning.
  • IceMetalPunk 5y ago
    To be fair, there would be a difference for awhile, as more training data makes its way into the system. But you're right that an effective machine learning system, or even just an effective Bayesian analysis, would be problematic if it didn't trend towards a the same answer as more data is obtained. And after 18 years and millions of users, yeah, if the initial training data still held any influence over its decisions, it would mean the whole system is broken.