1
$\begingroup$

I would like to showcase the value of machine learning compared to traditional rule based (symbolic AI) methods.

I would also like to break it down for Unsupervised, Supervised and Reinforcement learning. E.g.:

  • Why would you perform clustering using k-means, DBSCAN etc vs asking an expert of that field (assuming there is one) to give you some segmentation rules?
  • Why would you use a Random Forest or an XGBoost to do classification rather than again getting the rules from and predictions from an expert?
  • Why would you use Thompson's sampling to find the next action vs "downloading" the brain of an expert?

Ideally, I would like to move beyond the idea of "a lot of options/data/possibilities" for a human to take them all into account and process without adding their bias. Is there anything else that can answer this question without of course setting a competition between the expert and the machine?

$\endgroup$
3
  • $\begingroup$ Look at the scoreboard! If some method performs the best, that’s appealing. What beyond this would you want in an answer? $\endgroup$
    – Dave
    Commented Apr 16, 2022 at 0:25
  • 1
    $\begingroup$ Probably best to read writings of Gary Marcus. He has extensive examples. Even though his writings takes a critical thinking approach, but also provides examples whereby data-driven ML excels. $\endgroup$ Commented Apr 16, 2022 at 0:27
  • $\begingroup$ @Dave This is before looking at the scoreboard, before even setting up a match between the algorithm and the human expert. Imagine that you're trying to convince someone strongheaded to go ahead and set-up this experiment $\endgroup$ Commented Apr 16, 2022 at 11:09

1 Answer 1

1
$\begingroup$

Machine learning is fundamentally about pattern recognition. If you have sufficient data and the model is flexible enough, you can find the pattern (see universality theorems). An expert makes predictions based on their experiences (data) and some perceived relationship (pattern) in the data.

To me, ML and an expert are trying to do the same thing - with the crucial difference that human brains process the data differently and computers can crunch massive datasets whereas humans cannot. Clearly if the digitized data is a poor representation of reality or the model is inappropriate, the ML results will be worse. Alternatively humans are subject to cognitive biases and have limitations as well.

But why does it have to be exclusively one or the other? Why not embed the predictions of experts into an ML algorithm?

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.