A Machine Learning cheatsheet focusing on the most popular algorithms | Hacker News Reader

A Machine Learning cheatsheet focusing on the most popular algorithms

github.com

90 pointsmarban10 comments

Misleading right off the bat. “Linear” and “parametric” are orthogonal concepts. A parametric model can be linear or non-linear (I.e you can fit the parameters of a model f(x) = A * sin(w*x)). A non-parametric model can also be a linear combination of examples, or not.

Nothing included about cross-validation which is probably the most important concept in modern machine learning. Not sure what to think about that.

Could you elaborate? Any good resources on that?

In the Ordinary Least Squares section, this sheet says the vector of residuals is equal to the sum of the squared elements of that vector.

Since this will become obsolete in 5 years of sooner, it's just a great introduction. Not something hanging in my room/office.

Machine Learning is merely a bunch of statistical algorithms that have been around for centuries. So although the name/branding may indeed change in a few years, the content will be the same.

AlphaGo changed my worldview. It was the first sort of singularity that divides my history of the world into before and after. Now I don't doubt that we will rapidly continue to develop the algorithms and hw too that is all very different from what is believed to be known.

"Centuries" is a bit of a stretch, right? Or am I misunderstanding your comment?

Some of the most widely used algorithms in machine learning, for training or prediction, are much younger than that: - backpropagation (1970s) - Viterbi algorithm (1960s) - Markov Chain Monte Carlo (1940s) (Though, the Buffon's needle problem is from the 18th century.)

And things like conditional random fields, LDA, and LSTMs, were introduced after 2000.

edit: removed lone parenthesis

LSTM 1997

https://dl.acm.org/citation.cfm?id=1246450

I was referring mostly to Bayes, PCA and the regression algorithms which are widely used.

Reply on news.ycombinator.com