2 ms·

I was viva'd in 1998 an yea! verily, I used Iris and Abalone (and 10 others) in my thesis. But why, young padawan? why? Coz they worked. 16 years of industria

by sgt101 12y ago

I was viva'd in 1998 an yea! verily, I used Iris and Abalone (and 10 others) in my thesis. But why, young padawan? why? Coz they worked. 16 years of industrial machine learning research later... The world is not like the UCI repository Getting the data into the form that the analysis algorithm can parse, asking the question... that is the work. Ok, also dealing with the answer - that too.

tensor 12y ago
Data munging may be time consuming, but it is hardly difficult. The real challenge in ml is still the algorithms and theory.
- sgt101 12y ago
  No, I disagree. Algorithms can be selected, it's not hard to learn which technique is suited to which situation. What theory are you talking about? What do we have - Kearn's contributions on COLT and the descendant's there of (Michael Kearns is a great fella btw, and COLT is super - but so 1990's), or Vapnik and Structural Risk Minimisation? It's not really physics is it?
  - tensor 12y ago
    All the data munging in the world will not make progress of any sort in improving the state of ML/AI. All advances come from better algorithms and understanding of associated mathematics. Applying techniques to new problems involves data munging, and that is an important and useful task, but it is not difficult in the same way that doing new algorithmic work is, nor does it advance the state of the art.
    - sgt101 12y ago
      I still think you are wrong. There is no harm in developing a new algorithm, or coming up with new stories as to how it works, but to say that "data munging" will not advance the state of the art is to exclude the art! It's like saying that unless someone comes up with new physics it's impossible to build better spaceships. And if Nasa were to build a better spaceship would that advance the state of the art in spaceship building?

2 ms·

I was viva'd in 1998 an yea! verily, I used Iris and Abalone (and 10 others) in my thesis. But why, young padawan? why? Coz they worked. 16 years of industria

by sgt101 12y ago

tensor 12y ago
Data munging may be time consuming, but it is hardly difficult. The real challenge in ml is still the algorithms and theory.
- sgt101 12y ago
  No, I disagree. Algorithms can be selected, it's not hard to learn which technique is suited to which situation. What theory are you talking about? What do we have - Kearn's contributions on COLT and the descendant's there of (Michael Kearns is a great fella btw, and COLT is super - but so 1990's), or Vapnik and Structural Risk Minimisation? It's not really physics is it?
  - tensor 12y ago
    All the data munging in the world will not make progress of any sort in improving the state of ML/AI. All advances come from better algorithms and understanding of associated mathematics. Applying techniques to new problems involves data munging, and that is an important and useful task, but it is not difficult in the same way that doing new algorithmic work is, nor does it advance the state of the art.
    - sgt101 12y ago
      I still think you are wrong. There is no harm in developing a new algorithm, or coming up with new stories as to how it works, but to say that "data munging" will not advance the state of the art is to exclude the art! It's like saying that unless someone comes up with new physics it's impossible to build better spaceships. And if Nasa were to build a better spaceship would that advance the state of the art in spaceship building?