The Machinery of Thought - Working Memory

Studies of the brains of monkeys and, more recently, of humans are=20
revealing the neural underpinnings of working memory, one of the mind's=20
most crucial functions by Tim Beardsley, staff writer
_________________________________________________________________________=
____


In a darkened basement laboratory on the campus of the National=20
Institutes of Health in Bethesda, Md., volunteers earn $100 by lying for=20
two hours with their head inside a huge magnetic resonance imaging (MRI)
 machine while they gaze at a screen reflected in a mirror. The screen=20
periodically displays black-and-white pictures: some are faces, others=20
scrambled blocks of light and shade. When a face appears on the screen,=20
the subject signals by pressing buttons whether the face is a new one or=20
the same as one that was shown a few seconds earlier as a "target" to be=20
remembered.=20
As the test proceeds, the MRI machine bombards the volunteer's brain=20
with radio-frequency waves that excite hydrogen atoms in the=20
bloodstream, causing the atoms to emit signals of their own. Later, the=20
machine transforms the resulting electromagnetic cacophony into=20
color-coded maps of oxygen consumption levels throughout the subject's=20
brain. Because increased oxygen consumption results from heightened=20
neural activity, researchers can analyze these brain maps to learn what=20
parts of the brain work hardest when a person recognizes a face.=20

With experiments such as these, researchers are beginning to fathom the=20
neural processes underlying "working memory"--the limited, short-term=20
store of currently relevant information that we draw on when we=20
comprehend a sentence, follow a previously decided plan of action or=20
remember a telephone number. When we bring to mind the name of Russia's=20
president, for instance, that information is temporarily copied from=20
long-term memory into working memory.=20

Psychological studies have demonstrated that working memory is=20
fundamental to the human ability to reason and make judgments that rely=20
on remembered contextual information. There are compelling humanitarian=20
reasons for understanding working memory. Schizophrenia, one of the most=20
devastating mental illnesses, is believed to be caused in part by a=20
defect of this system. Studies of the molecular basis of working memory=20
"have implications for drug treatment in mental illness," says Patricia=20
Goldman-Rakic of Yale University, one of the most prominent=20
investigators of working memory.=20

An intensive research effort has started to produce detailed information=20
about the areas of the brain involved when we engage this vital=20
intellectual faculty and is illuminating the patterns of neural activity=20
that allow it to operate. The important role of specific brain chemicals
 in working memory is also becoming clear. Yet for all the progress,=20
researchers have still to agree on how working memory is controlled and=20
organized.=20

>From Electrodes to Fast MRI=20

The prototypical test for working memory involves what is called delayed=20
choice. An animal or a person signals where some specific cue was=20
previously seen, before an imposed period of waiting. Thus, a monkey=20
might be given a choice of two jars in separate positions and be=20
rewarded for pointing to the one in which it previously saw food placed.=20


The task provides no clue to the correct response at the time of=20
testing, so the monkey must rely on its recollection of the correct=20
location. A related challenge rewards an animal for remembering which of=20
several images it saw presented initially as a target. The NIH=20
volunteers who were recalling faces were engaged in a variant of this=20
test.=20

Technological advances have greatly enhanced researchers' ability to=20
probe the neural underpinnings of such capacities. Investigators began=20
studying cerebral activity in working memory some 40 years ago by=20
inserting electrodes into individual neurons within the brains of=20
monkeys. This method has its limits, however. Although monkey brains
 have clear anatomical similarities to human brains, the animals'=20
behavior is vastly simpler, making detailed comparisons with human=20
thinking problematic. Lacking language, the animals must be patiently=20
trained over a period of weeks to master tasks that a person would pick=20
up in a minute.=20

Electrode-recording techniques are also ethically unacceptable for use=20
on people. Researchers try to learn which parts of our species' brain do=20
what by studying the effects of damage caused by injury, disease or=20
therapeutic surgery. Yet patients have different medical histories--and=20
their brains vary in exact shape--so interpreting this clinical data is=20
tricky at best.=20

Earlier this decade, positron emission tomography, or PET scanning, made=20
enormous strides by showing which parts of the human brain are busiest=20
when performing different tasks, such as hearing words or speaking. But=20
PET requires exposing the human subjects to radioactive tracers, and to=20
keep radiation doses within acceptable levels, researchers have to use=20
techniques that can resolve brain areas only about a centimeter apart.=20
Also, during a delayed-choice task, PET scans are too slow to=20
distinguish between the neural activity pattern of a target being held=20
in mind and the pattern that follows a few seconds later when the target=20
is recognized.=20

The new technique used at NIH and elsewhere, called functional MRI, can=20
resolve the position of active neurons to about two millimeters and is=20
fast enough to study activity before and after the brain recognizes a=20
cue on a screen. The rapidly improving technique has over the past two=20
years become the state of the art for functional brain imaging.=20

Monkey Puzzle=20

Experiments involving electrodes implanted in monkeys still provide=20
crucial information, however, because they reveal in fine detail and on=20
a millisecond-by-millisecond timetable what happens as these primates=20
respond to cues and rewards. When animals perform such feats of working=20
memory, several brain regions can play a role, but as Joach=92n M. Fuster=
=20
of the University of California at Los Angeles showed in the 1970s, one=20
area that is always involved is the prefrontal cortex.=20

The prefrontal cortex is a layer of tissue that lies just behind the=20
forehead. With neural connections to almost all the areas of the brain=20
that process sensory information, it is well situated to maintain a=20
flexible store of information relevant to any task at hand. It is also=20
the part of the brain that has grown the most in humans, as compared=20
with monkeys. Monkeys missing some parts of their prefrontal cortex=20
preserve their long-term memory but perform miserably on delayed-choice=20
tests. Humans similarly afflicted suffer a reduced attention span and=20
ability to plan.=20

Fuster and, separately, Kisou Kubota and Hiroaki Niki of the Kyoto=20
Primate Center made electrical recordings from a variety of neurons in=20
the monkey prefrontal cortex, including some that apparently were active=20
only while the animals were holding information in working memory.=20
Subsequently, Goldman-Rakic and her colleagues have explored working=20
memory in monkeys with more sophisticated tests. They established that=20
prefrontal neural activity during a delayed-choice task indeed=20
corresponds well to the functioning of working memory.=20

Goldman-Rakic and her associate Graham Williams have taken the analysis=20
all the way to the subcellular level, showing that receptors for the=20
neurotransmitter dopamine pivotally influence the responsiveness of=20
cells in the prefrontal cortex and their actions in working memory.=20
"There is no other example I know" of research that spans the gulf=20
between behavior and subcellular function, Goldman-Rakic notes. She and=20
her colleagues have recently shown that administering antischizophrenic=20
drugs to monkeys for six months leads to specific changes in the numbers=20
of two different types of dopamine receptors in that region, further=20
evidence that schizophrenia--or its treatment--alters normal function=20
there.=20

Research by other scientists supports the view that the prefrontal=20
cortex could sustain working memory. Robert Desimone of the National=20
Institute of Mental Health, along with Earl K. Miller, Cynthia Erickson=20
and others, has discovered in the monkey's prefrontal cortex neurons
 that fire at different rates during the delayed-choice task, depending=20
on the target the animal saw previously. Neurons in other parts of the=20
brain generally "forget" the target when a distracting stimulus=20
appears--their rate of firing changes. Prefrontal neurons detected by=20
Desimone and his colleagues, in contrast, maintain their rate of=20
activity during a delayed-choice task even after the animal is presented=20
with irrelevant, distracting stimuli.=20

Activity in some prefrontal neurons, then, appears to embody directly=20
the temporary working memory of the appearance of a target the animal is=20
seeking. Other researchers have found prefrontal neurons that seem to=20
maintain locations in working memory: Giuseppe Di Pellegrino of the=20
University of Bologna and Steven Wise of the National Institute of=20
Mental Health have found prefrontal neurons that are busiest when an=20
animal has to remember where it saw a cue. Stimuli fail to excite the=20
same frenzy unless they are in the location that is the current target=20
for the task.=20

Neurons in the prefrontal cortex could thus apparently control how=20
animals respond in a delayed-choice task. Fuster, one of the pioneers in=20
the field, says the prefrontal cortex "serves the overarching function=20
of the temporal organization of behavior" by driving networks that=20
maintain currently important information in an active state. And neurons=20
in the prefrontal cortex might exert their influence in more subtle=20
ways, too.=20

Besides controlling directly the responses in delayed-choice tests,=20
Desimone believes, the prefrontal cortex might tune the visual and=20
possibly other perceptual systems to the task at hand. "What's loaded=20
into working memory goes back to sensory processing," he suggests.=20
Hundreds of experiments with both animals and people have shown that=20
organisms are far more likely to perceive and react to cues relevant to=20
their current needs than to irrelevant stimuli. This effect explains why=20
we are more likely to notice the aroma wafting from a neighbor's grill=20
when we are hungry than just after eating. If Desimone is right, the=20
prefrontal cortex could be responsible for focusing an animal's=20
attention and thus possibly steering awareness.=20

Imaging studies with PET and functional MRI corroborate the evidence=20
from brain injuries that the human prefrontal cortex, like that of=20
monkeys, is central to working memory. Several research groups have now=20
imaged activity in the prefrontal cortex when people remember things=20
from moment to moment. Different tasks may also require various other=20
brain regions closer to the back of the head, but for primates in=20
general, the prefrontal cortex always seems to be busy when target=20
information is kept "in mind."=20

The Devil in the Details=20

Having shown that the prefrontal cortex is crucial to working memory,=20
investigators naturally want to understand its internal structure.=20
Goldman-Rakic and her associates at Yale have found evidence that when=20
an animal retains information about a spatial location, the prefrontal=20
activity is confined to a specific subregion. A separate area below it=20
is most active when an animal is remembering the appearance of an=20
object. These findings, together with observations of the anatomy of=20
neural pathways, led Goldman-Rakic to propose that the prefrontal cortex=20
is organized into regions that temporarily store information about=20
different sensory domains: one for the domain of spatial cues, one for=20
cues relating to an object's appearance and perhaps others for various=20
types of cues.=20

There are, moreover, some indications that the human prefrontal cortex=20
may be organized along similar domain-specific lines. A PET study=20
reported last year by Susan M. Courtney, Leslie G. Ungerleider and their=20
colleagues at the National Institute of Mental Health found that in=20
humans, as in the monkeys studied earlier by Goldman-Rakic, certain=20
brain areas are especially active during exercises that challenge=20
working memory for visual details and for locations. Moreover, the most=20
active brain regions lie in similar relative positions in both species.=20

Goldman-Rakic's proposal about the organization of the prefrontal cortex=20
argues against the standard view of the various components of working=20
memory. The British psychologist Alan Baddely proposed in 1974 that=20
working memory has a hierarchical structure, in which an "executive=20
system" in the prefrontal cortex allocates processing resources to=20
separate "slave" buffers for verbal and spatial information. The memory=20
buffers were supposed to be well behind the prefrontal cortex. But=20
Goldman-Rakic is unconvinced that the brain's executive processes are=20
confined to any particular location. Moreover, in the traditional model,=20
memories organized by domain would lie somewhere behind the prefrontal=20
cortex, not within it.=20

The high-speed imaging capability of functional MRI is now able to help=20
resolve the question. A study that Courtney and Ungerleider and their=20
colleagues published in April in Nature pinpoints the part of the brain=20
that is liveliest while working memory holds an image of a face. That=20
region--the middle part of the prefrontal cortex--has been fingered as=20
the crux of working memory in a variety of studies.=20

Yet the face-recognition task Courtney and company used does not involve=20
any obviously executive functions, Ungerleider notes. Their findings=20
thus contradict the view that only executive functions reside within the=20
prefrontal cortex, but they do fit with Goldman-Rakic's scheme.=20
Similarly, Jonathan D. Cohen of Carnegie Mellon University and his=20
co-workers found a region of the prefrontal cortex partly overlapping=20
the one identified by Courtney that is active while subjects remember=20
letters seen in a sequence. The more the subjects had to remember in the=20
Cohen experiment, the more active their prefrontal regions. So Cohen's=20
result also suggests that working memories are actually stored, in part,=20
in the prefrontal cortex. Domain-specific organization "is the dominant=20
view" of the prefrontal cortex, Wise says.=20

Wise himself does not subscribe to that dominant view, however. He=20
points, for example, to a study reported in Science in May by Miller and=20
his associates at the Massachusetts Institute of Technology. The=20
researchers recorded from neurons in the prefrontal cortex of monkeys=20
while they solved delayed-choice tasks that required them to remember=20
information about both the appearance and spatial locations of objects.=20
Over half the neurons from which Miller recorded were sensitive to both=20
attributes, a result not expected if domain-specific organization=20
prevails. "It argues against Goldman-Rakic's view that identity and=20
location are processed in different parts of the prefrontal cortex,"=20
Miller says.=20

Goldman-Rakic responds that she and her colleagues have recently found=20
hundreds of cells in part of the prefrontal cortex that respond=20
selectively even in untrained animals to objects or faces--further=20
evidence, she asserts, that the information in that area is organized in=20
part by sensory domain. "We do feel the evidence is overwhelming that=20
the functions of neurons in the prefrontal cortex are dictated in large=20
part by the neurons' sensory inputs," she says. Moreover, Goldman-Rakic=20
believes technical problems cast doubt on Miller's experiment. She=20
maintains the targets he used were too close to the center of the visual=20
field, which could produce spurious firings.=20

Keeping Self-Control=20

Michael Petrides of McGill University, another leading figure in the=20
field, has mounted a different challenge to the standard view.=20
Petrides's studies point to two distinct levels of processing, both=20
within the prefrontal cortex. In his view the levels are distinguished=20
primarily not by whether they maintain information about place or=20
objects, as Goldman-Rakic holds, but rather by the abstractness of the=20
processing they perform. The lower level in the hierarchy--physically=20
lower in the brain as well as conceptually lower--retrieves data from=20
long-term memory storage elsewhere. The higher "dorsolateral" level, in=20
contrast, monitors the brain's processes and enables it to keep track of=20
multiple events. This higher monitoring level is called on when subjects=20
are asked, for example, to articulate a random list of each number from=20
1 to 10, with no repetition: a subject has to remember each digit=20
already chosen.=20

Petrides finds that both humans and monkeys with lesions in the=20
dorsolateral part of the prefrontal cortex are crippled in their ability=20
to monitor their own mental processes: they perform badly on special=20
tests he has devised that require subjects to remember their earlier=20
responses during the test. He also cites PET studies of healthy humans=20
that find heightened activity in the same region when subjects are=20
performing the tasks he uses. The finding is the same whether the tasks=20
involve spatial cues or not. "The material does not seem to matter--the=20
process is crucial," Petrides says.=20

Other researchers have found evidence to support the notion that the=20
higher parts of the prefrontal cortex are key for self-monitoring. In an=20
experiment by Mark D'Esposito and his associates at the University of=20
Pennsylvania, volunteers performed either one or both of two tasks that,=20
separately, did not require working memory. One task required subjects=20
to say which words in a list read aloud were the names of vegetables,=20
whereas the other asked them to match a feature of a geometric figure=20
seen in different orientations. Functional MRI showed that the=20
dorsoventral prefrontal cortex became active only when subjects=20
attempted both tasks simultaneously. And in April at a meeting of the=20
Cognitive Neuroscience Society, D'Esposito presented a meta-analysis of=20
25 different neuroimaging studies. The analysis supported Petrides's=20
general notion that tasks involving more computation involve higher=20
regions of the prefrontal cortex. "It was amazing that this came out,"=20
D'Esposito says.=20

D'Esposito's analysis also confirmed earlier indications that humans,=20
far more than monkeys, represent different types of information in=20
different halves of the brain. The meta-analysis did not, however,=20
detect the upper/lower distinction between spatial and object working=20
memory that Goldman-Rakic espouses.=20

Asymmetry of the human hemispheres is becoming apparent to other=20
researchers as well. John D. E. Gabrieli and his colleagues at Stanford=20
University have used functional MRI to study the brains of volunteers=20
who were solving pictorial puzzles such as those often found on=20
intelligence tests. The puzzles were of three types. One group was=20
trivial, requiring the subject simply to select a symbol identical to a=20
sample. A second group was a little harder: people had to select a=20
figure with a combination of features that was absent from an array of=20
sample figures. The third group contained more taxing problems that=20
required analytical reasoning.=20

Gabrieli's study sheds some light on the debate over the organization of=20
the prefrontal cortex. When volunteers pondered the intermediate class=20
of tasks, which most resembled the tasks other investigators have used=20
when studying working memory, the right side of the higher part of the=20
prefrontal cortex was prominently active. Moreover, the activity was in=20
areas that other researchers have found to be used when cues about=20
spatial location are stored. This result fits Goldman-Rakic's idea that=20
working memory for spatial location is stored in the higher regions of=20
the prefrontal cortex, because these intermediate tasks all demanded=20
that subjects visualize features in different locations.=20

When the volunteers in Gabrieli's experiment worked on the hard=20
problems, however, the prefrontal cortices of the subjects became even=20
more active, on the left as well as the right side. The added complexity=20
produced a pattern of activation like that Petrides has found during his=20
tests of self-monitoring.=20

Gabrieli's data thus provide some support for Petrides's theory of a=20
higher executive level in the prefrontal cortex, as well as for=20
Goldman-Rakic's view that domain-specific regions exist there. "There=20
are definitely domain-specific places," Gabrieli says. "And there are=20
others that rise above that." In other words, both sides in the debate=20
over domain-specific organization of the prefrontal cortex may have a=20
point. Yet in June, Matthew F. S. Rushworth of the University of Oxford=20
and his colleagues reported in the Journal of Neuroscience that monkeys=20
with large lesions in their lower prefrontal cortex could still perform=20
well on delayed-choice tests. The finding casts new doubt on the theory=20
that object working memory resides there and seems to support Petrides.=20

It may take years before the outstanding questions about the prefrontal=20
cortex are settled and the operation of the brain's executive functions=20
are pinned down to everyone's satisfaction. "If you put a theory out,=20
people will attack it," Goldman-Rakic muses. "Everyone is contributing."=20
And the modus operandi of the brain's decision-making apparatus is=20
slowly becoming visible. "We are getting," Goldman-Rakic observes, "to=20
the point where we can understand the cellular basis of cognition."=20


------------------------------------------------------------------------
Further Reading=20

"Cognitive Neuroscience." Special section in Science, Vol. 275, pages=20
1580-1610; March 15, 1997.=20

"Down-Regulation of the D1 and D5 Dopamine Receptors in the Primate=20
Prefrontal Cortex by Chronic Treatment with Antipsychotic Drugs."=20
Michael S. Lidow, John D. Elsworth and Patricia S. Goldman-Rakic in=20
Journal of Pharmacology and Experimental Therapeutics, Vol. 281, No. 1,=20
pages 597-603; April 1997.=20

Temporal Dynamics of Brain Activation during a Working Memory Task." J.=20
D. Cohen, W. M. Perlstein, T. S. Braver, L. E. Nystrom, D. C. Noll, J.=20
Jonides and E. E. Smith in Nature, Vol. 386, pages 604-608; April 10,=20
1997.=20

"Transient and Sustained Activity in a Distributed Neural System for=20
Human Working Memory." S. M. Courtney, L. G. Ungerleider, K. Keil and=20
James V. Haxby, ibid, pages 608-611.=20

"Integration of What and Where in the Primate Prefrontal Cortex." S.=20
Chenchal Rao, Gregor Rainer and Earl K. Miller in Science, Vol. 276,=20
pages 821-824; May 2, 1997.=20