Hacker News Reader

Comment by ekkeke | Hacker News Reader

Big fan of spaced repitition, especially for language learning. Unfortunately I feel like it fares worse for topics that require more application instead of memorisation, like mathematics or electrical engineering. Would love know if there was some super effective way to learn these similar to spaced repitiion.

So far, the only thing that really works for me is solving lots of problems until I have the technique mastered, but even then after a while I'm prone to forget how to solve them. Perhaps there some way to combine the problem solving with the spaced repition? It seems like it would be far harder to make a deck for this and I don't think most flashcard software handles it very well.

View on HN

I've been experimenting with "spaced free recall". So first, I'll read a section of a textbook. Then, I write down everything I can remember about it in a blank text file, organizing things in a way that makes sense to me. Next, look back at the section and compare to my recalled notes, filling in missing information and committing extra attention to missed spots. Repeat the process with increasing intervals between reviews.

From what I understand of the literature, free recall produces better learning compared to cued recall like flash cards. Part of the reason is that it forces you to organize information and associate it with existing knowledge.

Anecdotally, it's much easier to learn conceptual knowledge, and I don't really feel like my recall of specific facts has suffered compared to traditional SRS.

I actually used Anki cards to study LeetCode problems when preparing for interviews and it seemed to help. After doing a problem and solving it I created the card as such:

- Front of card is the entire LC problem statement

- Back is a bulleted list of the steps or key points (ie. first I notice this list is unsorted, so I would sort first, next I would do blah blah..)

- Back also contains the code solution that I might just glance through or look at a particular part of it.

I also benefitted a bunch from using Anki for LC problems -- I described the details in https://news.ycombinator.com/item?id=35517232

Maybe try drawing the key points instead of text cards. Idea sparked by the below, which is awesome but requires someone else who already understands to create the learning material first.

"Each 5-minute video, or 'cartoon', is the equivalent of 50 minutes of a university-level computer graphics class. ... there was no statistically significant difference in learning effectiveness between [cartoons & lectures] as measured by exam, homework, and project scores. In other words, the cartoons were just as effective as traditional classrooms for teaching the material."

https://g5m.cs.washington.edu/

https://youtube.com/playlist?list=PLWfDJ5nla8UpwShx-lzLJqcp5...

In my own experience using spaced repetition for math: math has both semantic and procedural knowledge. The procedural knowledge comes from doing problems and rewriting proofs. But the semantic knowledge is also important, and you can acquire and retain this through spaced repetition.

I was going to write some rules specifically about math but I might write those as a separate post because they got too long. I think I've benefited specially from memorizing the proofs of theorems, though refactoring proofs into multiple lemmas to make each proof small enough to fit in a flashcard is a tedious process.

"refactoring proofs into multiple lemmas to make each proof small enough to fit in a flashcard is a tedious process."

Can GPT/chatGPT help here ? If yes, how ?

I haven't tried it. But it's a two step process:

1. Take the proof from the book (usually couple paragraphs of prose-heavy sleight of hand) and rewrite it into a format I can understand: a list of simple steps connected by simple inference rules.

2. Split them up until each proof is 5-7 steps.

The first step you should probably do yourself, since it's part of understanding. The second step GPT can probably help with.

I use it and it's quite effective. I just paste text I want to summarise and just ask (GPT-4) to "Create Anki cards for these paragraphs. Keep the answers brief". It does quite a good job in distilling the knowledge.

And for cards creation in general, the ever-green "20 rules of formulating knowledge in learning" is always a good guide.

http://super-memory.com/articles/20rules.htm

On top of that, you can prompt it with the 20 rules so that it generates cards which would conform to the rules.

> require more application instead of memorisation, like mathematics or electrical engineering

I’ve dreamed of having some app that mixes in bite sized learning lessons with otherwise “fun” internet (social media, news, etc.)

I could imagine it could give you a little tutorial and then ask you a quiz (to force application). If you get it wrong it keeps you at the same concept and explains it a different way next time, maybe asks if you want to revisit prereqs.

Even if you can’t memorize the answers, you can change your understanding and intuition.

The concept of "Kata" seems to be a popular repetitive method for learning/practicing programming skills: https://docs.codewars.com/concepts/kata/

You can use SRS to schedule the review of problems you've understood how to solve.

Front of card: where to find the problem (e.g., book, page number, problem number).

Back of card: where to find a solution (e.g., solution manual, page number, maybe a personal notebook with cleanly written solutions, etc.).

I initially tried writing up the problem and solution in Anki, but that was too much of a hassle and realistically I'm not gonna be reviewing problems without the book in front of me anyway.

Same, I put links to online geography quizzes on the front and then record my times on the back (editing each time). I just put them in a separate deck I only do at the computer. I'm gonna add Leetcode links too I think.

General advice for spaced repetition is to make flashcards atomic i.e. as small as possible, as in the OP, but general advice for language learning is to always learn words in context instead of on it's own, for example in example sentences. Have you figured out a solution combining those two goals?

For language you might be interested in the Clozemaster[0] approach. Basically, you are shown a sentence, both in English and the language you want to learn, and one of the words in either one is a cloze deletion, e.g.:

    English: there are thirty days in April.

    French: il y a trente ___ en avril

And you have to complete the cloze with "jours".

The sentences are compiled automatically from Tatoeba[1], the cloze deletion is done on the least-common word[2]. This combines vocabulary with grammar.

I didn't like the Clozemaster UI so I wrote a script to make the clozes myself: https://borretti.me/article/building-diy-clozemaster

But automatic approaches are not great. Later I asked GPT-4 to make these flashcards for me, that gave me much better/more meaningful results.

[0]: https://www.clozemaster.com/

[1]: https://tatoeba.org/en/

[2]: https://www.clozemaster.com/faq#how-are-the-blanks-in-the-se...

Could you detail a bit more your gpt4 usage for language learning?

I was wanting to get back into French and thought about using chatgpt, but I'm worried about it's hallucinations and teaching me wrong.

I asked it to list an outline for a French course, then for each item in the outline I asked it to make a table of English-French sentence pairs of increasing complexity.

This is very nice.

For language I've always used the sentence in target language (the one i want to learn) in the front of the card and the translated sentence in the back of the card but I've always wondered if it should actually be the other way around.

Your suggestion with the cloze is another good approach

> I've always wondered if it should actually be the other way around

It should be both ways round. This is especially true for languages (where your brain needs e.g. French -> English for reading/listening and English -> French for writing/speaking). It's also useful even where you only need one direction because learning both directions actually strengthens the memory for the direction you need.

This is a common problem. My preferred solution is to quiz myself on that specific word, then see the word being used in a context with example sentence(s). That could be extra info on back of the card. While it is right to make flashcards atomic, one might misunderstand that so as to not include information that doesn't directly play a role in Question -> Answer.

Simply spoken, get questioned on the word alone, then see it in context. I've found that sufficient to solve this problem.

As an alternative, you can question yourself on a sentence and the word by its own. Note that sentence alone wouldn't cut as you'd memorize the sentence and not the word and would be unable to remember the word otherwise, most likely.

Could you save “representative problems” to your cards? Eg a particular integral that uses a particular method etc.

This has not been effective for me. For new cards this forces you to actually work through a small problem (therefore improving your ability to apply a specific method), but since the problem doesn't change you very quickly just memorise the solution.

On further reflection I found that those cards just were an attempt at avoiding actually practicing something, but this isn't really possible. If you want to be good at solving integrals, than you do actually have to solve lots of integrals. Anki will not make you proficient at this if you don't also put in the time to frequently solve integrals. Instead what it can do for your is keep the various techniques for integral solving close to the surface, so that you can relearn them much quicker if you haven't solved any integrals in months. You skip the step of having to rediscover all the techniques.

Is Duolingo basically spaced repetition for language learning?

Duolingo's spaced repetition is poor to non-existent. The whole point of spaced repetition is to prompted to remember information at the right time, but Duolingo relies on you to decide what to do and when. In the new Duolingo layout there is some pseudo-SRS in that the lessons are ordered such that concepts will be presented a few times with increasing gaps between them, but it's still on you to decide whether and when to do new lessons, and once you've done all the repetitions of the lesson you will not see it again unless you decide to go back and revisit it (in which case there is no help to decide when to go back to it).

Personally, I dislike having every app implement (or not) it's own version of SRS, so I combine language apps with Anki. For each lesson in an app, I make an Anki card which simply tells me to revisit that lesson. I then put those cards in a special deck with customised settings with larger review gaps so that I'm not overwhelmed with time-consuming lessons.

Reply on news.ycombinator.com