Where to start on neural networks

Question

First of all I know the question may be not suitable for the website but I'd really appreciate it if you just gave me some pointers.

I'm a 16 years old programmer, I've had experience with many different programming languages, a while ago I started a course at Coursera, titled introduction to machine learning and since that moment i got very motivated to learn about AI, I started reading about neural networks and I made a working perceptron using Java and it was really fun but when i started to do something a little more challenging (building a digit recognition software), I found out that I have to learn a lot of math, I love math but the schools here don't teach us much, now I happen to know someone who is a math teacher do you think learning math (specifically calculus) is necessary for me to learn AI or should I wait until I learn those stuff at school?

Also what other things would be helpful in the path of me learning AI and machine learning? do other techniques (like SVM) also require strong math?

Sorry if my question is long, I'd really appreciate if you could share with me any experience you have had with learning AI.

Answer 1

No, you should go ahead and learn the maths on your own. You will "only" need to learn calculus, statistics, and linear algebra (like the rest of machine learning). The theory of neural networks is pretty primitive at this point -- it more of an art than a science -- so I think you can understand it if you try. Ipso facto, there are a lot of tricks that you need practical experience to learn. There are lot of complicated extensions, but you can worry about them once you get that far.

Once you can understand the Coursera classes on ML and neural networks (Hinton's), I suggest getting some practice. You might like this introduction.

Answer 2

I would say... it really depends. You may need to:

• use machine learning algorithms: this will be useful for specific applications you may have. In this situation what you need is some programming skills and the taste for testing (practicing will make you strong). Here maths are not so much required I would say.
• be able to modify existing algorithms. Your specific application may be reticent to regular algorithms, so you may need to adapt them to get maximum efficiency. Here maths come into play.
• understand the theory behind algorithms. Here maths are necessary, and will help you increase your knowledge of the field of machine learning, develop your own algorithms, speak the same langage as your peers... NN theory may be primitive as said by @Emre, but for instance this is not the case for SVM (the theory behind SVM requires e.g. to understand reproducing kernel Hilbert spaces).

On the mid term for sure you will need strong maths. But you don't need to wait for them to come to you, you can start right now with linear algebra, which is beautiful and useful for everything. And in case you encounter (possibly temporary) difficulties of any sort with maths, keep on practicing the way you already do (many people can talk about the perceptron but are not able to make a perceptron in Java), this is very valuable.

Answer 3

Neural Networks are not a great introductory model, simply because of the complexity that you describe. If you're trying to get your feet wet, boosted decision trees tend to perform well by comparison, and are a bit more intuitive. If you want a description on this method, and are already familiar with Coursera, The University of Washington has an introductory course on data science which explains it quite well.

Answer 4

For sure you need to learn some maths. However you should also make an effort to gain some broader engineering and science skills. There are far too many people going into computer science and all they know is a few programming languages and math. The end result is a very boring person with little in the way of creativity to do anything new. Take a year out when you are 18 or 19 to travel the world.

Answer 5

This is a hell good book. Adrian is going to make black friday sale, so it's great chance to pick it up. It's step by step guide through deep learning with math, intuition and code. It focuses mostly on computer vision, but it'll give you nice start.