Live data from Hacker News

Understanding Machine Learning: From Theory to Algorithms (2014)

cs.huji.ac.il

41–50 of 58 posts

Re: Understanding Machine Learning: From Theory to Algorithms (2014)

#41
post #33
post #30

Earlier quoted context omitted.

> A needle of length L is dropped at random onto a sheet of paper ruled with parallel lines a distance L apart. What is the probability that the needle will cross a line? Thickness of line is needed right? Otherwise P approaches 100% as thickness approaches 0?

The lines are infinitely thin. Equivalently you can imagine the paper is divided into regions of width L, and the question is whether the needle will cross a region boundary ( https://en.wikipedia.org/wiki/Buffon's_needle ).

I don't think that page explains it very well, but have poor math background. I imagined notebook paper with horizontal lines spaced L apart and then the needle dropping at any angle. When the needle is vertical the probability it cross a line is 1, when horizontal it is zero. The length of the needle L is the hypotenuse of a triangle. If we call the angle from horizontal x, the "height" of the needle can be anywhere within h=Lsin(x) for x between 0 and pi/2.

The "lines" are like a sample of a point from a uniform distribution U with width L, and h is an interval inside U. The probability a number sampled from a distribution of width L will fall within interval h is h/L. Substituting for h gives p(cross|x) = sin(x).

Then assuming the needle is equally likely to drop at any angle, for any one angle theta we get probability density p(theta=x) = 1/(pi/2-0)= 2/pi.

The probability the needle drops at angle x AND crosses a line is the product of p(theta=x)p(cross|x)= (2/pi)sin(x). As mentioned, x can range between 0 and pi/2. To get the probability the needle drops at angle x1 OR x2 OR x3, etc and cross we need to sum all these. So take the integral of (2/pi)sin(x) from 0:pi/2. This gives 2/pi.

Re: Understanding Machine Learning: From Theory to Algorithms (2014)

#42
post #26
post #7

I feel like the barrier to machine learning for me, as I've seen in many tutorials and books and is an immediate discouragement, is the massive amount of math thrown in your face. Many of us didn't just graduate, need glasses and fall asleep at 8pm on the couch when the kids are in bed... Math is this distant fragment of memory buried under years of everything not Math. It feels like machine learning is only taught b…

I interview a lot of developers for ML positions at our company. The first red flag is always a lack of math. Candidates who come in with API-level competence ie. can implement an ML algo using this,that or the other API, without any understanding of some basic math behind it, always fare poorly. Atleast in ML, not having an understanding of math is pretty much like claiming expertise in riding a bicycle by watching…

Not to detract from your point that math is important, but in that example proper methodology (e.g., cross-validation), proper feature engineering, and especially domain knowledge are probably even more important. You can be aware of the strengths & weaknesses of different machine learning algorithms without being intimately familiar with the math. Ideally, ML methods are not treated as a "black boxes", but some aspects are inherently black box, even if you do know the math (e.g., parameter tuning).

Re: Understanding Machine Learning: From Theory to Algorithms (2014)

#43
post #26

Earlier quoted context omitted.

I interview a lot of developers for ML positions at our company. The first red flag is always a lack of math. Candidates who come in with API-level competence ie. can implement an ML algo using this,that or the other API, without any understanding of some basic math behind it, always fare poorly. Atleast in ML, not having an understanding of math is pretty much like claiming expertise in riding a bicycle by watching…

A lack of mathematical intuition is a serious problem for many people from engineering to biology to economics. It certainly plagued me throughout my engineering bachelors studies and is something I continually work to get better at. In my opinion, physics students learn the best framework for thinking and get a very good mathematical intuition. For example, here's a problem from an introductory QM book that really t…

> A lack of mathematical intuition is a serious problem for many people from engineering to biology to economics.

This is true. I can't really say why, but after my Discrete Mathematics class a lot of my Computer Science problems became a lot easier to reason about.

Re: Understanding Machine Learning: From Theory to Algorithms (2014)

#44
post #26

Earlier quoted context omitted.

I interview a lot of developers for ML positions at our company. The first red flag is always a lack of math. Candidates who come in with API-level competence ie. can implement an ML algo using this,that or the other API, without any understanding of some basic math behind it, always fare poorly. Atleast in ML, not having an understanding of math is pretty much like claiming expertise in riding a bicycle by watching…

A lack of mathematical intuition is a serious problem for many people from engineering to biology to economics. It certainly plagued me throughout my engineering bachelors studies and is something I continually work to get better at. In my opinion, physics students learn the best framework for thinking and get a very good mathematical intuition. For example, here's a problem from an introductory QM book that really t…

For such a problem, usually "at random" means a uniform distribution. But on the plane, there is no uniform distribution. So, the "paper" can't be the plane. So, it might be fair to ask the size of the paper and what happens with the needle near the edges? E.g., on a rectangular sheet of paper of finite size, the needle can land in a position so that it does not cross a line but would on a larger sheet of paper.

Re: Understanding Machine Learning: From Theory to Algorithms (2014)

#45

I'm a math geek, but I'm also a mostly self-taught data scientist. "The Elements of Statistical Learning" ( https://web.stanford.edu/~hastie/local.ftp/Springer/OLD/ESLI... ) is far and away the best book I've seen. It took me hundreds of hours to get through it, but if you're looking to understand things at a pretty deep level, I'd say it's well-worth it. Even if you stop at chapter 3, you'll still know more than mos…

Would you say elements is superior to Bishop's book?

Re: Understanding Machine Learning: From Theory to Algorithms (2014)

#46
post #7

I feel like the barrier to machine learning for me, as I've seen in many tutorials and books and is an immediate discouragement, is the massive amount of math thrown in your face. Many of us didn't just graduate, need glasses and fall asleep at 8pm on the couch when the kids are in bed... Math is this distant fragment of memory buried under years of everything not Math. It feels like machine learning is only taught b…

Not trying to detract from worthwhile discussion here ... but for a hack-first introduction, try this: https://github.com/hangtwenty/dive-into-machine-learning

Re: Understanding Machine Learning: From Theory to Algorithms (2014)

#47
post #19
post #7

I feel like the barrier to machine learning for me, as I've seen in many tutorials and books and is an immediate discouragement, is the massive amount of math thrown in your face. Many of us didn't just graduate, need glasses and fall asleep at 8pm on the couch when the kids are in bed... Math is this distant fragment of memory buried under years of everything not Math. It feels like machine learning is only taught b…

> the massive amount of math thrown in your face I can skip the math part - we don't all invent new algorithms - but what I really need is a large enough & gradual set of problems to solve (datasets + verification scripts). I mean, start from the simplest and teach people how to use the already available software. Machine Learning should be assimilated practically, too much theory with too little application is usele…

Try this :) https://github.com/hangtwenty/dive-into-machine-learning

Re: Understanding Machine Learning: From Theory to Algorithms (2014)

#48
post #12
post #7

I feel like the barrier to machine learning for me, as I've seen in many tutorials and books and is an immediate discouragement, is the massive amount of math thrown in your face. Many of us didn't just graduate, need glasses and fall asleep at 8pm on the couch when the kids are in bed... Math is this distant fragment of memory buried under years of everything not Math. It feels like machine learning is only taught b…

Andrew Ng's Coursera ML course is supposed to be pretty accessible. I've also heard good things about Machine Learning for Hackers ( http://www.amazon.com/Machine-Learning-Hackers-Drew-Conway/d... ). Ultimately, ML is a mathematical discipline. You can ask for a gentle approach that gets you to the foot of the mountain, but "if you want to learn about nature, to appreciate nature, it is necessary to understand the la…

I'm a high school Senior (17) and I'm currently taking it. It is ridiculously understandable and I often see myself yearning for more. But I've also had Calc 3 + Linear Algebra by now, so its understandable that not everyone would get it. The intuition is simple, however.

Re: Understanding Machine Learning: From Theory to Algorithms (2014)

#49
post #26
post #7

I feel like the barrier to machine learning for me, as I've seen in many tutorials and books and is an immediate discouragement, is the massive amount of math thrown in your face. Many of us didn't just graduate, need glasses and fall asleep at 8pm on the couch when the kids are in bed... Math is this distant fragment of memory buried under years of everything not Math. It feels like machine learning is only taught b…

I interview a lot of developers for ML positions at our company. The first red flag is always a lack of math. Candidates who come in with API-level competence ie. can implement an ML algo using this,that or the other API, without any understanding of some basic math behind it, always fare poorly. Atleast in ML, not having an understanding of math is pretty much like claiming expertise in riding a bicycle by watching…

As an undergrad interested in machine learning, what areas of math specifically should I be focusing on?

Re: Understanding Machine Learning: From Theory to Algorithms (2014)

#50
post #26

Earlier quoted context omitted.

I interview a lot of developers for ML positions at our company. The first red flag is always a lack of math. Candidates who come in with API-level competence ie. can implement an ML algo using this,that or the other API, without any understanding of some basic math behind it, always fare poorly. Atleast in ML, not having an understanding of math is pretty much like claiming expertise in riding a bicycle by watching…

As an undergrad interested in machine learning, what areas of math specifically should I be focusing on?

1.math prereqs at UW:http://courses.washington.edu/css490/2012.Winter/lecture_sli...

2. math prereqs at UCSC:http://people.ucsc.edu/~praman1/static/pub/math-for-ml.pdf

3. math prereqs at UMD: https://www.umiacs.umd.edu/~hal/courses/2013S_ML/math4ml.pdf

Post reply on HN