Struggles adjusting to FSRS

Anki FSRS
21

Probably so.

The first experience with FSRS went pretty badly. I left desired retention on 90% and only used Again when I completely forgot a word and couldn’t remember it.

That led to it setting intervals meant to let me completely forget 10% of my words.

FSRS wasn’t “wrong”, it was probably accurately modeling what would lead to me having a 10% chance of being completely unable to recall a word. I just asked it for the wrong thing, because a 10% chance of being unable to recall a word is definitely NOT what I wanted.

I probably should have asked for help with that sooner, because instead it led to a lot of frustration and loss of trust in the algorithm while I came up with different ways to fight it.

Now I just need to figure out if it’s possible to ask it for what I actually want.

Which is tricky, because to me “being able to remember the word” and “being able to remember the word instantly/effortlessly” feel like two extremely different things. So it is hard to trust that FSRS is going to be able to handle both when there is no way to tell it about both, just a single “Again” button and hope it figures out that I press it for two different reasons.

But pressing “Again” for both reasons is the one thing I haven’t tried yet, so I think it makes sense to set up a test deck, try it there for a while, and see what happens.

22

That is a very interesting point.

By the standard I had been using of “a lapse is a completely inability to remember a word”, I very rarely had lapses after the first few weeks with a card. I would preemptively shorten my intervals with “Hard” long before cards could get to that point.

I haven’t had many opportunities to see what FSRS does when a card with an interval of more than a week or two lapses, because I am in the habit of going out of my way to avoid that.

I’ll be curious to see how this goes with the test deck.

23

It will learn the difference pretty quckly in 2 ways:

  1. 2nd type of words will lead to more again presses.
  2. if you’ll use Easy button if you feel like the recall was too Easy and the algo could have waited longer - 1st type of words will lead to more Easy presses.

Further more - it will learn true difference, i.e. when in fact your memory crosses the boundary between the state you want and want you dont want. At the moment you just guess that a specific completely forgot card will be forgoten quicker than a specific not well enough card. This could be wrong for many expected and unexpected reasons, for example, because you feel more emotions and put more effort into remembering the c9mletely forgotten cards.

24

So far so good with the test deck.

I’ve noticed that for same-day reviews, “Hard” can decrease card stability (and therefore interval), which seems to be a nice way to tell FSRS the difference between “I can’t remember” and “I don’t remember as well as I wanted to” while cards are still growing to reach that “remember as well as I wanted to” state and earn their first “Good”.

Then once cards have reached that state, hitting “Again” when they fall below that state will feel justifiable.

It will take some time for the test deck to get any mature cards, but I played around with the FSRS visualizer and I like what I see there. It’s a clever algorithm, and that visualization of how it’s going to behave with different patterns of answers is helping to restore my trust in it (and to understand how/why things went wrong the first time).

25

Just point me to the message if you already answered this and I missed it, but I’m curious - have ever used FSRS for a decent amount of time, fully trusting it to do the right thing, and been unhappy with the results?

26

Yes, but there was some miscommunication with FSRS at the time that probably led to it producing the results I was unhappy with. (The fighting the algorithm didn’t start until after it was producing results I was unhappy with.)

The miscommunication was that I left desired retention set at 90% which was lower than my actual desired retention, and I didn’t press “Again” enough and was using “Hard” for “I got the word right but it was a struggle”.

I’m doing a new trial with a test deck now changing the way I use FSRS to see if that produces better results.

27

This is what I use Hard for personally. Again is for when I get it wrong or don’t know.

28

You may want to read the rest of the thread for a discussion of why I should have been using “Again” instead. Especially Danika’s post.

If a card was too difficult, “Hard” will still lengthen the interval (just not by as much). Unless it’s a same-day review, the only button that can shorten the interval is “Again”.

To appropriately communicate to FSRS that I was not remembering the card well enough, I needed to press “Again”.

29

To clarify, the only reason I gave that suggestion was because you decided that your goal was a shorter interval. That’s not general advice that anyone else would need to follow.

In general, @mason is correct – struggling to remember and remembering correctly is a successful outcome, and Hard would be the best grade to give there.

30

This is not always the case, sometimes “Hard” will shrink my interval by a little bit. But regardless, why does it need to be made smaller every time? If FSRS thinks it’s okay to make it longer, it probably is.

31

mason, I think this is solved for now.

I am trying things out in an FSRS test deck and will ask back here if I have more questions later.

32

See this page Setting up Anki | All Japanese All The Time | AJATT

The bottom line is that FSRS is a dead end. Even its creators have said so. The best solution would be to deprecate it and develop a new scheduling algorithm. I use Anki’s default algorithm, which is based on SM2. I don’t think FSRS will ever be “ready.”

33

Everything changes, and even Anki hasn’t been the same for a long time, and the article “December 28, 2020 — Tatsumoto Ren”

And don’t get me wrong, I’m not defending the “FSRS” algorithm. I see problems with every algorithm, just as there are always problems with theories, and practice is more complicated. But if one algorithm works better for the majority, then it’s chosen. Let the minority decide for themselves, and they can even write their own algorithm.

34

Still keeping at it with the FSRS test deck.

I’ve realized the reason I use “Hard” as an interval-shortening button in SM-2 is probably because SM-2 increases the intervals too quickly for me. And that even with SM-2, Hard as an interval-shortening button doesn’t work once I hit longer intervals, which is probably why my retention for mature cards isn’t as good as for young cards.

If FSRS can learn not to give me too-long intervals in the first place, I won’t need to worry about wanting “Hard” to work as an interval-shortening button.

I’ve also put the FSRS test deck in its own profile, and hidden the times shown above the answer buttons for that profile. That will make sure I give it a fair try and don’t try to fight the algorithm.

I have two test decks, one where I use “Again” if a word doesn’t feel strong enough in my mind, and another where I use “Hard” for that and reserve “Again” for words that I have actually forgotten.

There has been some unpleasantness along the way:

  1. With the default setting of one Relearning step, cards are shown once 10 minutes after I fail, then not again until the next day. Apparently that’s a built-in Anki behavior, but it was really causing problems for me. If I forget a card the day after I learn it, reviewing it once per day doesn’t make it stick. So then I tried removing the default 10m relearning step and letting FSRS manage relearning, but that seemed to cause Ease Hell++++ with cards being stuck at what felt like 1-4 hour scheduling for multiple days. That got bad enough that I started the test deck over with new cards and am trying not press Again quite so much this time around.
  2. I feel like I see some cards twice, don’t see them again for days, and then forget them. That’s very frustrating, but hopefully with time and parameter optimization it will stop doing that.

But hopefully test deck version 2 will go better.

35

For this reason, people set their own intervals until they actually learn a word (sentence, etc.). Here’s an example from this add-on: https://ankiweb.net/shared/info/1873164405

3m 10m 1h 6h 1d 3d 7d 21d

But this method isn’t for FSRS and it violates Anki’s logic!

Why is this so? I can’t say for sure, but our teachers say, “To remember something, you need to encounter it (encounter it, recall it) at least 5-7 times during your studies.” Others say that to form a habit, you need to do it frequently and for at least a month, and some even say 21 days. My opinion is that physiology is at work here. If you do it every time, at the same time, with the same incoming stimuli and associations, the brain remembers it and even salivates in advance (Pavlov’s experiment—for a dog to respond to a bell or light (a neutral stimulus) as if it were food, it took several days to several weeks of regular repetitions).

But I think it’s more complicated than that. It could be a simple word, “jduenaodkjk,” or it could be “fantastical,” but we can’t remember one, while we can remember the other because it’s already in our heads…or parts of it are already in our heads, that is, pre-existing connections. The algorithm can’t know this; it would have to check how difficult the word is for you, and how many words you can remember in a given time period. So, in reality, there can’t be the same amount of time and the same number of words for every person. We need to adapt to each person, and FSRS tries to do just that. Perhaps AI will do this later, just as a good teacher will initially assess your knowledge and memory and tailor the course to your needs.

But you know, FSRS tries to minimize time so you’ll remember a lot more, but in reality, if a person studies, they do it constantly, they don’t get tired because they’re interested. There, the number of repetitions is much greater, and they’re not really even repetitions; they’re about finding new connections between this knowledge and what you already know. Essentially, you’re strengthening the connection between this new knowledge and the likelihood of you forgetting it will be slim. And the purpose of our memory isn’t simply to remember—there’s such a word. A computer knows everything, but AI is smarter precisely because it’s not just important to know, it’s important to build all the connections and be able to create new connections based on them. But on exams, we often just have to say the answer to a question, so Anki can help, but that won’t lead you to become like Einstein, who said, “Imagination is more important than knowledge.” It’s all because you need to use your knowledge. You can memorize 1,000,000 words but not be able to combine them into sentences, or you can memorize 1,000 but be able to combine them in different ways to create meaning. Many here know this, and for this reason, they recommend studying words in sentences rather than individually.

Maybe you already know all this, but maybe someone else will think about it. It’s good that you’re searching… as the saying goes, “He who seeks will always find, and he who doesn’t want to seek will always find an excuse not to do anything.”

36

That’s an interesting take on it.

I have seen advice before to do learning outside of Anki and only use Anki for review. But that’s hard to do when working with a big premade deck.

I’m also reading in my learning language and doing an hour of conversation practice with a native speaker 1-2x per week, so Anki isn’t my only practice.

The conversation practice is part of where the frustration comes in, because words I “know” in Anki sometimes fail to come into my head when I need them in conversation. Especially words that have grown to have longer intervals. But I can’t practice my entire 3000+ word vocabulary with < 2 week intervals, I’d have no time to do anything else. Many words that grow to have longer intervals are still possible to remember in conversation, but there’s a sizable subset that aren’t.

That probably relates to the fact that my retention is lower for mature cards than for young cards, which I’m hoping FSRS will eventually help with.

But first I need to figure out how to get cards through their first weeks with FSRS, because “do the card twice and then fail it 8 days later” isn’t very fun. Hopefully with time and parameter optimization that will change.

37

Yep, that’s its job. It will learn that if that’s what your review history shows.

You can add another re/learning step if you think you would benefit from that. Or perhaps your re/learning step should be longer? Either of those would be a better choice than trying to game the algorithm by modifying your grading.

But the other obvious response to frequent lapses is to consider – what are you doing when you get a card wrong to give yourself a better chance to get it right the next time you see it? Re/learning is active, not passive. If you just click Again and hope for the best, you’re skipping an important step.

FSRS doesn’t have a model for near-term memory – so blanking out your steps isn’t recommended for most users – Reddit - Please wait for verification [and this still applies to FSRS-6].

The downside of your insistence on segregating FSRS from the rest of your collection is that it doesn’t have the chance to give you customized scheduling based on your own needs. Until you’re willing to optimize your parameters with a full, robust review history, FSRS might not be able to give you the results you’re hoping for.

38

Normally recalling a card after 10 mins is eneugh to recall it one sleep later.

If you cant most probably that means that the cards have too much information on them.
Even if you have one word on card it can be split into a few cards.

For example, one struggles with confusing letters e and i usually, they can make a card “translate blabla - whats the 3rd letter in the translation?”
or “blabla is blobl[…]blu”.
Or one struggles with even starting remembering the word, they can make a card “translate blabla - whats the first letter?” or “translate blabla - whats the root word?”

Concidering your wish to fast and easy recall this could be exactly what you want.

P.S. Some ppl make it 20mins test instead of 10mins - this could help.

39

That is a very good point. For actual lapses, I tend to say the word aloud a few times, but maybe I should be doing something more.

For the deck where I’m trying out clicking “Again” when I remembered a word but didn’t remember it easily enough, I am just counting on more frequent repeats to make it improve.

Ah-ha, good to know!! So multiple re-learning steps isn’t bad the way multiple learning steps is?

(Or maybe neither is bad as long as they stay under 1d in length?)

How many reviews does it take to give FSRS a full robust review history?

Part of the reason I’ve segregated it to a set of test cards is because I thought my existing review history (where I used Hard as an interval-shortening button with SM-2) would not be useful/correct input for FSRS. Especially because my past use of “Hard” has been kind of subjective and inconsistent.

So I set up two test decks (one with “Again” for “didn’t remember well enough” and one with “Hard” for that) with fresh new cards to try to give FSRS clean fresh consistent data not affected by my past strange usage patterns with SM-2. But if it’s going to take more than a few hundred cards to give FSRS enough data, I may need to change that plan.

40

I’ve been spending some time reading about spaced repetition and long-term retention, and I think I have been doing spaced repetition completely wrong.

When I talked about “not remembering a word well enough”, I was describing recalling with effort. That thing that apparently plays an important role in strengthening memories long-term. By attempting to keep my intervals short enough to prevent that, I have been totally shooting myself in the foot.

FSRS was right and I was wrong.

What FSRS does felt like strain the same way working out feels like strain, but apparently that’s as necessary for memory as it is for exercising. Doing 1000 reps of lifting a paperclip doesn’t make you stronger. (Which probably explains why I can’t remember passwords I used to type on a daily basis 2 years ago.) If I try to keep all words tip-of-my-tongue fresh memories at all times, I never get the benefit of having to make an effort to recall them.

I thought the point of stretching intervals out as long as possible was just so you didn’t have to spend as much time reviewing, I didn’t really grasp that the longer intervals were actually necessary and beneficial for strengthening the memory.

I still hate the idea of stretching reviews out so far I end up outright forgetting 10% of my words even temporarily, but if it’s necessary for building long-term recall then I guess I’ll just have to get used to it. (And I guess 90% desired retention probably is better for building long-term recall than 95%?)

OK, I’m an FSRS convert.

And I guess I’ll be sticking to using “Hard” when a card was hard to remember instead of switching to using “Again” so it will give me a shorter interval. (Unless maybe it was really hard to remember.)

The only question now is what to do about my strange learning history. Let FSRS try to make sense of it, or stick to the default parameters until I have enough learning history from doing spaced repetition properly with newer cards to that I can have FSRS optimize parameters while ignoring older cards?