I’ve been using Anki for about two weeks now for solidifying my knowledge of English vocabulary (yes, it’s my first language). I’ve already noticed a difference in my everyday speech.
I’d be interested in hearing your tips on learning list-like material.
For algorithms or math proofs, I typically put the whole algorithm on the card minus the current step, so something like this:
ElGamal:
Alice chooses: sk = x pk = g^x
Bob encrypts: C = (c1, c2) = (g^y, m * g^(xy))
Alice decrypts: s = [...] m = c2 / s = m * g^(xy) / g^(xy) = m
Anki’s cloze deletion makes this really easy. Just copy the full list, select one item, press F9, card done. Paste the list, select another item, next card done.
When you later try to reproduce the list (say, in an exam), you only need to remember one item. This item will work as your context from which you can retrieve the rest in either direction. This keeps the whole list connected in your memory.
Second, make those items really small. Again, it should contain one bit of information. How large this bit is depends on your proficiency. ;) For new algorithms, I might put each variable or operation in one step, but if it’s based on something I already know, I might just put “do a Dijkstra” in one step. You’ll know whether your steps are too large if you can’t put all of it into your working memory. Don’t worry about creating tons of cards. Reviews will be really fast and reinforce each other.
I also break down lists hierarchically if possible. Say you want to learn the Noble Eightfold Path, but all those items sound so similar. So you first group them into “Wisdom”, “Conduct” and “Concentration” and make a card for each. Then you break “Wisdom” into “Right Understanding” and “Right Intention”, learn that as a 2-item list. That way, each individual list is small and has a maximal contrast between items.
Third, and most importantly, chain backwards. I took this from Don’t Shoot the Dog. It only applies if the list has a fixed order.
I’ll just quote Karen Pryor on that:
Behavior chains break down and the behavior goes to pieces if there are unlearned behaviors in the chain, or behaviors that have not been brought under stimulus control [i.e. are directly associated with an automatic trigger]. You can’t reinforce the subject with a cue if it doesn’t recognize the cue, or if it cannot accomplish what the cue indicates. This means that behavior chains should be trained backward. Start with the last behavior in the chain; make sure it has been learned and that the signal to begin it is recognized; then train the next-to-last one, and so on. For example, in memorizing a poem, a piece of music, a speech, or lines in a play, if you divide the task up into, say, five sections, and memorize the sections in reverse order, starting with the last, you will always be going from weakness to strength, from the stuff you’re not quite sure of yet into the great, reinforcing, well-memorized stuff you know cold. Memorizing material in the order in which it is written and will be presented necessitates plowing continuously from familiar ground into the more difficult and unknown, a most unreinforcing experience. Treating memorization tasks as behavior chains not only shortens the needed memorizing time considerably, it also makes the whole experience more pleasant.
This really works. If you chain forwards, you always have to walk through steps you already know and when it’s time for the new one, you forgot it already. You are constantly unsure what to do next. But going backwards, learning the last step first, means you can directly practice the new step and add it to an established one. You will always know how to go on and can focus on the one new component.
Pryor gives the example of training a dog to fetch a Frisbee:
What people complain of is that their dog, when encouraged, will leap for the Frisbee and try to grab it as it is waved around, but when they throw it, the dog just stands there and watches it go. Or the dog chases and grabs it, but never brings it back.
There are two training problems in this game: [...] The second is that the game is a behavior chain: First the dog chases the Frisbee, then the dog catches the Frisbee, then the dog brings the Frisbee back for another throw. So each behavior must be
trained separately, and the last behavior in the chain, retrieving, must be trained first.
You can teach retrieving over very short distances—indoors, even—with something easy to hold: an old sock, maybe. [...]
When the dog will carry things to you on cue and give them up, you shape catching the Frisbee. First you get the dog all excited about the Frisbee, waving it around his face. You let him take it and have him give it back a few times, praising him madly for returning it, of course. Then you hold it in the air, let him have it when he leaps for it, and make him give it back. Then you toss it momentarily into the air and make a big
fuss when he catches it. When he has the idea, you can start shaping the first behavior of the chain, the chase, by tossing the Frisbee up and out from you a few feet so the dog has to move off after it, to catch it. And now you are on your way to having a great Frisbee dog.
This is a general rule: learn the behavior first, then add the cue. Otherwise you won’t associate cue and behavior (because you don’t know what to do) and won’t get anywhere.
Of course, your Anki cards will eventually become randomized in your reviews, so this applies only to adding new ones. It helps to occasionally just recite the whole list from scratch, too.
(Oh, and should you want to learn π, group digits into n-tuples (3<=n<=5 is typical) and assign a vivid mental image to each tuple. Then chain them as a simple story. And if someone asks you for π, stop reciting after 4 digits or everyone will know what a giant nerd you are.)
[From great-grandparent]: I
have a todo somewhere in my org files about writing up all my Anki experience
in a nice comprehensive article. One day I’m gonna get to it.
Gwern wrote the article I wanted to write. :) Everything worthwhile he doesn’t talk about I either already mentioned here or it’s very specific to a given purpose, like the use of morphological analysis to sort language cards.
Thanks. Gwern’s piece was a huge influence on me finally getting around to
using spaced repetition. Also influential:
variouspeople
on LW, including you, but a particular shout-out to
a couple
lukeprog
posts
that, while not directly relevant to my hardware, made me realize how quickly I
could take the first step.
Now I am better at doing arithmetic in my head and I know my immediate family’s
birthdays! (And more.)
I’ve been using Anki for about two weeks now for solidifying my knowledge of English vocabulary (yes, it’s my first language). I’ve already noticed a difference in my everyday speech.
I’d be interested in hearing your tips on learning list-like material.
There are three components that I found crucial to learning lists.
First, add plenty of context before and after the current item. For lyrics, about 2 lines before and after work well. A question looks like this:
(I wrote a simple python script to do this automatically.)
For algorithms or math proofs, I typically put the whole algorithm on the card minus the current step, so something like this:
Anki’s cloze deletion makes this really easy. Just copy the full list, select one item, press F9, card done. Paste the list, select another item, next card done.
When you later try to reproduce the list (say, in an exam), you only need to remember one item. This item will work as your context from which you can retrieve the rest in either direction. This keeps the whole list connected in your memory.
Second, make those items really small. Again, it should contain one bit of information. How large this bit is depends on your proficiency. ;) For new algorithms, I might put each variable or operation in one step, but if it’s based on something I already know, I might just put “do a Dijkstra” in one step. You’ll know whether your steps are too large if you can’t put all of it into your working memory. Don’t worry about creating tons of cards. Reviews will be really fast and reinforce each other.
I also break down lists hierarchically if possible. Say you want to learn the Noble Eightfold Path, but all those items sound so similar. So you first group them into “Wisdom”, “Conduct” and “Concentration” and make a card for each. Then you break “Wisdom” into “Right Understanding” and “Right Intention”, learn that as a 2-item list. That way, each individual list is small and has a maximal contrast between items.
Third, and most importantly, chain backwards. I took this from Don’t Shoot the Dog. It only applies if the list has a fixed order.
I’ll just quote Karen Pryor on that:
This really works. If you chain forwards, you always have to walk through steps you already know and when it’s time for the new one, you forgot it already. You are constantly unsure what to do next. But going backwards, learning the last step first, means you can directly practice the new step and add it to an established one. You will always know how to go on and can focus on the one new component.
Pryor gives the example of training a dog to fetch a Frisbee:
This is a general rule: learn the behavior first, then add the cue. Otherwise you won’t associate cue and behavior (because you don’t know what to do) and won’t get anywhere.
Of course, your Anki cards will eventually become randomized in your reviews, so this applies only to adding new ones. It helps to occasionally just recite the whole list from scratch, too.
(Oh, and should you want to learn π, group digits into n-tuples (3<=n<=5 is typical) and assign a vivid mental image to each tuple. Then chain them as a simple story. And if someone asks you for π, stop reciting after 4 digits or everyone will know what a giant nerd you are.)
John Conway on memorizing pi.
A Toothpaste for Dinner comic on how many digits of pi people have memorized.
That would be nice. I liked these comments a lot.
Gwern wrote the article I wanted to write. :) Everything worthwhile he doesn’t talk about I either already mentioned here or it’s very specific to a given purpose, like the use of morphological analysis to sort language cards.
Thanks. Gwern’s piece was a huge influence on me finally getting around to using spaced repetition. Also influential: various people on LW, including you, but a particular shout-out to a couple lukeprog posts that, while not directly relevant to my hardware, made me realize how quickly I could take the first step.
Now I am better at doing arithmetic in my head and I know my immediate family’s birthdays! (And more.)