Evaluate pedagogical tool#
A score for a learning resource relative to similar resources on the same topic. The score for a particular learning resource should theoretically be based on the slope of the Learning curve for the set of actions you want to able to do after finishing learning. A steeper slope would theoretically produce a higher score.
Value#
We often make massive time investments in learning from a particular topic; a textbook can take a year to cover in detail. If we can learn more efficiently, we can do more efficiently. Most of the time, we are evaluating learning resources to help us figure out how to answer a question.
Writing and organizing your own notes depends on an understanding of how to teach effectively, even if you are only communicating with your future self.
Cost#
Prefer a single language#
The major advantage of a textbook is that the author can define terms early on they use later in their material. In exchange for a major time investment, you get a (perhaps temporary) language to communicate in.
Read the table of contents. What percent of it are you going to read? It helps to read the whole book, because often authors make connections between different parts of the book. For example, they’ll promise to discuss something in a later chapter, or refer back to older chapters in explaining a new concept. If you don’t finish a book, you’ll invest a lot of time loading the first half of the book into memory and then throw it away when you leave the book.
Still, don’t feel you need to finish any textbook despite the major sunk cost you put into reading a large fraction of it; be flexible and constantly reevaluate the decision of whether to finish. You may discover while reading it (sometimes, before you start) a better learning resource or a better learning direction. Delete notes aggressively; delete the book aggressively (by not reading it).
Your own notes are a major learning resource that is already in a language you clearly understand. A textbook is the author’s notes. In your own notes you link between articles; the author of a traditional textbook is forced to link between sections.
Wikipedia is a major learning resource where, most of the time, the language is standardized across articles as the community of authors follows links and makes connections.
In summary, the major advantage and disadvantage of a paper book is that you feel like you need to read it once you start it. You sometimes don’t organize your notes while you’re working on it, because you need to stay on focus within the author’s world (rather than your own) to answer his questions efficiently. You’re afraid to get e.g. a topology book because you might focus too much on it.
Scale evaluation#
Scale your evaluation to the size of the learning resource. The number of words in a book is a good measure of how long it will take to read; you may also want to fudge your estimate based on an understanding of how “heavy” or new the material is.
Is there really any difference between an article and a book, besides that the latter is longer? You should be able to evaluate many static pedagogical resources the same way, just considering length as one factor.
Prefer primary sources#
This usually conflicts with a preference for newer tools, but regardless, an old primary resource is often-not-always better than a nearly as old non-primary resource.
Prefer multi-modal resources#
For a textbook, can you get an ebook version? Can you find a DRM-free ebook? Can you get a pdf? Can you get an audiobook version? If there are technical terms in the book, you’ll learn to pronounce them. Consider all the modalities used in ML; see also Perceiver: General Perception with Iterative Attention - 2103.03206.pdf.
Is there associated code? Should you start learning from the documentation or the code? You usually want to start from the documentation; the natural language documentation (usually) prefers the big picture view. Natural language is also often easier to parse since you spend the majority of your time in it (you almost always think in natural language). Natural language is significantly more common than artificial languages and more backwards-compatible; it will likely remain the same after programming languages are gone. Consider how quickly Statistical Rethinking | Richard McElreath has gone through different programming languages.
Still, you may move to the code earlier or initially if you’re more familiar with the programming language. Many concepts cannot be described or cannot be described precisely in natural language.
Prefer Jupyter notebooks, which avoid the Split attention effect when it comes to code and natural language. That is, for at least these two modalities the content is kept together. Tools like Sphinx are another way to keep these modalities together.
Can you get an offline (paper) version? To keep your own bookshelf empty, can you get it from a nearby library when you need the paper copy? It’s likely you’ll use it on a camping trip or vacation, then never touch it again. If you want the paper version again, you can get it from a library again. Once you finish reading a textbook, you rarely refer to anything but an electronic version again.
See also:
Learn from both primary and secondary sources#
Examples of the regular conflict between primary and secondary resources:
Jonathan Hui’s articles on RCNN models and the Faster-RCNN paper
Web articles usually include a date of publication. All else being equal (e.g. whether the source is primary or not), a somewhat reliable heuristic is that newer articles are better than older ones. Newer textbooks can refer to newer research, allowing you to skip over historical details. Would you start learning from a book where the source code was last edited years ago? The downside to these resources is that they are typically secondary (unless you are on the bleeding edge).
Learning from the original paper is going to be more valuable if you need to reference it; it’s better to reference the original than secondary sources because secondary sources add unnecessary indirection and potentially bias. On Wikipedia it’s still acceptable to cite secondary sources however; see Wikipedia:No original research - Primary, secondary, and tertiary sources.
Instead, prefer neither older nor newer pedagogical material; analyze the original with supplementary explanations (use them all, in focused work). Make up for bad explanations in one with better explanations in another. In the end, your internal representation will be richer if you learn from multiple sources.
If one resource doesn’t use language that is easy for you to consume, the other might. In the process, you’ll likely learn new language when e.g. one resource used more compact (unfamiliar) language and another used expanded language you understood. This is similar to how documentation makes code more understandable; either representation of the model should be acceptable (if both are complete) but it helps to see two versions. Most of the time, you’ll have a paper and Wikipedia to inspect in parallel; sometimes you’ll have SE questions, code, and alternative websites found in a web search.
In the same way that multiple resources can help you understand the same concept from multiple angles (use them all, not just one) all the answers to a SE question can be helpful. Give them all a plus or minus when you get invested in a question. Should SE questions be where you collect all your links, rather than Wikipedia? It’s almost like Wikipedia is just one high-quality answer in a list of possible answers. Multiple resources often include multiple modalities, which allows for zero-shot learning.
Migrate to a single learning focus#
How can you directly apply what you learn from this book? Prefer to focus on similar concepts at work that you are studying at home. Prefer to listen to the same audiobook you are reading, and read articles on the same topic the textbook you are studying is on. If you’re doing it right, you should see yourself taking notes in one area of your own notes at a time.
Most learning tasks should actually start from a TODo in your own notes, so you can expand (e.g. record answers to practice questions) or clean up your own notes as you work. What question are you answering? Will you be able to easily integrate the book’s mental functions into your own personal notes (mental functions)? Be aware of the Expertise reversal effect; if you are an expert already then reduced guidance often results in better performance than well-guided instruction. In general, if you pick too simple a tool to learn from you’ll also likely struggle to build knowledge efficiently for the simple reason that you’ll have to skim for material you don’t know, or you’ll relearn the same material in new words without ever connecting it to the words you already have (mental duplication).
Prefer interactive tools#
Can you find a video or browser game that covers the topic? Don’t assume you can’t for advanced topics; web search for e.g. “topology game” (don’t forget to test the plural in web searches, i.e. “games”) and you’ll find some interactive material. If you (physically) know someone else interested in the topic, you could buy a board game.
Is the material editable? Can you or the author easily fix issues you find? Ask the author a question. Does he/she respond at all? You almost always generate errata while reading a book. Collect these to provide public feedback to the author. Does the author ignore your suggestions, add them to a master errata, or provide a fixed version of the PDF or eBook you’re reading?
When you are a significant way through a book, contact the author with any other artifacts (e.g. chapter solutions) you generate so he can link to them. Why not help others? You may even get feedback from him/her.
Prefer copyleft licenses#
If the material comes with source code, it’s usually editable. However, you could copy and paste any web article (in e.g. natural language) and edit it. Is the content also have a permissive or copyleft license? If not, you won’t be able to share your modifications and will need to worry about fair use if you e.g. need to reproduce part of the material to review it.
It can be hard to tell if a website is based on open-source content. For example, even open source content based sites can require you to log in (though they are less likely to do so). Dive into Deep Learning looks like it isn’t at first, but you can see GitHub links at the top.
Look for “Terms” or “About” at the bottom of the page, most of the time. You can search for “opyright” in the Terms for information on how e.g. they own everything you submit to their platform. Read the terms before “signing” (in this case by reading the contents of the page, i.e. creating a mental link). Another option is to web search for their content license, i.e. “quora content license” for the website Quora.
If you can’t tell in one minute, block them. Clearly they’re not advertising their license, which they would if it was focused on their user.
Prefer free tools#
It’s difficult to refer to e.g. books behind a paywall even if you have access (these systems often have poor support for e.g. linking to a specific page). Your own readers will surely be blocked.
Books are often an example of the fallacy that the more you pay, the better a product you are getting. Because you can’t inspect a book completely without buying it (assuming it is not free), it’s often impossible to evaluate whether the book is good until after you’ve purchased it (unless someone writes a review). Who says online resources aren’t as good as books? You tend to not read long web articles because you think you should focus on focused learning in a book instead, but this is almost surely not the right decision in many cases. Free resources are almost surely getting more feedback from more people, simply because they’re free.
Language that has been tested by a large number of people is common; not standard. That is, it occurs in the mental networks of a large number of people because it has proved useful in the minds of a large number of people. For example, Koine Greek was well-tested because so many people spoke it. This is related to the The Cathedral and the Bazaar argument; there are more people in the bazaar because it’s free to get in. Open source code tends to be well-tested, if not in automatic tests then in tests by who really matters, people (and who can also provide some value feedback/tests).
Check for known issues#
Is there a public place for other readers or consumers to provide feedback? How many issues are reported on the code examples, and are they being resolved? See e.g. Issues · rmcelreath/rethinking. Is there an errata for published material? If there’s no errata, it’s likely there are a lot of issues. See Introduction.
Ask for an independent evaluation#
Who do you trust to tell you about the quality of a learning resource? A sharp coworker? Your professors? What shows up you web search for reviews? You can also find lower quality book reviews on e.g. Amazon.
Well-tested instructions#
Ideally, you would not like to be the first person reading through instructional material. If you are the first person reading through it, you can expect to discover typos, errors, and other issues. See Introduction.
How do you test, in general, that instructional material has been “peer-reviewed” by others? When we want to test that natural language is peer-reviewed, we test it. That is, peer-reviewed material is human-tested. If we ported (translated) the concept to an artificial programming language, then we can use computers to test as well.
If language is both computer-tested and human-tested, this will cause it to be minimal (not vice versa, for sure). If you’ve ever worked in a company’s proprietary codebase you’ve probably discovered large amounts of code that no one is using e.g. because only one person has ever seen the code (you’re probably the second).
This is a problem, because there’s a conflict between minimizing material and providing a thorough explanation. Everyone comes to new material with a different background; what unwritten dependencies does the material have? This is related to the “single language” requirement; a long textbook lets the author state all their dependencies. The more dependencies we cut out, the less useful material is for learning.
There’s a difference between learning dependencies (e.g. you need a middle school rather than an undergraduate education) and programming language dependencies, though. The first is a dependency between recursively formed mental states (what you know, versus what you will know after following the learning instructions) and the second is a dependency on what you need to perform a task you already know how to do.
Learning is not the process of doing some task, it is the process of learning to do some task. It literally involves the formation of new language; you should be acquiring new words when you learn (new definitions i.e. new words built on old words). It’s more analogous to writing new code, in a series of commits in your version control system, that rely on old code.
Said another way, your “peers” when you are learning are those who came to the same material without about the same level of background as you did when you arrived it. If every student came to the teacher with the same background, having taken the same e.g. courses, then teaching is easier. Of course, universities do this efficiently by providing a list of prerequisities and fitting students into a Cohort (educational group). The more diverse a class, the more “conflict resolution” the teacher will have to do to get them to the same place in the end.
In this analogy, instructional material is like a branch in version control history, but a more robust version of it. You test the material (branch) on students. Teachers provide the continuity between the continuity between versions of the branch that are applied on different students, because no two students come to the material with the exact same mental state. Teachers here are analogous to a developer who constantly takes the same code commit and rebases it on the same parent commit (or different parent commits) over and over. If you enjoy rebasing your code to make your reviewer’s life easier, perhaps you should be a teacher (better, don’t get so far ahead of your reviewers, or just fork the code i.e. create a long-lived branch).
For someone reading through a branch, certain series of commits are going to be better to help them understand how to go from what they know (commit A) to the end of the new branch. Said another way, instructional materials don’t throw away the history of how to get from A to Z; and because they are tested on many students, a single person (the teacher) has a good idea of what series of commits are best for bringing the student from mental model A to mental model Z.
In the same way, your own notes are a record of how to go from what you know you know (what you effectively have memorized) to a place that you’ve reached once but may not always be at. No one repeats what they already know they know in their notes; you don’t retell your future self how to do elementary algebra because you assume he has not forgotten it. Said another way, your own notes only have the dependencies you need to get from e.g. T to Z. If you’re already at T, then you don’t need someone to explain commits A to T.
Should technical instructors provide not only source code, but a history of how to edit source code to their students? Students would effectively “review” a long branch of commits, alternating between new problems and the solutions to them (worked examples).
You don’t want to maintain too many of your own notes for similar reasons to why you don’t want to maintain your own software; your explanation of how to remember how to go from T to Z is probably not as good as an instructor’s explanation and would require you writing a significant amount of material that essentially duplicates what your instructor wrote to teach you (drawings, etc.). Better, rely on others to provide materials that you take from your memorized state to the state you’ve been to once in the past, only appending your own notes on the instructor’s notes. To try to forget only what you want to forget, you should also aggressively clean up your notes and throw away instructional material you aren’t going to forget. For more details, see Organize notes.
To summarize, you’re looking for instructional material provided by someone in education, who has spent time actively working with modern students (similar to yourself) and seeing their struggles.