Image Occlusion and the Problem of Excess Context in Flashcards

On AnkiWeb[.]net, I had a discussion with @sorata about the 20 Rules of Knowledge Formation (also known as Wozniak’s 20 rules). During our discussion, @sorata, the creator of the NEET UG deck, expressed a preference for creating large flashcards with multiple cloze deletions, arguing that having more context within a single flashcard is beneficial.

However, from my own experience with flashcards, I believe that too much context can be detrimental to efficient learning.

“Writing context is good, but providing too much context is bad.”

I have noticed that in the NEET UG deck, many flashcards contain multiple cloze deletions within large paragraphs. I believe this is a poor approach for the following reasons:

1. Memorization Instead of Learning

When a learner reviews a flashcard with multiple cloze deletions in a long paragraph, they tend to memorize the orderof the missing words rather than understanding the content. This creates a false sense of mastery.

2. Predictable Pattern Recognition

Because of the abundant surrounding context, the learner can see the lines above and below the cloze word. This makes it easy to recall the missing word not because of understanding, but simply because it is positioned between known words. Over time, the learner relies on this pattern rather than truly learning the material.

3. Interference with Spaced Repetition

In an ideal spaced repetition system, we should be reviewing a specific fact after a delay (e.g., 2 days → 7 days → 15 days, etc.). However, if a flashcard always exposes additional context (such as multiple cloze deletions in the same paragraph), we are effectively seeing other parts of the card too frequently, which disrupts the spaced repetition process.

4. The Problem with Occluding Lecture Slides for Cramming

Some students take entire lecture slides, occlude large chunks of text, and use them for cramming. This is a bad ideabecause:

• The occluded parts are still surrounded by too much context , making recall pattern-based rather than concept-based .

• Lecture slides are often dense with information, so occluding parts of them does not create atomic flashcards, leading to poor long-term retention.

• This approach is inefficient for spaced repetition, as the learner is not engaging with individual key facts but rather memorizing the slide layout itself.

A Better Approach: Optimizing Flashcard Design

To avoid these pitfalls, I recommend the following strategies:

1. Image Occlusion Flashcards

• Use the “Hide All, Guess One” approach rather than just hiding a single label at a time.

• This prevents the learner from seeing other parts of the answer, ensuring they focus only on the specific fact being tested.

2. Cloze Deletion Flashcards

• Instead of creating long paragraph cloze deletions, test one or two key facts per flashcard.

• Ensure each card is atomic, meaning it tests a single, well-defined piece of information.

• This improves recall without overwhelming the learner with too much information at once.

For context, we do not need a whole paragraph. A single word within the atomic flashcard can provide just enough context for proper retrieval without unnecessary distractions.

Example Deck and Further Discussion

I wish I could share an example deck with you to illustrate these concepts better. However, I am still learning how to post on Anki forums and use Discord effectively.

Additionally, I recommend checking out USMLE and MCAT decks to see examples of high-quality flashcards that effectively use minimal context while maximizing retention.

Let’s discuss—how do you structure your flashcards to balance context without excessive clues? Looking forward to your thoughts!

If you are talking of cards with lots of bullet points, most of these cards were created by Anubis Nehet back in 2023 and I’m afraid I can’t do anything to change them now. I also don’t think they’re terrible cards either. Anubis says they didn’t study much in 11th and only seriously started NEET study in 12th. The around 6 moths prep with this deck was enough to get a really good college with 660 in UG exams.

One of the things I provide with the cloze cards is a randomization feature. If you feel that, you can add alternative texts separated by a ¦. Other things include turning on the burying feature.

It’s a downside I guess. This was the motivation behind why the randomisation script was added. But nonetheless, I doubt it’s as serious in this deck as it could be in some other cases.

I think there’s some degree of accounting for this phenomenon when you’re using FSRS. If the RMSE is good, which it is in my case, I consider the effects negligible with scheduling.

To slightly elaborate, imagine with me what happens when this “disruption” happens. It slightly increases your chance of memorising the next slibling of card that you’ll review. This essentially affects the retrievability (R) of other cards. In any scheduling that’s targetting a specific target retention, the “disruption” thus reifies into a bad RMSE value.

I mostly agree with the other points you’ve bought up. Here, however, I think I’ll need to clarify that “context” meant helping people see the big picture.

If you have tons of tiny flashcards, it sometimes become hard to fit everything into a larger chunk of meaning. To give you a personal example, I some time ago was struggling with the picture of a angiosperm reproduction. I knew all the individual factoids, but somewhere was missing how the parts make up the whole. So, that’s what I was trying to mean with “context”.


I will point out a few more things while summarising everything I’ve said here:

  1. It has been pointed out by some people that in the long-term, pattern recognition plays less and less of a role in retrieval.
  2. I try to create seperate cards for things that don’t fit into a large card. Everytime I feel that’s not necessary, I let the cards be what they are.
  3. The surrounding context can provide unnecessary retrieval cue to people, but I’m not sure that’s always the case.
  4. RMSE is almost always a good way to check how good the scheduling is.

For whatever downside there is, these cards have helped a lot of people including me personally. I think a compromise has been made, I’ve tried to have less notes and more data in a single chunk which to be fair, isn’t what I normally do. But the cards Anubis made worked really well and I like them too. So, I didn’t think they’re particularly bad.


Addendum:

There is this thing that I requested: Provide a 'hide all, show one' option for cloze deletions? · Issue #3626 · ankitects/anki · GitHub

Hopefully, when it happens we both can find a middle ground. As you’ve pointed out, IO cards work better with “Hide all, Guess one”. Though not all the time, I think this also works best for normal cloze cards. That should certainly solve the problem for you (although I don’t think there is much problem for me at least).