Helping to memorize information more efficiently using AI

Tübingen – Flashback to the days when one tried learning a second language. Whether adult or child – a person’s ability to remember those nouns, verbs and adjectives depends critically on the number of times the vocabulary is reviewed, the hours or days in between each review, as well as the time elapsed since the last repetition.

Deciphering the way human memory stores information has always been a fascinating topic for scientists. One primary method for successful memorization of information is known to be “spaced repetition” of content. Many empirical studies have been conducted to assess the appropriate spacing, the best intervals between repetitions to form an optimal strategy for the process of acquiring for instance a foreign language with minimum effort. These studies inspired flashcards, small pieces of information a learner repeatedly reviews, following a certain schedule.

Needless to say, the physical flashcard has long been replaced by the flashcard 2.0. In the past decade, an entire industry of e-learning platforms spewed out of the ground, providing spaced repetition course-schedules. However, most of these online learning providers use spaced repetition algorithms that are simple rule-based heuristics with a few hard-coded parameters. To give an example: a learner may be given a word to review in predetermined intervals of 2 days, then 4, then 8 – independent of the individual’s abilities and various difficulties of content under study.

Appropriate spacing: It´s all about the intervals between repetitions

The classical methods do not leverage the automated fine-grained monitoring and greater degree of control offered by modern online learning platforms, a group of scientists from the Max Planck Institute for Intelligent Systems in Tübingen and the Max Planck Institute for Software Systems in Kaiserslautern suggest in their recent work. According to this group, modern spaced repetition algorithms should be data-driven and adapt to the learner’s performance over time. Staying with the example given earlier, if the student is learning a difficult word he or she may be asked to review that content again in 1 day instead of 2 days while for an easier content the student might be asked to review it again in 3 days – all depending on the probability that this interval maximizes the probability that the learner recalls that content successfully with minimum effort.

In their work with the title "Enhancing human learning via spaced repetition optimization" published in the prestigious journal PNAS on Tuesday, lead author Behzad Tabibian together with Utkarsh Upadhyay, Abir De, Ali Zarezade, Bernhard Schölkopf and Manuel Gomez-Rodriguez developed a mathematical frame-work, that – given an existing machine learning model of human memory – gives the optimal solution on how to decide when is the best time to review a particular piece of content. At its core, their algorithm estimates when is the best time to review a word a student wants to permanently remember, all the while avoiding too many attempts, in order to make the learning process as efficient as possible.

To evaluate their scheduling algorithm, they relied on already existing datasets. “We used a dataset released by Duolingo (an online language learning app) a few years ago. It is a comprehensive data-set that contains fine grained records of learners’ practices over a period of two weeks”, Tabibian explains.

Avoiding unnecessary effort

The empirical evidence presented in this work suggests that spaced repetition algorithms derived using this new framework are superior to alternative heuristic approaches. “Our optimized spaced repetition algorithm, which we dubbed “Memorize”, helps learners remember more effectively, while also avoiding unnecessary effort, compared to learners who follow schedules determined by alternative heuristics”, says Tabibian and gives an example: “Let´s say you want to learn the German word ‘Brötchen’ (a bun, or bread role). Based on the experience of everybody who has tried learning this word, the algorithm estimates a baseline for how fast learners usually forget this word. When a person studies the word ‘Brötchen’ for the first time, the algorithm suggests that you should review this word in a day and a half. Based on whether you can recall the meaning of this word at that time, the algorithm suggests when would be the best time to review that word again and so forth. The algorithm decides differently based on the learner’s past experiences which makes this method both adaptive to individuals and specific to each individual piece of content.”

To sum it up: the research project is about estimating when is the best time to review. The tradeoff is not wanting to review something too many times too often, while at the same time wanting to practice something enough times with appropriate spacing so it is permanently remembered. “Our algorithm takes up these two prerequisites and comes up with an optimal learning strategy.”

About us

Behzad Tabibian is a Ph.D. student in the Empirical Inference department at the MPI for Intelligent Systems in Tübingen and a member of the Network Learning group at the MPI for Software Systems, co-advised by MPI-IS-Director Bernhard Schölkopf and MPI-SWS-Faculty Manuel Gomez-Rodriguez. Prior to joining the MPI he completed his Masters degree in the US and his Bachelor in the UK. The main focus of Tabibian's research is developing Machine Learning methods for information systems and large-scale web data. In his other projects during his Ph.D. he worked on tackling problems related to spread of misinformation on social networks using Machine Learning techniques.

Prof. Dr. Bernhard Schölkopf, born 1968, studied physics, mathematics and philosophy in Tübingen and London and received his doctorate in computer science from the TU Berlin in 1997. From 2001 to 2010, he was Director and scientific member of the Max Planck Institute for Biological Cybernetics in Tübingen. As one of the founding Directors, Schölkopf has been Director of the Max Planck Institute for Intelligent Systems in Tübingen since 2011. He has received numerous prizes, including the Max Planck Research Prize in 2011, the Academy Prize of the Berlin-Brandenburgische Akademie in 2012, the Milner Award of the Royal Society in Great Britain in 2014, the Leopoldina Award in 2016, the Association for Computing Machinery (ACM, the world's largest association of computer scientists based in New York City), the ACM Fellowship at the end of 2017, and the Leibniz Prize, Germany's most prestigious research prize, in 2018.

Schölkopf is also an honorary professor at the Institute of Mathematics and Physics at the University of Tübingen and an honorary professor of computer science at the Technical University (TU) Berlin. In December 2018 he was appointed affiliate professor for Empirical Inference at ETH Zurich.