What's the best way to self teach machine learning and AI?

[wasab]

I'm about to graduate and know nothing about AI and machine learning. I guess it's kinda my fault because I keep picking electives in Unix operating system and development instead of machine learning like computer vision, natural language processing, and whatnots. What's the best way to self teach these things? I am kinda interested in data science aspect of machine learning algorithms. 

[Sauron]

I guess youtube tutorials...? ML and AI aren't that complicated to understand, it all comes down to getting the base idea and remembering the most used types of network. Unless you're specifically doing AI research with massive compute farms dedicated to that you will just be using premade models anyway.

[wasab]

1 minute ago, Sauron said:

I guess youtube tutorials...? ML and AI aren't that complicated to understand, it all comes down to getting the base idea and remembering the most used types of network. Unless you're specifically doing AI research with massive compute farms dedicated to that you will just be using premade models anyway.

From what I've seen, ML seem to involves lots of math. Wouldn't I miss out on a lot of underlying theories and concept by watching YouTube? 

[straight_stewie]

7 minutes ago, wasab said:

What's the best way to self teach these things? I am kinda interested in data science aspect of machine learning algorithms. 

For someone at your level of experience the best way to self teach is probably just to build the application you want. Maybe follow an introductory tutorial or two at the beginning of your learning to get the basic approaches and the vernacular down.

There are more than a few organizations that distribute already classified datasets of various types to the general public. These should prove useful for your learning, as you wouldn't have to classify 10 bajillion images or anything crazy like that. The Google Search "AI competition" or "AI challenge" will turn up plenty of these datasets of various types.

[Slottr]

MIT has tons of lectures freely available to watch

 

https://ocw.mit.edu/courses/electrical-engineering-and-computer-science/6-034-artificial-intelligence-fall-2010/index.htm

[Sauron]

1 minute ago, wasab said:

ML seem to involves lots of math

The theory yes, the practice... not at all really. Again, since developing a model requires way more processing power than any single person has access to and a lot of trial and error, most people (and most companies hiring ML "experts") just end up using premade models without needing to know almost any theory beyond the basics of how you train the model (or more commonly how you retrain the last few layers) and how you import it in python.

[Devin92]

This is where I started. It's pretty old but gives the introduction for what is neurons, neural network, and back propagation etc.

http://neuralnetworksanddeeplearning.com/

 

then this for introduction to CNN

http://cs231n.stanford.edu/

 

I think anyone want to join the machine learning area cant avoid these two fundamentals.

 

[Raytsou]

8 hours ago, Slottr said:

MIT has tons of lectures freely available to watch

 

https://ocw.mit.edu/courses/electrical-engineering-and-computer-science/6-034-artificial-intelligence-fall-2010/index.htm

Was about to suggest the same thing. MIT offers most of its courses online for free to the public via ocw (OpenCourseWare). 6.034 is intro to AI and 6.036 is intro to ML. Can't speak for 034 but 036 is a great class, the lecturer is great. Would definitely also recommend doing the problem set assignments as they help build a very deep understanding of the material. There's also 6.867, which is the graduate class for ML but is also available online.

 

8 hours ago, wasab said:

From what I've seen, ML seem to involves lots of math. Wouldn't I miss out on a lot of underlying theories and concept by watching YouTube? 

It really isn't. If you've taken any linear algebra course the math is really trivial. It's just matrix manipulation to store data, but you can easily visualize it a different way, the math is just a convenient method to work with complicated things. You don't really have to understand it, just know how it works to use it.

I highly recommend the youtube channel 3Blue1Brown. He has a series on linear algebra, followed by a series on neural nets (also a good watch to get started on ML).

8 hours ago, Sauron said:

The theory yes, the practice... not at all really. Again, since developing a model requires way more processing power than any single person has access to and a lot of trial and error, most people (and most companies hiring ML "experts") just end up using premade models without needing to know almost any theory beyond the basics of how you train the model (or more commonly how you retrain the last few layers) and how you import it in python.

Very true. The actual implementation of ML is mostly just importing standard ML packages, reading the documentation, and plugging in your data set. What's difficult is fine tuning your constants to achieve the desired local minima. This requires a bit more understanding of the theory behind to be able to understand how to adjust things in expected ways.

[igormp]

20 hours ago, wasab said:

I'm about to graduate and know nothing about AI and machine learning. I guess it's kinda my fault because I keep picking electives in Unix operating system and development instead of machine learning like computer vision, natural language processing, and whatnots. What's the best way to self teach these things? I am kinda interested in data science aspect of machine learning algorithms. 

I assume that you have some basic knowledge on calculus and linear algebra since you're graduating. If you really want to get on with the theory and whatnot, Andrew Ng's course on coursera is amazing. Knowing that is useful if you ever need to fine tune some small details in your model (as a ML engineer, that's a large chunk of my daily work along with deploying stuff).

 

Other than that, following simple tutorials on how to do simple classifiers on youtube should give you a nice head start.

[wasab]

13 hours ago, igormp said:

I assume that you have some basic knowledge on calculus and linear algebra since you're graduating.

I used to know differential, integral, and multivarible calculus as well as differential equations. I forgot most if not all by now. Although if i look at my old homeworks and exams, it will come back to me quick. Linear Algebra i dont know much beyond working with matrices and finding eigenvector and eigenvalue. 

 

13 hours ago, igormp said:

If you really want to get on with the theory and whatnot, Andrew Ng's course on coursera is amazing. Knowing that is useful if you ever need to fine tune some small details in your model (as a ML engineer, that's a large chunk of my daily work along with deploying stuff).

Any resources that are free? 

[wasab]

On 1/16/2020 at 4:23 PM, Slottr said:

MIT has tons of lectures freely available to watch

 

https://ocw.mit.edu/courses/electrical-engineering-and-computer-science/6-034-artificial-intelligence-fall-2010/index.htm

I really cant stand college style lectures. These usually put me to sleep. I prefer youtube videos like khan academy.  

[wasab]

On 1/16/2020 at 4:24 PM, Sauron said:

The theory yes, the practice... not at all really. Again.

 

On 1/16/2020 at 4:21 PM, straight_stewie said:

For someone at your level of experience the best way to self teach is probably just to build the application you want. 

I wish to know the theory though. 

[Slottr]

9 minutes ago, wasab said:

I really cant stand college style lectures. These usually put me to sleep. I prefer youtube videos like khan academy.  

Not to be an ass, but I’d say suck it up lol

 

thats the most direct form of information I think you’ll find on such a large subject 

[Sauron]

6 hours ago, wasab said:

I wish to know the theory though. 

Well then proper lectures are the best way, video tutorials rarely go sufficiently in depth.

[igormp]

7 hours ago, wasab said:

I used to know differential, integral, and multivarible calculus as well as differential equations. I forgot most if not all by now. Although if i look at my old homeworks and exams, it will come back to me quick. Linear Algebra i dont know much beyond working with matrices and finding eigenvector and eigenvalue. 

 

Any resources that are free? 

Andrew's course is free and has enough theory along with exercises.