Dr Piotr Wozniak, November 2014

What is incremental reading?

Incremental reading is a technique that helps you convert texts, e.g. from the web, into life-long knowledge. While reading incrementally, the student extracts the most important portions of texts and converts them with a click to questions that are later used to test student's memory along the spaced repetition scheme. This is how the acquired knowledge can be retained for life.

What is the fastest reading method?

To a casual observer, incremental reading looks like ordinary reading combined with lots of extra computer operations. In other words, it looks slow and cumbersome. To a newbie, the impression of sluggishness may be amplified by a never-ending battle with complexities of software (i.e. SuperMemo), difficult article selection choices, prioritization woes, and lack of experience in choosing the right materials for extracts or question keywords.

In this text, I want to prove that no reading can beat the speed of incremental reading. I also want to demonstrate that the speed does not need to come at the cost of comprehension or recall. Let's make it even more radical by saying bluntly:

Now that I got your attention, let me add some qualifiers that will explain that this claim is true but not as rosy as it seems at first. I would like to tell you a story in which incremental reading went into a contest with classical speed-reading. I will include some hard-earned data to illustrate my points.

Why can incremental reading be very fast?

Incremental reading can be very fast mostly due to the fact that you do not need to worry about neither thorough comprehension nor recall during your first pass through the text. This contrasts it with speed-reading where comprehension is a measure of reading quality and maximum focus is essential for speed-reading to make sense. In other words, incremental reading can be employed as skimming without a speed limit, while comprehension and recall become a concern at later time. At the very extreme end of the reading speed, you can import a Wikipedia article, automatically split it into sections, fish for a few sections of maximum interest and set their priorities high for further processing in the incremental learning process. All that should take mere seconds even if the article is a megabyte long! Nothing can beat that speed. Naturally, your initial inflow of new knowledge to your memory will be microscopic. Perhaps just a few general ideas on the subject.

For example, if you wanted to know what Biosphere 2 is, you might execute a speed-learning procedure that would take you just 2 minutes, give you a general understanding of the Biosphere project, help you illustrate it with 2-4 pictures, and prioritize the knowledge for future consumption. With hundreds of things to learn, very often, such a speedy rough outline is all that you need at the moment of first reading. In incremental learning, all well-prioritized pieces of information will compete for your attention. Their understanding and recall will depend on how you manage the learning process and how you shape your learning strategy. It is possible you will never go beyond the ABC of Biosphere 2. Perhaps you will even forget it. However, if the subject is important, you can incrementally dig into individual pieces of the imported text and ensure excellent recall for as long as you are ready to invest in repetitions with SuperMemo.

Speed-reading contest

When a friend from America, Matt, paid me a visit in January 2006, he boasted of his improving speed-reading skills and his 2000 words per minute record. He was ready to give me a demo. I pulled a book on American History from a shelf and picked a test chapter to process. He wheezed through the text fast enough for me to be pretty certain that his recall of facts would be rather superficial. However, his comprehension was remarkable. We did not do any measurement to confirm his 2000 wpm capability, but it was definitely the fastest reading performance I have ever witnessed personally. One detail needs to be mentioned though. Matt's knowledge of history is a bit too good to make for a reliable test. For example, he included Alabama when listing out a group of southern states, while Alabama was not actually mentioned in the text. Secondly, the book was a fun read that was not too rich in hard-to-recall or hard-to-process facts. I asked Matt to take on a harder text from my own SuperMemo collection, where we could measure the exact speed in a controlled environment. We found an exemplary article about sleep that included a dozen of sleep drug names. Slightly tired due to the late hour, Matt admitted that the text was much harder and his reading speed crawled down to 200 wpm, i.e. a tenth of his best achievement. His comprehension was also worse than in the case of American History, partly due to his lesser knowledge of the subject matter. Matt's story shows that speed-reading techniques can really help one devour books at the speed of light. However, when it comes to harder fact-rich texts, even the best speed-reading will slow down to make it possible to combat complexity and volleys of facts.

Next day, I thought of giving the same text a test in SuperMemo with my own incremental reading skills. As mentioned earlier, there is virtually no limit on the speed of reading in incremental reading. You can skim and extract long portions of text without sacrificing comprehension in the long run. Those skimmed texts will simply come back in the learning process later. Perhaps in a day or in a week. Matt needed maximum focus and 7 min. 20 sec. to rush through the article. I knew that I can process it faster with no concern about focus and possible distraction by rambling thoughts. In other words, the first big advantage of incremental reading is that it is entirely stress free (at least for a pro). If you miss details today, you can recover them later. Reading without the pressure of focus is fun, and this, in turn, paradoxically, improves focus! I managed to rush through the text in 2.5 minutes. During that first reading pass, I generated 13 extracts and ignored little or no text. My comprehension was ok, but the recall of details was pretty poor. In other words, in a hurry, I can easily beat a speed-reader as long as I defer the actual acquisition of knowledge. The best part about fast reading with incremental reading is the freedom, lack of stress, and 100% guarantee that not a single detail will be missed unless I choose so with my own rational decision. Obviously, a good speed-reader will do better in incremental reading than a typical user of SuperMemo without speed-reading training. Many years of incremental reading experience are a form of a speed-reading training substitute. I am not sure how effective a substitute though.

The value of speed with limited comprehension

Your question might be now "What's the point of speed without comprehension?". In your first pass, you usually focus on (1) prioritizing the material and (2) finding golden nuggets of knowledge. With good speed you can read a hundred articles in a day rather than just a few. On the next day, you can start slowly with the most important finds of your speed-reading blitz. More often than not, you will interleave high speed with slow reading depending on your needs:

Incremental reading: 8 years later

After my speed test, I decided to repeat it on the same text on the same day. This time I had to read 13 extracts generated in the first pass. The total text size would be 35% larger due to the fact that extracts often include context that repeats between individual elements. This second pass took 5.6 minutes. This means that my two first passes of the article took only 11% more time than Matt needed to process the article in a single pass. The best news came from the fact that my comprehension and recall increased ten-fold. With an 8 minute investment, I did not employ any solid measure of comprehension. So my estimates were based on guesswork. My mastery of facts was pretty solid, almost certainly better than Matt's, however, sleep is one of my favorite subjects, so the battlefield was pretty uneven. After the two passes, we had double good news: (1) first pass was very fast, and (2) second pass delivered solid comprehension.

Intrigued with the results I decided to keep repeating the reading test. However, for a good comparison with incremental reading, we need to involve the actual test of knowledge retention. In SuperMemo, this means the need to generate cloze deletions. For that reason, I decided to repeat the test in increasing intervals that would give active knowledge carried by SuperMemo items a chance to solidify for long-term retention (along the principles of spaced repetition). As the original goal of the test was to measure the speed of reading, I could not let the natural incremental learning process take care of the extracts generated from the article. The goal was to read all texts in a single pass on a single day for each review. For that reasons, I gave the article and all its extracts lowest priority possible to ensure minimum interference from my daily learning. I then returned to the whole body of generated knowledge yet five times in the years 2006, 2007, 2008, 2010 and 2014. At each pass, I would randomize the elements to eliminate the sequence effect. Naturally, items generated from a single article do interfere with each other and falsify the retention measurements. However, in SuperMemo, one can choose his or her own desired retention level and the retention was not part of the experiment. We can safely say that in SuperMemo, whatever the extent of the testing time, the retention is always excellent. My measurements also produced excellent retention (94%) that was lowered by an arbitrary choice of intervals and increased by item interference.

All incremental readers know that the cost of processing long texts tends to increase over time once the student gets to the details of individual paragraphs and starts generating cloze deletions. However, because of the usual learning overload, the full processing cycle of a single article is not too visible and not too obvious to an average user. Overtime, the topics tend to disappear entirely from the process. The cost of topic processing has a bell-shaped nature over time. The height of the cost curve peak depends on one's meticulousness, size of the text, workload, text's factual richness, etc. The full processing time may extend to lifetime, but it can also take just days or weeks, esp. for high priority articles, short articles, sparse articles, easy articles, etc. Once topics disappear from the process, the decline in the cost of item review is theoretically exponential. In other words, the cost of knowledge acquisition in incremental reading, for a single text, is a superposition of a bell-shape and exponential curves with a short-lasting increase in workload and a gradual dissipation of costs over time.

My little experiment was to plot the cost curve in terms of processing time. Naturally, the superposition of topics and items was pretty artificial as I had to focus on speed-reading and topic processing, while items were reviewed on the experimental schedule largely inconsistent with spaced repetition. My item review proceeded in arbitrary intervals that did not reflect the actual nature of learning with SuperMemo.

Now that I consumed the entire text and deleted all topics, from November 2014 on, I will focus on proper item repetitions. I will report on the future progress, perhaps in a decade, however, anyone who knows SuperMemo knows that the outcome is pretty clear. I should be able to recall 95-98% of the information when tested at any time of the day in any moment of the future. Naturally, this will happen only if I opt to keep all the article information in high priority knowledge ranks, which I plan to do for the sake of the experiment.

In the meantime, I asked Matt how much he remembered of the article we read together 8 years ago. I was not surprised that he even had problems with recalling the subject matter. He remembered reading American History though! This only shows that we can recall past events as long as they are prominent enough. Perhaps Matt remembered how impressed we all were with his performance?

RESULTS

Size of texts

In incremental reading, the total size of texts in topics expresses the reading that still needs to be done for full comprehension. The total size of texts in items expresses the total size of knowledge acquired for life. Those two numbers keep changing in the learning process where items constantly feed on and consume the texts included in topics.

The total size of texts in topics will keep increasing for a while. The increase is caused by retaining context carried by some paragraphs. It may take 3-6 reviews for that increase to plateau. After that, there is a rapid decline in the size of topics as they get converted to items and gradually deleted (with Done in incremental reading). The size of items keeps increasing as long as new items are generated. After that, the size of items will tend to shrink due to the simplification process in re-formulating items and ditching redundant information.

[imagelink]

Cost of reading and learning

In incremental reading, the cost of learning in time is initially dominated by reading, however, after a few reviews, the time devoted to answering questions carried by items becomes dominant. What cannot be seen in the graph, however, is the fact that this cost drops exponentially for well-formulated items. Once the article is fully processed, the cost of review will drop to negligible levels after just a few months. The total 8-years cost of review in the presented experiment amounted to 60 minutes. That's much more than Matt's 7 minutes at 200 wpm. However, if you compare recall-to-cost ratio, the superiority of incremental reading becomes obvious.

[imagelink]

Element count

The number of topics will increase at first, but will ultimately drop to zero. Items, on the other hand, are on an increase in the period of topic processing, however, their impact on the cost of review is mitigated by increasing review intervals in spaced repetition.

[imagelink]

What are the bottlenecks of speed-reading?

My primary interest revolves around reading for creative purposes. It is not how fast you can read fiction without missing on the plot. I am more interested in the bottlenecks in the creative process where an individual uses reading as a source of information or creative inspiration to generate value: new ideas or long-term knowledge. This angle of interest undermines the significance of the speed of reading due to the fact that true bottlenecks in the creative reading process are: (1) processing information, and (2) long-term retention of knowledge.

In creative reading, once you hit upon an insightful piece of information, you definitely do not want to rush ahead and load on more information that might interfere with the just soaked inspiration. Just the opposite, to let the mind wander is the key principle of creativity. Many parallel subconscious pathways are activated in the brain. Given sufficient freedom to roam, those pathways will bob up to consciousness with new ideas that can change the world. That subconscious creative information processing is hardly under our rational control. Its speed is roughly pre-set and rushing it might have the same effect as giving a chess player a short deadline for a complex move. Rushing in with more information is like blowing more wind into the clouds that are just forming a meaningful structure in the sky. The inflow of information must be carefully controlled. More speed on barren mind, less speed in a fertile moment. More speed on low-quality texts, less speed on high-quality texts. For more see: The 1000-Word Dash.

The second bottleneck, the long-term retention of information requires spaced repetition, which in turn is executed best along the minimum information principle. This means that attentive and meticulous approach to formulating knowledge for active recall will dramatically cut the lifetime cost of retention.

In speed-reading, there is a never-ending trade-off between speed, comprehension and retention. In incremental reading, you can rationally control all components of the creative reading process: speed, comprehension, creativity and long-term retention. In that light I must reiterate the original claim:

Conclusions

• incremental reading can be substantially faster than speed-reading, however, for an inexperienced reader, this is largely accomplished with skimming at the cost of comprehension and recall
• despite the high speed of reading, over time, the comprehension can quickly be brought up to 100% (or more precisely, to the maximum, which may depend on student's knowledge and article's quality)
• in the long run, the retention of knowledge will approach 96% by default (different levels can be programmed in SuperMemo in cost-vs-retention trade-off). This 96% level can be maintained for lifetime at low review cost
• the time cost of incremental reading increases initially due to the proliferation of topic extracts and due to the process of generating new cloze deletions. The cost curve over time is a superposition of a bell-shaped curve (topic proliferation), linear inflow of new items, and exponential decline in the cost of item review
• once all topics are fully processed, topic costs drop to zero. Depending on student's strategy, this may happen in just one review, or never. In the presented example, costs of topic processing dropped to zero after 7th pass through the text
• in the presented example, the cost of topic processing and the size of topic texts peaked in 5th review (15 min and 18 kB respectively, up from the original reading time of 2.5 min in the first review for 9 kB text)
• once topics are processed, cost of item review becomes negligible within a few months
• speed-reading training is a great introduction to fast reading with incremental reading, however, long-term incremental reading should also help develop natural speed-reading skills
• the lifetime cost of incremental reading may be ten times the cost of speed-reading, but the dismal recall record of speed-reading makes the difference in actual knowledge vast and growing with each month
• incremental reading may be difficult to master, however, it should be a dream tool for speed-readers who care about long-term retention of knowledge