Teaching Philosophy Data Science and Deep Learning from fast.ai


Rachel Thomas, co-founder of fast.ai, professor at the USF Data Institute

Paul Lockhard, Ph.D. in mathematics at Columbia University, former professor at Brown University and middle school math teacher, in his important essay “Crying math” describes a terrible world where children can’t listen and play music until they spend decades mastering musical notation and music theory, transposing notes to a different key in their classes. In drawing lessons, students learn colors and brushes, but can only start painting in college. Sounds absurd, right? This is exactly what mathematics is taught – we require students to spend years cramming, studying dry, unrelated “foundations”, which, we argue, will pay off in the future, when most people will leave the subject itself.

Unfortunately, this is where the mention of some of the few deep learning resources begins; learners are asked to know the Hessian definition and Taylor series expansion theorems for the loss function, but they never give examples of actually working code. I do not blame the math. On the contrary, I love mathematics and even taught it at the college level, but I do not think that this can be called a good or useful introduction to deep learning.

So many students drop out of math because of how boring it is to teach with tedious, repetitive exercises and a curriculum that puts aside so much fun (e.g. graph theory, calculations and permutations, group theory) until later, when all students are with the exception of mathematical specializations – drop the subject. And the guardians of deep learning gates do something similar whenever they ask for a multiple chain rule or theoretical justification of the Kullback-Leibler distance before they teach you how to use a neural network for your own projects.

We will use the best available research on teaching methods, trying to solve problems with technical training, including:

  • Training “Whole game” – starting with a demonstration of how to use a complete, working, very useful, modern deep learning network to solve real problems using simple expressive tools. And then gradually dig deeper and deeper into the understanding of how these tools are created, and how the tools are created that create these tools and so on …
  • Always teach with examples: make sure there is a context and purpose that you can understand intuitively; it’s better than starting with algebraic manipulation of characters.
  • Simplify as much as possible: we spent months on the creation of tools and teaching methods that at times simplify topics that previously seemed complicated.
  • Remove barriers: to date, deep learning has been very exclusive.
  • Now we are making it more accessible so that everyone can participate.
  • …and many others. Some of our techniques I have discussed in more detail below.

In the end, we are talking about quality education. This is what excites us the most. Below are our thoughts on quality education:

Quality education begins with “Whole games”

Just as children have an idea of ​​baseball even before the first training session, we want you to have an idea of ​​deep learning in general long before you start learning mathematics and the chain rule. We will move away from the general picture of deep learning to details (that is, move in the opposite direction compared to most educational institutions that try to teach all the elements separately, and only then add them to the general picture). A good example would be inJeremy’s talk on recurrent neural networks – he starts with a three-linear RNS, using a high-performance library, then removes the library, costs his own architecture using the GPU framework, and then removes it and builds everything from scratch in the smallest details using basic python.

In a book that inspires us, David Perkins – a professor at Harvard University with a doctorate from MIT (Massachusetts Institute of Technology) in the field of artificial intelligence – calls the disease “elementitis” approach, in which a person does not begin complex work until he first learns all the elements separately. It is a lot like training with a ball without any idea of ​​baseball. Elements may seem boring or meaningless when you do not know how they fit into the big picture. And it’s quite difficult to stay motivated when you can’t even solve the problems that concern you or understand how the technical details fit into the whole. Perhaps that is why research has shown a decline in the internal motivation of schoolchildren, starting from 3rd to 8th grades (the only studied range of years).

Quality education gives you the opportunity to work on issues that interest you

If you can’t wait to determine if the plants are sick by photographs of their leaves, automatically generate knitting patterns, diagnose tuberculosis by X-ray, or understand when your raccoon is using your cat’s front door, we will teach you deep training to solve your problems as soon as possible (using the preliminary other people’s tasks), and then gradually delve into more detailed information. During the first thirty minutes of the first lesson, you will learn how to solve your own problems perfectly using deep learning! There is a dangerous misconception that deep learning requires computing resources and data sets the size of what Google has, but this is not the case.

Quality education does not require too much effort

Have you seen how Jeremy introduces advanced deep learning optimization techniques in Excel? Not yet? Then look (start at 4:50 min in the video) and come back. This is considered a complex topic, but after several weeks of hard work, Jeremy was able to simplify it to such an extent that it became as clear as possible. If you are really good at something, you can also give an accessible explanation, or maybe embed it in Excel! Intricate jargon and stupid technical definitions arise because of laziness or when the speaker is not sure of the essence of what he says and simply hides behind his peripheral knowledge.

Quality Education – Inclusive

It does not erect unnecessary barriers. You will not feel guilty if you do not code from the age of twelve, if you have an unconventional background, if you cannot buy a Mac, if you are working on an unusual problem or have not attended an elite educational institution. We want our course to be as accessible as possible. I am very concerned about the issue of inclusivity, I spent months studying and writing each of my popular articles with practical advice about how do we increase diversity in technology as well as a year and a half taught software development women. Currently deep learning is even more homogenouslythan the whole industry as a whole – scary enough for such a strong and efficient field. But we will change it.

Quality education motivates the study of basic technical concepts

Understanding the big picture provides you with a framework for posting the basics. The best motivation during the tedious parts of training is an understanding of what deep learning is capable of and how you can personally use it.

Perkins writes: “Playing basketball is more interesting than practicing with the ball, playing music is more interesting than learning scales, doing some kind of historical or mathematical research at the initial level is more interesting than remembering dates or solving problems.” Building a working model for the problem you are interested in is much more interesting than writing a proof (for most people!).

Quality education encourages mistakes

IN TED’s most popular performance Education expert Sir Ken Robinson claims that our school systems destroy the inherent creative abilities of children by publicly blaming mistakes. Robinson says: “He who is not ready to make mistakes, is not able to create, cannot think in an original way.”

Teaching deep learning using a code technique in interactive jupyter notebook – This is a great starting point in order to try new things, make mistakes and easily make changes.

Quality education leverages existing resources

No need to reinvent the wheel – good training materials already exist. Khan Academy for those who need to refresh the memory of matrix multiplication. If you are interested in X and want to know more information, we recommend that you read Y. Our goal is to help you achieve your goals in deep learning, and not be the only source for achieving these goals.

Quality education encourages creativity


Lockhart is sure that it’s better not to teach mathematics at all than to do it in such a distorted form that scares away most of its beauty. He describes mathematics as a “rich and exciting journey of the imagination” and defines it as the “art of explanation”, although it is rarely taught that way.

Your biggest breakthrough in deep learning will be the moment when you can apply it to the external areas of your expertise and the problems that concern you. This requires creativity.

Quality education teaches you to ask questions, not just answer them

Even those who, at first glance, get more advantages with the traditional teaching method, sometimes still suffer from it. For the most part, I studied with the traditional approach (even though I had several outstanding teachers at all levels of instruction, especially in Sworthmore). I excelled at school, passed the exams very well and generally enjoyed the learning process. I loved mathematics, which resulted in a doctorate in mathematics at Duke University. I did an excellent job with assignments and exams, but the traditional method did me a disservice when I began to prepare for my doctoral research and my professional career. Teachers no longer set clear objectives for me. I could no longer study every step block before starting work on the assignment. As Perkins wrote about his problems with finding a good topic for a dissertation, so I only knew how to solve the tasks set for me, but I could not independently search for what interests me. Now, I perceive my academic success as a weakness that I had to overcome professionally. I liked reading mathematical theorems and evidence at the very beginning of deep learning, but they did not help build deep learning models.

Quality Education Based on Evidence

We love data and the scientific method, and are interested in research-based approaches.

Interval repetition training is just one of these approaches, backed up by evidence, in which students periodically return to the topic before forgetting it. Jeremy himself used this method and achieved impressive results in chinese. The game learning method goes well with the interval repetition method – we return to the topics, each time delving into more and more low-level details, but always return to the big picture.


Learn the details of how to get a sought-after profession from scratch or Level Up in skills and salary by completing SkillFactory paid online courses:

Read more

  • The coolest Data Scientist does not waste time on statistics
  • How to Become a Data Scientist Without Online Courses
  • Sorting cheat sheet for Data Science
  • Data Science for the Humanities: What is Data
  • Steroid Data Scenario: Introducing Decision Intelligence

Similar Posts

Leave a Reply