Lone's notes

Recently Updated

Open Problems in the Theory of Deep Learning - MITCBMM

Dec 08, 20241 min read

  • Video: Panel Discussion: Open Problems in the Theory of Deep Learning - YouTube

  • Thoughts:

    • ML might be bad for science and the importance of understanding how ML works

      thought

      thougths while listening to Open Problems in the Theory of Deep Learning - MITCBMM

      I recently listened to this talk where one of the panel made a point of why ML can be bad for science

      the argument is that ML gives us a seemingly impossible solution to many problems without the need to understand such solutions

      and scientists across various fields are using ML in that way

      where they just collect a lot of data and let the magic of ML do it work

      yes they need some understanding to design the model and data and training algorithm

      but that’s mostly a ML job

      and it gives them, for the most part, a black box solution for their problem

      and they don’t really understand the solution, just need to know that it works

      so they don’t really learn any new thing on the actually subjected in question

      so in some sense it hinders the growth of science, and we will be understand less and less

      also getting the solution in that way does not feel satisfactory

      so yeah, symbolic representation that you were (and still are I suppose) doing actually helps advance science

      not related to my research though, just a philosophical idea that’s a good food for thought

      Link to original

    • what ML models learns and what's the real solution

      thought

      Thought while listening to Open Problems in the Theory of Deep Learning - MITCBMM

      In The Bitter Lesson where it says

      This is a big lesson. As a field, we still have not thoroughly learned it, as we are continuing to make the same kind of mistakes. To see this, and to effectively resist it, we have to understand the appeal of these mistakes. We have to learn the bitter lesson that building in how we think we think does not work in the long run. The bitter lesson is based on the historical observations that 1) AI researchers have often tried to build knowledge into their agents, 2) this always helps in the short term, and is personally satisfying to the researcher, but 3) in the long run it plateaus and even inhibits further progress, and 4) breakthrough progress eventually arrives by an opposing approach based on scaling computation by search and learning. The eventual success is tinged with bitterness, and often incompletely digested, because it is success over a favored, human-centric approach.

      Link to original
      So if using our inductive biases (the researcher tries to incorporate their knowledge into the agent) gives worse results than statistically model a large amount of data, does it mean that our inductive biases are wrong ? For example, we came up with theory of language and from that derive syntax trees to analyze text, but language modelling using statistical methods outperform that. Does that mean our theory of language is wrong ? Can we make such deduction ?question

      Consequently, does this mean that the solution produce by the statistical model is the true solution ? Or is it just an approximation ? What define a real solution and does one exist ? Probably not.question

      Does a good approximation lie in a path toward the true solution ? In other words, if an approximation is good (i.e. produce good empirical results), does it similar to the real solution ? Could it be that those approximations exists in a whole different space than the real solution, it just happens to give good results but will never reach the best possible results.question

      Link to original

Graph View

Backlinks