Mass Immersion Approach
Low-Key Anki: No Penalties or Boosting
What I am calling “Low-Key Anki” consists of two Anki add-ons which were created by ja-dark. The first of these two add-ons, “No Penalties or Boosting,” removes the ease factor changes that are normally applied to cards while reviewing. What this means is that, when using this add-on, no matter what grade you give a card (“again,” “hard,” “good,” “easy”), the ease factor will always remain unchanged; only the interval will be modified.
I explained in a prior section that because aiming for 100% retention is inefficient and unrealistic due to the probabilistic nature of the forgetting curve, effective SRS use consists of strategically forgetting a certain portion of your cards. Let’s call forgetting that occurs as a byproduct of intentionally aiming for a retention rate of less than 100%, “anticipated lapses.” Anticipated lapses occur even when the algorithm is functioning perfectly, and have nothing to do with inappropriate ease factors. On the other hand, lapses can also occur due to an inappropriately high ease factor causing the interval of a card to grow too quickly. Let’s call lapses that happen for this reason “ease factor lapses.”
In Anki’s default algorithm, anticipated lapses are, by definition, unjustly penalized in the form of ease factor reduction, hurting efficiency and wasting the time of the user (the ease factor problem). The No Penalties or Boosting add-on remedies this by wiping out ease factor reductions altogether. The potential problem with this is that now ease factor lapses will go untreated: because ease factors will never be changed from their starting value, if the intrinsic difficulty of a card is significantly lower or higher than the bulk of the other cards in the deck, it is going to suffer from intervals growing at a rate that is faster/slower than what is optimal. So by alleviating the unjust damage done to anticipated lapses, we are causing new damage in the form of nonoptimal interval growth in cards whose intrinsic difficulty diverge from the bulk of the deck.
Let’s consider whether this tradeoff is worth it. Really, it comes down to how heterogeneous is the intrinsic difficulty of the cards in the deck? If the intrinsic difficulty of the cards in a deck are largely homogeneous, then we can expect the benefit of using the No Penalties or Boosting add-on to greatly outweigh the cost. For example, in the case of sentence mining, if all of the sentences you learn are i+1, then we can expect that all of the cards in your deck will have very similar intrinsic difficulties. This means that nearly all lapses that occur will be anticipated lapses, AKA, due to either poor initial learning, a random brain fart, or a simple fluke of memory. By using the No Penalties or Boosting add-on, instead of having this fluke permanently stunt the growth of the card, you will be able to see the card again once, properly relearn it, and then have the card’s interval go back to growing at a normal rate.
But what about those minority of cards that are significantly more or less intrinsically difficult than the rest of the deck? Well, for cards that are relatively less intrinsically difficult, you will end up seeing them more often than what is optimal, but compared to how much time you will be saving not seeing cards that were anticipated lapses more often than what is optimal, what you are losing here is negligible. Furthermore, in the context of language learning, if a card is so easy that you only need to see it half as often as all your other cards, then you probably don’t really need that card in the first place (as immersion will retain your memory of the word/construction).
What about cards that are relatively more intrinsically difficult? We can expect that the interval of these cards will grow too quickly, leading to ease factor lapses. And because the ease factor won’t be adjusted, we can expect these cards to continue to lapse and give us trouble. Well, we can remedy this issue by setting a low leech threshold. In Anki, a “leech” is a card that has lapsed a given number of times. The idea with leeches is that if you fail the same card a significant number of times, something about that card probably needs to be changed. So, after a card becomes a leech, Anki will notify you of this and automatically tag and suspend the card, such that you can continue your current study session, and then properly handle the leech later on. By default, the
leech threshold, or the number of times a card must lapse before it becomes a leech, is eight, but this can be changed within the deck options. By setting the leech threshold to something low (like four), we can swiftly catch these intrinsically difficult cards and modify them to reduce the intrinsic difficulty. “When you fail cards, it’s commonly a matter of the quality of your memory encoding; that the interval was too long was a symptom of this, not the cause. Difficult cards are primarily a matter of quality control, not quantity.” To go back to the example of sentence mining, if a card becomes a leech, you can try finding a new sentence that contains the target word, finding a different definition for the target word, or adding audio. By modifying cards to lower the intrinsic difficulty in this way, we allow its interval to grow at the same rate as the other cards in the deck. After reformatting a leech, you may want to also reset the interval of the card in order to give it a fresh start, but I will explain this more in the next section.
Compare this process to what happens to a relatively intrinsically difficult card in Anki’s default algorithm: it gets graded “again” and “hard” a few times, the ease factor is reduced, and interval growth is slowed down enough for the content of the card to be kept in memory. So, instead of reformatting the knowledge contained in the card in a way that makes it easier to remember, allowing its intervals to go more quickly so that you can spend less time reviewing the card from that point onward, you see the card more often indefinitely. Clearly, the first option is more desirable. Now, there will also be times where no amount of reformatting is enough, and a card simply won’t stick. In cases like these, it’s best to just delete the card. According to the SuperMemo website, “Eliminating 10% of the most difficult items in a generic material may produce an increase in the speed of learning of up to 300%”; the time that stubborn leeches eat up simply isn’t worth the time they take away from learning other material.
So, it’s clear in my eyes that the No Penalties or Boosting add-on constitutes a significant upgrade to Anki’s algorithm, given that the intrinsic difficulty of the cards in a deck is largely homogeneous. But what about decks with more heterogeneous intrinsic difficulty? Although we can expect largely homogeneous intrinsic difficulty of cards when dealing with the domain of language learning, if you are using Anki to study a subject like math or science, perhaps the intrinsic difficulty of the different cards in your deck will be significantly heterogeneous. In cases like this, whether or not using the No Penalties or Boosting add-on will be beneficial is not as clear-cut. With significant intrinsic difficulty heterogeneity, the resulting amount of untreated ease factor lapses may lead to an unmanageable amount of leeches, a significant portion of which will simply not be re-workable in a way that lowers the intrinsic difficulty. In a case like this, the benefit of removing unjust penalties from anticipated lapses may not outweigh the cost of the increased number of ease factor lapses.
That said, because the benefit of using the No Penalties or Boosting add-on when the intrinsic difficulty of the cards in your deck is largely homogeneous is so great, what I would recommend to people with heterogeneous intrinsic difficulties in a deck is to attempt splitting that deck into multiple subdecks. For example, let’s say that we are studying Japanese and have one giant deck that contains a mix of kanji, sentence cards, and production cards. Although the intrinsic difficulty of all the kanji cards is likely to be similar to each other, the average intrinsic difficulty of a kanji card may be greater than the average intrinsic difficulty than a sentence card or production card. So if we split this one big heterogeneous deck into smaller subdecks (one for kanji, one for sentences, and one for production cards), we would end up with three relatively homogeneous decks, and become able to benefit from using the No Penalties or Boosting add-on, making our studies much more efficient.
Now at this point, you may be wondering: if we have three relatively homogeneous subdecks, but each of those decks have different average intrinsic difficulties, then how are the intervals going to grow at appropriate rates for each of the different subdecks after the feature of ease factors has been removed? This is where the Interval Modifier comes in. When using the No Penalties or Boosting add-on, instead of calibrating interval growth to match intrinsic difficulty on the basis of individual cards using ease factors, you use the Interval Modifier to calibrate interval growth on the basis of entire decks. The Interval Modifier is a feature within the deck options that allows you to modulate the speed at which the intervals of all the cards in a deck grow; it’s the deck version of ease factor.
By assigning each deck its own Option Group, we can independently adjust the Interval Modifier for each of our decks. By intermittently checking our retention rate and adjusting the Interval Modifier accordingly (making intervals grow faster for less reps with a lower retention rate, or making intervals grow slower for more reps with a higher retention rate), we can move towards our ideal retention rate over time. This way, intervals will grow at a pace that matches the mean intrinsic difficulty of the deck. This intermittent Interval Modifier modulation may sound like more work, but as I explained earlier, effective SRS use is about strategically deciding how much you want to forget; if you let Anki make this decision for you, the result is likely to be far from optimal. Anki is calibrated such that most people learning most things will have forgotten 10% of the cards that come up for review. Not only is it plausible that you won’t perfectly fall into the category of “most people most of the time,” but even if you do, in many circumstances, 90% is not the optimal retention rate to aim for (I will be covering optimal retention rates in the near future). You can also expect an add-on which will automate the process of Interval Modifier modulation to come out soon. Until then, I recommend this short article by eshapard for learning how to modulate the Interval Modifier.