So, SRS algorithms are such that the more you try to get the optimality, the more the returns diminish because there's significant variability in the human population (bias-variance tradeoff sort of applies). The Ebbinghaus curve itself is a very crude approximation.
(I'm a language learner and an active user of SRS)
The notion of an outdated algorithm corresponding to worse or better results to me is neither here nor there. There's evidence that the repetition interval itself is not the critical factor in retention. Empirically we know there are things that we remember with 0 repetitions (the vivid details of a traumatic event), while other things we struggle to remember even with SRS when there's no emotional connection to what is being memorized.
When you're doing self-hacking, you have to understand you're working on a highly variable, complex adaptive system that defies optimality and exactness. Much of what is known are approximations.
I’m in strong agreement with your points. It always surprises me when I read people trying to come up with a more optimal algorithm for testing one’s retention when there are so many factors out of one’s control that would influence the forgetting curve. For instance, the notion of judging your ability to recall something and providing it as input to Anki is such a gross approximation. Choosing the wrong value from 1 to 5 would affect your recall schedule considerably.
In my opinion, what’s more important to a recall strategy is just pure habit. You have to practice regularly, and anything that makes the software easier to use for the long term is something that would have a greater effect than the optimality of the recall algorithm. (Incidentally this is why I’m working on a flash card app that is easier to use, to help ensure one comes back to it regularly.)
Have you tried SuperMemo? Intervals matter because they correlate with rep loads. The more accurate the intervals, the lower daily rep loads can be.
The SuperMemo creator has spent 3 decades and collected plenty of data, it's hard to imagine that he wouldn't be able to improve on an algorithm from 1990.
SRS isn't my main method for learning -- I use a variety of other methods. I only use SRS for (independent and disconnected) atomic-type knowledge that I need to drill. My rep loads are generally not onerous, so it's not a problem that necessarily affects me too much.
(Example: I learn through cognitive dissonance so I use it primarily to deprogram certain associations in my head. For instance, the word "oficina" looks like "office" to my English brain, but the word actually means "mechanic's workshop" in Portuguese. Or "pretender", which looks like pretend in English, actually means "intend". SRS helps me drill the Portuguese meaning into my skull. I use SRS to memorize bits of knowledge that don't fit anywhere and that are hard to create a context for.)
I have looked into SuperMemo in the past because I was attracted to the prospect of a general pro-level memorization tool. I use complex pro tools all the time and was not fazed by the learning curve, but I felt the UI was, let's say, overfitted to the creator's personal preferences.
It was a purely emotional decision but I knew I wouldn't use it regularly, and the best SRS tool is one that one would use over one that one would not.
> SRS isn't my main method for learning -- I use a variety of other methods. I only use SRS for (independent and disconnected) atomic-type knowledge that I need to drill. My rep loads are generally not onerous, so it's not a problem that necessarily affects me too much.
Agree. Also agree that actually using something is more important than how good it is. I just strongly object to:
>outdated algorithm corresponding to worse or better results to me is neither here nor there. There's evidence that the repetition interval itself is not the critical factor in retention.
claiming that algorithm does not matter, especially from people without experience actually using both.
>(Example: I learn through cognitive dissonance so I use it primarily to deprogram certain associations in my head. For instance, the word "oficina" looks like "office" to my English brain, but the word actually means "mechanic's workshop" in Portuguese. Or "pretender", which looks like pretend in English, actually means "intend". SRS helps me drill the Portuguese meaning into my skull. I use SRS to memorize bits of knowledge that don't fit anywhere and that are hard to create a context for.)
>I have looked into SuperMemo in the past because I was attracted to the prospect of a general pro-level memorization tool. I use complex pro tools all the time and was not fazed by the learning curve, but I felt the UI was, let's say, overfitted to the creator's personal preferences.
It was a purely emotional decision but I knew I wouldn't use it regularly, and the best SRS tool is one that one would use over one that one would not.
That is understandable. I will say: I personally like the interface. To anyone reading this, I might seem insane but for a high-level user it just clicks and works (with few quirks here and there). Likely because I lack imagination I can't imagine how I'd make it look nicer while still maximizing usability.
It is of course terrible to share with newcomers. But I've taught many people SuperMemo, I think the main issue with the UI for most people is just that you don't understand it. Once you get over that I think it's not so bad.
> claiming that algorithm does not matter, especially from people without experience actually using both.
Here's the basis of that claim [1].
That said, I have not tested it empirically myself so I think that is cause for moderating my claim -- I place a high value on empirically-tested knowledge. Take my claim as partially withdrawn.
But to make the argument more precise, an algorithm producing a lower rep load because it uses intervals more efficiently does not imply it produces a significant delta-difference in retention. It merely means the algorithm lets you fit more stuff into a memorization schedule. That seems to be to be quite different from improving retention itself of say a smaller inventory of facts. Just wanted to be sure about the actual claim that is made.
Based on a skim of the paper, this experiment seems to only plan out how individuals study material on the same day, not distributed over a period of time that lasts multiple days.
In order to test how much of the material was retained in long-term memory, they tested the subjects on the material again after one week.
Spaced Repetition software (anki, mnemosyne, and supermemo) will distribute the learning across multiple days regardless of the algorithm they use. I don't believe the paper cited is using the same testing method that SRS systems use.
Piotr Wozniak (the creator of Supermemo) has collaborated on a couple of papers regarding this topic. I have not found many researchers who are studying it and creating algorithms in quite the same way.
There are a few papers cited in a blog post / self-published supermemo wiki here [1] that discuss a proposed two-component model of memory that explain in greater detail why learning on increasing intervals of time (days, not in the structure of repetitions on the same day) can lead to a higher stability of memory.
Yeah, that's definitely a problem.
I will say though, if you use a VM you can get a not too bad experience on both mac and linux (I've tried both and found it usable enough for daily use).
Additionally SM is much more feature rich once you get into incremental reading and fully optimizing your learning experience.
Sure the UI isn't a nice flashy Electron.js app, but SM is a titan compared the terminate that is Anki.