Tips to get a campus quality education using youtube, I call it "University of Youtube."

I did my master in Stochastic Finite Element Analysis at IITM. FEA and Deep Learning have few similarities. Both have discretization (functional decomposition), matrix formulation, lots of partial differentiation and optimization. This assisted me to get started into learning using data.

What really helped was maintaining academic rigor and focus on math to gain a grip on Symbolist, Analogizers and Connectionist tribes of machine learning algorithms and RNNs, CNNs, and GANs in Deep learning. For this, I used videos on youtube and content on the web. I call this ... "University of Youtube".

In this blog, I will give few tips on how to do it. Long time back the activities I used to do on the campus are now entirely possible with youtube and web. As a student, my life used to revolve around three facilities on the campus for all my academic needs. Let us focus on how to substitute first two using the web.

1. The teaching facility
2. Library
3. Computer Center.

Teaching Facility:-

Some of the most important Deep Learning and Machine Learning classes taught on campus are also available on Youtube. I can fill the whole page with links. Let me give one example that I used a couple of years back.

e.g., Caltech's Machine Learning Course - CS 156 by Professor Yaser Abu-Mostafa This was one of the courses which taught me why learning is possible with data. 

With online videos, the problem is not the lack of ability to communicate back to the prof. There is a highly qualified community willing to help you on Quora and Stackoverflow. If you email, the prof might also respond. Problems are lack of a sense of urgency and video hopping. Before I realized I was switching videos like I switch TV channels on the first encounter of a difficult concept. Some tips to address these problems.

Tip1:- Like on campus, have a game plan with the help from mentor(s). Explore the videos from profs from many reputed universities on youtube. Decide on one or two at most and put a firm date and complete them. 

Where to find the mentors? Unlike on campus, you will not have one guide or a General test committee. A mentor can be anything, a blog, a video, advice from a friend, courses like Coursera, Edx, and Udacity, sites like fast.ai. It is an ensemble of all these week recommendations and your own heuristics which will give direction.

Tip2:- Never fail to click "More details" below the video. You will find links to Prof's course page. Visit the course page. Look at the assignments and notes. These links are invaluable. e.g., Caltech Machine Learning Course page

Tip3:- Make sure you go through all the comments on the youtube video. Start learning from the community. If prof has written a book, read it. Look around you might find a free copy the prof himself had provided or a digital copy at some student's home directory or GitHub.

Tip4:- You will find the prof teaching the same course on Edx or Coursera. It will cost you $49 to $99. Believe me, it is a steal. Enroll.

Auditing Supplemental Courses - 

On campus, it is common to audit many courses to supplement the research and also to perform better on the main course. It is done in a planned way. You do not hop from one class to another. Even the courses you audit, are taken seriously. Follow the same process when you do it online as well. 

Tip5:- When you encounter a difficult concept say "Lagrange multipliers" for SVM use Wikipedia or 15min videos and come back to course don't go on a tangent into "Constrained Optimization". Collect these topics, in a very deliberate and planned way make a choice based on your time and audit these courses as you do on campus. Topics like gradient techniques, discrete optimizations, and integer programming are very complex. Just because videos are available do not digress into them. They require planning and preparation. 

Tip6:- For Deep Learning in my opinion "Stanford courses" are the best to learn the concept. They are a bit more on to the application rather on pedagogy. Combine Udacity with Stanford Courses for good results.

Tip7:- Once I finish a topic, I will look for a related course on nptel. This is my personal choice. I like to validate my mathematical intuition, these are a continuous stream of math equations and derivations with blackboard and chalk by an expert. They are heavy on math and one can only understand if they got grip on the fundamentals of the topic. 

e.g., clustering - "Pattern Recognition by Prof. C.A. Murthy" NN - Prof.S. Sengupta RL Prof. Ravindran

Tip8:- There are some TAs doing recitation. Do not miss them. 

Discipline and planning are key to get campus quality teaching for deep learning and machine learning using online content. I say it is possible.. go for it...

Library:-

Follow the same rules you follow on campus when doing research apply when you grokking papers on the web. I used to visit the library with one paper in mind and in no time I end up in referring journal after journal. Fortunately, as the resources in the library were limited my exploration comes to an end and I used to go back to working on the main paper. 

On the web, it is a different story. All the important papers are freely available in unlimited supply. You will find a free copy of the ones even behind a paywall. If you do not plan properly you will never complete a paper and never restart your computer because you have a browser with tons of tabs open. 

Tip8:- Develop your own grokking workflow. I like to pocket and download them to pdf expert on my iPad and liquid text for later reference and focus on the main paper. 

Tip 9:- There are tons of blogs with DL papers. Use subreddits to find new papers. Here is a link to Machine Learning subreddit. There are similar ones for AI, DL etc.,.

Why is reading papers important? one example

Without reading the tensorflow paper I would not have discovered true reasons for computational graph and automatic differentiation. This helped me to understand other DL papers. Now, in a DL paper, I look for the optimization strategy and zero-in on the objective function and see how the computational graph is built. This approach makes it easy to understand the Network architecture, and not worry about the framework it is implemented in. Often in ML and DL the techniques are empirical and it better to hear the tips and techniques directly from the horse's mouth.

Tip 10:- Where to start to learn Deep Learning? 

I will take a different analogy. Compare web, youtube, and GitHub to an actual theme park where the true action is. Udacity is like the information center. It gives all the info needed to get started. But it would be foolish to just return home after visiting the information center. You need to enter the theme park, explore and exploit. 

I have signed up for my next adventure. I hope it will improve my current understanding to Vision and NLP as it did for Deep Learning. 

I want to end this blog by saying times are changing, continuous learning is no more optional. I hope some of these tips would help you to get you started.