During week 1 - Supervised vs Unsupervised Machine Learning a question was asked in the video “What is machine learning” which goes as follows:
If Arthur Samuel’s checkers-playing program had been allowed to play only 10 games (instead of tens of thousands games) against itself, how would this have affected its performance?
The answer is ‘Would have made it worse’
This is the obvious answer at first however, what is performance relative to in this question? If the computer played more than ZERO games I would say that the performance has still become ten games better than not playing at all.
In any case, I believe that the answer should be as follows:
“The checkers playing program would perform worse than if it played itself 10,000 times.” and not “would have made it worse”
I am in agreeance with you for the simplicity but I think the answer should be rephrased to “the checkers-playing program would perform worse” since we are comparing to a different amount of times ‘played’. The word “made” sounds like the person asking the question is making something that had already learned something, lose that knowledge and become less knowledgeable. If we are talking about learning, I would logically think that we are always starting from point zero, zero being no knowledge/data to begin with therefore any knowledge learned creates better performance than no knowledge at all.
Thank you for sharing your view. I will let the course team know your opinion so that they can make their decision about the wording.
In machine learning, we don’t always have to start from ground zero (there is a topic called Transfer learning but not covered by this course and not relevant to this question), however, in the context of this question, I think we are starting from ground zero, so I agree with this statement.
I have thought about whether I would recommend any reading, and here is my comment. This Supervised Vs. UnSupervised ML section meant to give a very brief overview of them but not to dig deep.
For the sake of going through these courses, I think we can just watch whatever is presented to us and try to make some notes ourselves. Hopefully, “the existence of labels” is among your list.
This is sufficient for now, because the rest of the Course 1 and most of the Course 2 are not about their difference. Even though they will only cover Sup. Learning, the idea of Sup. Learning is not the main theme. The main theme is Gradient Descent and some accompanying techniques that we use in training every neural network.
These courses are more like preparatory courses for you to learn Nerual Network, rather than focusing on Sup Vs. UnSup. However, I believe the course still wants to give you a heads-up to the two terms and something simple to digest for starter. Therefore, there is no need for us to dive deep and read more either, for the sake of these courses. Indeed, throughout this 10-week, 3-course series, you will only find 1 week of material to be about UnSup. learning.
This is my reason not to recommend more reading to dive deep, but if you just want some general discussion, I think googling “Supervised Vs. Unsupervised Learning” will give you many such results. However, if I were you, I would not struggle for too long in them, but to quickly move on to Gradient Descent which is what these courses are for.