Going From Not Being Able to Code to Deep Learning Hero

A detailed plan based on experience

Before you can learn to do deep learning, you must become a developer. Being a developer is not only about programming. In the larger scheme of things, being able to write code is just a small part of what a developer can do.

If I were to start all over, I would optimize this part of the journey for fun. I would want to get into the tinkering mindset and start feeling comfortable messing around on my computer. There is no better way to achieve this than by working on something you truly enjoy.

I would take it easy. When it comes to learning, consistency beats intensity every time. When you are just starting something, there is no better way to stay consistent than to not exert yourself too much. I would still make quick progress, because this time around, I would know what to focus on. This is what the beginning of this chapter will cover.

Harvard CS50 has one great thing going for it: On top of teaching you how to program, it also teaches you about the computer and how it works! Two birds with one stone. Still, regardless of which MOOC or programming book you chose, it is hard to go wrong. Programming has been taught for many generations, and many good teachers are out there. The biggest mistake you can make at this stage is to do more than a single beginner programming course. Striving to learn any particular programming language really well would be a mistake.

It doesn’t make sense to invest all of your time into learning calligraphy if you have nothing to write about! Likewise, it doesn’t make a lot of sense to work on building a table but focus all your effort on a single leg and then complain that the table will not stand!

All we need at this stage is some familiarity with programming concepts — enough so that we can perform meaningful online searches. The programming skills we need are understanding the value of Stack Overflow, documentation, and how to reach both. This might sound tongue in cheek, but these skills are genuinely at the core of a programming career. Add knowing a few basic things about programming, and you have a very firm platform on which to build.

The next leg of the table we are assembling is our ability to move effectively in code. This is not something that most programming courses cover, but it is an essential skill to have.

This boils down to learning to use a code editor really well. Any major code editor will do. Some people swear by textmate, sublime text, atom, vscode, emacs, or vim. It doesn’t matter which one you choose. Focus the same attention on learning it as you would to writing code and you will be alright. If you are absolutely lost as to which one to pick, chose vscode. Do not worry about the choice too much — it is not one you must stick with forever. I do not know of a single developer who has not tried multiple editors in the course of their career. However, I also don’t know a single developer worth their salt who doesn’t believe that learning to use an editor is an important skill.

To work on our next leg, we move to version control. Code is written in small iterations, often by more than one person. Additionally, code is so expensive to produce that we want to make sure we don’t lose our work and that we can explore new directions without destroying the work we have done up to now. All this is very hard to achieve without specialized tools. Thankfully, we have a software solution that addresses all the needs listed above and then some!

The tool goes by the name of git. It is a small command line program that can change any directory into a git repository. In the process, it endows the directory with superpowers, such as an ability to travel back in time or create alternate timelines!

Git comes with an online counterpart, called GitHub. While git excels at managing the source code on your computer, GitHub is great for code sharing and collaborating with others. Pushing your code to GitHub also provides you with a backup. If some accident should befall your PC, your work will be preserved at another location.

The last leg of our four-legged table is probably the hardest to construct. The challenge lies solely in the lack of good learning resources. This part deals with learning how to use the computer in the context of writing code. For deep learning, this goes beyond your local setup. How do you spin up a cloud VM? How do you ssh into it? How do you move data around? How do you build your own home rig for experimentation? The good news is that you do not have to learn all of this at once. A great starting point might be to learn how to navigate the file system on the operating system of your choice. As you get more comfortable using the command line, you can gradually add more commands to your arsenal.

Maybe a day will come when a single course will cover all the basics of being a developer. For now, unfortunately, we must piece together our own curriculum. The good news is that it is enough to get just a taste of each of the four disciplines I outline above; there is no need to become an expert at any of them.

In fact, for many, the course goes beyond showing one how to be a good deep learning practitioner. Through taking the fast.ai courses, I learned how to learn. I also started to share my work and became an active member of the deep learning community. The courses not only taught me nearly everything practical that I know about machine learning but also showed me how to thrive in the digital age.

It all begins with watching a lecture. The next step is to open the lecture notebook and figure out how all the pieces fit together. The idea is to run each line of code and look at the outputs. If you come across a function you have not previously encountered, this is a good time to read its documentation. How will the performance differ if you pass in different arguments or slightly tweak the hyperparameters?

Once you understand the bigger picture, it’s time to reproduce the results. To do so, open a new notebook and try to recreate the training pipeline that the lecture demonstrated. It’s an open-book exercise — it’s okay if you need to look back at the lecture notebook, but the less you do so, the better. On the other hand, there are no limitations on reading the documentation or searching online. Both of these activities are encouraged and they closely resemble what programming feels like when it’s done completely on your own.

The idea is to systematically grow your skills. You start in a controlled setting of the lecture notebook, learning the ins and outs of the technique that the lecture discussed. With every new exercise, more of the training wheels come off, to the point where you’re doing everything on your own from start to finish.

This approach to learning works like a charm. There are no hard and fast rules here and you are more than welcome to modify it as you go. Being creative and coming up with exercises that are challenging and interesting to you is a very valuable skill in its own right; acquiring it will take your learning to the next level. The result is always the same: to arrive at a place where you can comfortably apply the techniques you learn in the lectures to problems you haven’t previously seen.

While practice is the only way to improve your ability to solve real-life machine learning problems, moving toward this way of learning does not require you to revolutionize your life. A couple of minutes a day, over multiple days, is all you need to get started. With every consecutive day of learning, you build momentum. Soon you start to notice the results, and the ability to do things you thought were out of reach not too long ago creates a lot of excitement. Very soon, you will find yourself losing track of time as you explore this or that technique.

The way of learning I describe above was completely new to me. However, after experiencing what it can do, I have never ceased being a fast.ai student. Any other approach feels like a waste of time.

It wasn’t always like that. Initially, I thought that the fast.ai way of learning could not work. The exercises proposed in the lectures didn’t feel like learning to me. How could I make any progress spending so much time doing? To me, learning was this process in which you sat down with a book for an hour or two and came out on the other end being able to discuss something highly complex and abstract.

I was very lucky, as at the time I found fast.ai I was at the end of my rope. Nothing was working, so I was ready to try something completely new.

Stumbling across fast.ai turned out to be the adventure of my life.

Related posts:

My husband and I have been snowbirding it for the past seven years, but decided we’d like to spend this winter in Mexico rather than at our homestead south of Quartzsite, Arizona. We chose to travel…

If you want to be a more productive writer, incorporate these six tweaks into your daily routine. Based on lessons learned by writer Nicholas E. Barron.

Linear regression is a statistical method used to model the linear relationship between a response variable and one or more predictor variables. In a linear regression model, we aim to predict the…