In Matter and Memory, philosopher
Henri Bergson had an interesting argument about the importance of memory for
perception. Briefly stated, he suggested that memory must be involved in every
act of perception, for if it wasn’t, every time we perceive something it would
be as if we were perceiving it for the first time. But if that was the case,
then every perception would amount to a new learning experience, and we clearly
don’t learn afresh each time we perceive. Of course, this argument does not
work for all sorts of reasons—including, for instance, the fact that perceiving
something for the first time and learning are two different things, or the fact
that often learning actually involves the repeated perception of encoded
material.
Nonetheless, there is an
important kernel of truth to Bergson’s reasoning. Somehow, when we perceive,
memory must be able to tell whether what we are perceiving is new or is old,
for if it is new chances are you need to encode it, and if it is old, you may
need to integrate it to information already stored. Following Marr’s (1971)
characterization, computational neuroscientists call the first process “pattern
separation”, referring to memory’s capacity to encode information in a way that
does not overlap with previously stored representations in order to prevent
interference and/or overwriting effects. In contrast, the second process, which
is known as “pattern completion”, refers to memory’s capacity to incorporate
newly acquired information into existing representations. What is remarkable
about these two processes, is that both appear to be dependent upon the same
brain structure: the hippocampus. How does the hippocampus manage to
differentiate between these two processes? One of the most promising hypothesis
suggests that the hippocampus can bias the processing either toward pattern
completion or pattern separation via neuromodulatory mechanisms (e.g., Hasselmo
et al, 1995) that are highly dependent on the encoding context.
In a recent paper published in Science, Duncan, Sadanand and Davachi (2012) tested a
prediction that follows from this hypothesis. But first, full disclosure: I
look forward to Lila Davachi’s papers with almost the same exhilarating
anticipation with which, I presume, fans of K.J. Rowlings must’ve awaited for
each volume in the Harry Potter series. Her studies are usually
methodologically impeccable, and her questions are, well, pretty great. Which
explains why I was so eager to read the Duncan et al. paper—and it didn’t
disappoint. This paper reports three experiments that capitalize on the fact
that the neuromodulators allegedly responsible for biasing the hippocampal
processing are relatively slow (Hasselmo and Fehlau, 2001). As a result, Duncan
et al hypothesized that “if switching between pattern completion and separation
biases is, in fact, mediated by hippocampal neuromodulatory input, it follows
that a processing bias should linger in time and, thus, influence subsequent
mnemonic processing” (p. 485). In other words, they conjectured that if such neuromodulators
are indeed responsible for switching between pattern completion and pattern
separation, then right after each switch takes place, the hippocampus must
remain slightly biased toward the kind of computational process it was just in.
So, in the first experiment,
participants studied a bunch of pictures of objects. Then, during retrieval,
they were presented with three kinds of pictures: studied objects, non-studied
objects, and objects that were similar but not identical to those studied.
Participants were asked to respond “old” if the object had been studied, “new”
if the object hadn’t been studied and was NOT similar to a studied object, and
“similar” if the object wasn’t studied but was similar to a studied object.
Accordingly, Duncan et al. hypothesized that if the memory system was biased
toward pattern completion, similar but non-identical pictures would be more
likely to be wrongly identified as old, since the memory system would be in the
“mood” for incorporating similar information into already stored memory
representations. However, if the memory system is biased toward pattern
separation, similar items would be more likely to be judged as new, as the
memory system would be in the “mood” for highlighting differences between the
perceived stimulus and stored representations. To test this idea, Duncan and
collaborators compared similar trials that were preceded by “new” responses
versus “old” responses, the idea being that during “old” responses the memory
system is biased toward pattern completion, whereas during “new” responses it
is biased toward pattern separation. And lo and behold, they found that if
similar trials were preceded by “new” responses there were fewer false alarms
that when they were preceded by “old” responses—which is really cool. Notice
that I said "responses" and not "trials" because the effect was evident not only when participants correctly identified
old objects as old and new objects as new, but also when they false alarmed,
suggesting that the subjective memory decision rather than response accuracy
was responsible for the processing bias—which is even cooler.
The second experiment was
identical to the first experiment except that this time Duncan and
collaborators wanted to see how long this bias—this “memory penumbra”, as they
called it—would last. So they varied the time between stimulus—the
“interstimulus time interval” or ISI—between 0.5, 1.5 and 2.5 seconds, and
found that the effect was time dependent, being only apparent with short but
not long ISIs—which is really the coolest. Remember how I mentioned that one of
the most promising hypothesis about the way in which the hippocampus shifts
from pattern completion to pattern separation rests on a bunch of possible
neuromodulators? Well, as you may imagine, there is some controversy as to
which precise neuromodulator may be responsible for the shifting, and one of
the reasons for this controversy is the varying temporal scales of the possible
candidates. So what I really like about this result is that the decaying time of
the effect coincides with the temporal scale of acetylcholine, one of the most
likely candidates for this sort of neuromodulation! So here you have it: a
beautiful piece of behavioral evidence, in humans, that simultaneously speaks
to a computational hypothesis and a neurobiological hypothesis about memory
retrieval. Pretty cool, ah?
The third experiment is slightly
different, and it requires a bit more background, so I’ll talk about it in a
subsequent post.
Duncan K., Sadanand
A., & Davachi L. (2012). Memory’s penumbra: Episodic memory decisions
induce lingering mnemonic biases. Science, 337(6093), 486-487.
