Comments by "Kevin Street" (@Kevin_Street) on "Computerphile" channel.

  1. Fascinating! I think the central problem is with "e" in the diagram. You say it's "...a numerical fingerprint for the meaning in these two items..." But can "meaning" be reduced to a numerical fingerprint? It doesn't seem likely. The result of their paper sounds just like what you'd get when you're trying to achieve greater and greater performance from a system that has an ultimate limit. In this case the dataset is limited to the meaning humans put into it. Every picture can be equated to a series of words, by putting together sequences of words you can describe pictures, and so on. The fact that the dataset contains actual pictures and words that correspond with each other, and not just random noise is meaning that humans injected into the data. By representing the data numerically the computer can try out all the possible combinations until it finds the ones humans say are correct. It's like a blind watchmaker randomly putting together parts until someone tells him he built a working watch. The more parts there are, the longer it takes him to randomly build the watch. That's like what's happening in the paper. More and more data means it takes longer to find the inherent meaning, but the computer can't put any meaning of its own into the data (it can't build a "watch" that isn't already implicitly "there" in the parts), so eventually it approaches an ultimate limit.
    2
  2. Very interesting video! Thank you for this.
    1