Questions from Keybase user

[fasiha]

Hi Ahmed! 👍 👎 😂 😎 🎉 I hope this message finds you well, and I get a response from you 😅 I am extremely interested in your Ebisu project. I am an Anki user, and searching for better algorithms. I am here to request from you to explain in simple terms how the Ebisu algorithm works, if you don't mind. I am not a developer nor a mathematician so be easy on me please.

Another questions that I will be extremely grateful if you could reply to:

• How is Ebisu better than the SM18 in SuperMemo18 or SM2 in Anki?
• Where is Ebisu among the other algorithms that use AI to schedule?
• What are the best options right now to use Ebisu for learning?

Thanks a lot!

[fasiha]

Hi, thanks for writing! I'm going to answer here and send you a link to this in Keybase because I almost never check Keybase 😅.

I am here to request from you to explain in simple terms how the Ebisu algorithm works, if you don't mind.

In decreasing order of importance, here's how I describe how Ebisu works:

1. Ebisu applies some common statistical tools to model quizzes as coin flips with an unknown probability of heads or tails, and then uses your sequence of quiz results to estimate on that underlying hidden probability.
2. This is cool because you can ask "what is the probability that I've forgotten this card" for any card at any time. This is different than the SuperMemo family, which explicitly predict when to review each card: they don't think about odds, so the plugins to cram or space out reviews have to use more manual rules.
3. It works pretty well. I use Ebisu in all my apps, and it works reliably.
4. I'm working on a new approach, to combine the two techniques above: use statistics but also acknowledge that your memory strengthens with each review (i.e., it's not a static coin that you flip, the coin's weight changes as you study it).

How is Ebisu better than the SM18 in SuperMemo18 or SM2 in Anki?

As mentioned above, Ebisu really elegantly handles over-reviewing and under-reviewing. It lets you get away from scheduled quizzes, and this is really important to me because I don't have time to allocate time each day to answer all the cards Anki has scheduled for that day. On some days I review for hours, other days I have five minutes to review.

In terms of "is Ebisu better or worse than other algorithms for memory retention", I'm not sure, we don't have any evidence either way. #33 talks about benchmarking it against other algorithms but we haven't done it yet.

Where is Ebisu among the other algorithms that use AI to schedule?

Ebisu just uses basic Bayesian statistics, not AI or machine learning which use a large database of quizzes and results to predict when to schedule. Duolingo does it this way—#33 talks about this too—but machine learning training algorithms tend to have long runtimes, which I think will push them out of users' computers/phones and into the cloud where companies like Duolingo can do the analysis for you.

My personal thinking is this: when it comes to how well you learn something, SRS algorithm is one of the lesser factors. Much more important is the source, how deeply you engage the knowledge and the quality of your non-flashcard practice, etc.

What are the best options right now to use Ebisu for learning?

Some members of the community, Jacob and Arthur, worked very hard to package Ebisu into an add-on you can use with Anki: see the repo and the Reddit announcement for more details.

I'm afraid I don't use Anki but from the comments on Reddit, it looks like some people had success converting their decks to Ebisu. But note there seem to be lots of caveats: you need to use Anki on a desktop to run their add-on and it might corrupt your old reviews, so make sure you read the instructions.

Also note, if I understand it correctly: it's still Anki, so it's going to use Ebisu to schedule your cards in the future. Apps that use Ebisu more fully wouldn't schedule cards with due dates: such apps would just calculate which card is most in danger of being forgotten and present that to you when you go to review.

Let me know if this helps! I'm happy to answer more questions! From your questions so far though, I'd honestly recommend just sticking to the normal Anki workflow, making really good cards, studying energetically and deeply, surround yourself with the knowledge you're trying to learn, whether it's a language or medicine or whatever.

[Positron010]

Hey Ahmed! I am the Keybase user, and I'm here to thank you for replying to my humble questions, didn't expect that!

1. Following up on the implementation of Ebisu, during my humble research I found 3 options right now, with one being a WIP:

• meguro

I struggled a bit when I installed it, not being a developer, but the main issue with it for me right now, is that it uses a command line interface.

• Leaf

I was unsuccessful in running this program unfortunately, and will appreciate it if someone could help me because there is a sever version with GUI.

• Ebisu Flashcards

This project is apparently a WIP, and is not functional yet.

From these options, which option would you recommend to use @fasiha ?

2. Do you think that AI Spaced repetition is better than the non AI algorithms?

Thanks again!

Edit:

• Found another CLI application : memograph.

[fasiha]

From these options, which option would you recommend to use @fasiha ?

Hmm, I'm afraid I'm not sure, I wrote Meguro but the others I'm not familiar with at all. It might be the case there's no production-ready Ebisu-backed SRS yet 😞?

Do you think that AI Spaced repetition is better than the non AI algorithms?

Hmm, the way I personally think about this is, again, I think that, while just having an SRS is very important, the specific details about the SRS are not that important for learning a subject, being dwarfed in importance by good organization of the material, a lot of contact with the material outside of flashcards, etc. Specifically, I absolutely cannot use Duolingo: I tried it with French and Japanese, both languages I kind of knew, and found it was very disorganized and didn't work for me, so I have no idea if the machine-learning-heavy SRS helped or not.

Something to keep in mind is that, for Anki and Memrise and other general-purpose flashcard apps, the flashcards are totally opaque: the apps don't know about any relationships between cards or contents of the cards. For these, I would be surprised if machine learning spacing algorithms turned out to have an amazing advantage of SM2 or even Ebisu (in the scientific experiment sense of "advantage").

In contrast, when your flashcard app knows a lot about the underlying material (like Duolingo for languages, or like legal or medical teaching apps), then I wouldn't be surprised of machine learning SRS algorithms have a major advantage over SM2/Ebisu, because the SRS could see you failing one card, and infer that it should quiz you on similar cards in case you've forgotten the whole topic or not.

There's a ton of research that could be done on this topic, maybe in the next twenty years we'll make some progress 😄.

[fasiha]

Closing the issue but feel free to reopen or to ask more questions!