I want to learn maths for artificial intelligence. I heard that main areas of mathematics for AI are Multivariable Calculus, Linear Algebra, Probability and Statistics and Discrete Mathematics. But what level of these subjects do I need?

For example the easy level is like Calculus of Gilbert Strang and 18.01 Calculus from MIT Open Course Ware. The hard level is Apostol's or Spivak's Calculus.

Do I need easy level of maths or hard level to be good at artificial intelligence and its areas (machine learning, pattern recognition).

From what I've seen in different papers pertaining to AI, there's lots of graph theory, probability and statistics, and calculus, but I don't think you need much rigor and formalism (what I think you're referring to by hard level) for the AI/ML applications. Computer vision requires lots of linear algebra, which you may be interested in too.

There's a bit of a misconception about math in comp-sci field; a lot of people tend to think that high level math is necessary to solve problems, but the truth is that using the monte carlo method can often achieve results close enough to where you don't need to actually spend a lot of compute time on getting exact values. There are obviously specializations where those are not completely true statements, but there are places where it is, for example game and game engine development (i find this to be the are where this misconception applies the most). If you want to learn math, if you like math (i hate the way that math is taught), the above suggestions by Sato are very true. If you are however really unsure of what you want to work on, it seems that learning high level math is more of a solution in search of a problem. Once you outline the problem, it may be easier to determine what it is you need to learn to solve it...

There's a bit of a misconception about math in comp-sci field; a lot of people tend to think that high level math is necessary to solve problems, but the truth is that using the monte carlo method can often achieve results close enough to where you don't need to actually spend a lot of compute time on getting exact values. There are obviously specializations where those are not completely true statements, but there are places where it is, for example game and game engine development (i find this to be the are where this misconception applies the most). If you want to learn math, if you like math (i hate the way that math is taught), the above suggestions by Sato are very true. If you are however really unsure of what you want to work on, it seems that learning high level math is more of a solution in search of a problem. Once you outline the problem, it may be easier to determine what it is you need to learn to solve it...

This holds true especially for cellular automation. Conway's game of life does not require any major mathematical concepts, but only relies on theory of cellular automation.

There's plenty of methods. The one I use is just linear algebra; it's basically just counting. I prefer not to use graphs unless I'm trying to find patterns in sound. Some major companies prefer that you have experience using praat for that task, it's openware so try it out if you're working with sound. I like to go straight to the source though and use bits as opposed to graphs because it seems to be a lot easier to work with (and MUCH more precise. It's basically 100% confidence with identification), but if you're using python, you can import graph modules. Making probabilities, or using statistical inference, is mostly useful for determining the boundaries between units of knowledge so you don't accidentally interpret more than what's usual. Beyond that, it's just making the program pretty (like by adding confidence thresholds). If you have any specific questions I'll help. It's amazing what people are able to do with machines now that we've discovered pattern recognition.

 

http://youtu.be/pp89tTDxXuI

 

I think I said this in a different thread, but most of what I've been seeing from the intellectual community are toys. I know that google and amazon have ambitions but from the google frontier, I really haven't seen any advances for some time. Last one I noticed was about 3 years ago with speech recognition. They do have quick answer boxes too but it's not anything like AI at this point. The FBI is using facial recognition now, so that's good news, but the potential for this type of software is huge. You just need to be able to sell yourself.

Edited March 25, 201411 yr by Popcorn Sutton

By far my favorite quad copter talk