What Is This

Welcome! We made this for you.

Why? Because there’s a flood of information out there on machine learning and it is woefully insufficient in two ways that we all recognize.

The first is that while there is a terrific amount of resources online for picking up machine learning in 24 hours, there is not nearly as much that treats this as a ten year endeavor. When you start thinking that way, you really want to consider the trunk of your tree foremost before each convolutional leaf.

The second is that there is no bridge from where we came from before, to where we are today. Without that, it is really hard to insightfully propose directions on where to go. For example, it is one thing to understand how an algorithm like TRPO works but all together another to grasp why the authors made the choices that they did and what they knew before that to influence those choices.

This public repository of depth-first study plans is the outcome of an effort to address those shortcomings in the machine learning research world. DFL was designed to be shared with and grown by the community. If you are interested in developing your own curriculum for a paper that we do not already host, we will help you! Check out Github for examples and open an issue detailing which paper you want to study. We want your help in growing this resource and making it richer.

We were inspired to create DFL to satisfy a missing part of other resources. Each of the following were north stars and are complementary to DFL:

  1. Textbooks: The de facto method of learning. We cannot match the depth that they cover. Instead, we aim to to create a curricula for a single modern paper.
  2. Online courses: Online courses are fantastic at reaching almost everyone in the world while having a similar profile to textbooks.
  3. Blogs: Blog posts are a copious resource in our community. They provide a counter to textbooks in that they frequently cover just a single paper or idea and can be a very welcome accompaniment to our own reading. In contrast to DFL, they generally are one-off and insufficient for deep understanding.
  4. distill.pub: Distill.pub has beautiful and thorough explanations for machine learning phenomena, oftentimes even with new insights about the underlying mathematics. Each topic may span multiple papers and cover a broader area of interest to machine learning practioners. In contrast, we try to unify resources needed to understand a single paper deeply by providing the right resources and topics from which to study.
  5. Metacademy: Metacademy is the closest parallel to Depth First Learning. We have similar goals in improving the learning process through breaking down concepts into their parents. However, we ultimately have different foci as DFL is built for understanding significant machine learning papers and consequentially increasing the strength of the field itself.