Spaced repetition can allow for infinite recall · EFAVDB
My friend Andrew is an advocate of the “spaced repetition” technique for
memorization of a great many facts [1]. The ideas behind this are two-fold:
-
When one first “learns” a new fact, it needs to be reviewed frequently in
order to not forget it. However, with each additional review, the fact can
be retained longer before a refresher is needed to maintain it in recall. -
Because of this, one can maintain a large, growing body of facts in recall
through daily review: Each day, one need only review for ten minutes or so,
covering a small number of facts. The facts included should be sampled from the
full library in a way that prefers newer entries, but that also sprinkles in
older facts often enough so that none are ever forgotten. Apps have been
written to intelligently take care of the sampling process for us.
Taking this framework as correct motivates questioning exactly how far it can
be pushed: Would an infinitely-long-lived, but forgetful person be able to
recall an infinite number of facts using this method? \(\ldots\) Below, we
show that the answer is: YES!
Proof:
We first posit that the number of days \(T\) that a fact can be retained before
it needs to be reviewed grows as a power-law in \(s\), the number of times it’s
been reviewed so far,
\begineqnarray \tag1\label1
T(s) \sim s^\gamma,
\endeqnarray
with \(\gamma > 0\). With this assumption, if \(N(t)\) facts are to be recalled
from \(t\) days ago, one can show that the amount of work needed today to retain
these will go like (see appendix for a proof of this line)
\begineqnarray\tag2\label2
w(t) \sim \fracN(t)t^\gamma / (\gamma + 1).
\endeqnarray
The total work needed today is then the sum of work needed for each past day’s facts,
\begineqnarray \tag3 \label3
W(total) = \int_1^\infty \fracN(t)t^\gamma / (\gamma + 1) dt.
\endeqnarray
Now, each day we only have a finite amount of time to study. However, the
above total work integral will diverge at large \(t\) unless it decays faster
than \(1/t\). To ensure this, we can limit the number of facts retained from
from \(t\) days ago to go as
\begineqnarray \tag4 \label4
N(t) \sim \frac1t^\epsilon \times \frac1t^1 / (\gamma + 1),
\endeqnarray
where \(\epsilon\) is some small, positive constant. Plugging (\ref4) into
(\ref3) shows that we are guaranteed a finite required study time each day.
However, after \(t\) days of study, the total number of facts retained scales as
\begineqnarray
N_total(t) &\sim & \int_1^t N(t) dt \\
&\sim & \int_0^t \frac1t^1 / (\gamma + 1) \\
&\sim & t^ \gamma / (\gamma + 1). \tag5 \label5
\endeqnarray
Because we assume that \(\gamma > 0\), this grows without bound over time,
eventually allowing for an infinitely large library.
We conclude that — though we can’t remember a fixed number of facts from each
day in the past using spaced repetition — we can ultimately recall an infinite
number of facts using this method. To do this only requires that we gradually
curate our previously-introduced facts so that the scaling (\ref4) holds at
all times.
Appendix: Proof of (2)
Recall that we assume \(N(s)\) facts have been reviewed exactly \(s\) times. On a
given day, the number of these that need to be reviewed then goes like
\begineqnarray \tagA1\labelA1
W(s) \sim \fracN(s)T(s).
\endeqnarray
where \(T(s)\) is given in (\ref1). This holds because each of the \(N(s)\)
facts that have been studied \(s\) times so far must be reviewed within \(T(s)\)
days, or one will be forgotten. During these \(T(s)\) days, each will move to
having been reviewed \(s+1\) times. Therefore,
\begineqnarray \tagA2 \labelA2
\fracdsdt &\sim & \frac1T(s)
\endeqnarray
Integrating this gives \(s\) as a function of \(t\),
\begineqnarray \tagA3 \labelA3
s \sim t^1 / (\gamma + 1)
\endeqnarray
Plugging this last line and (1) into (A1), we get (2).
References
[1] See Andrew’s blog post on spaced repetition
here.
Jonathan grew up in the midwest and then went to school at Caltech and UCLA. Following this, he did two postdocs, one at UCSB and one at UC Berkeley. His academic research focused primarily on applications of statistical mechanics, but his professional passion has always been in the mastering, development, and practical application of slick math methods/tools. He currently works as a data-scientist at Stitch Fix.