Roy proposes that the only difference between distributed representation
and localist representation brain models is that localist neurons have meaning
by themselves, and distributed neurons do not. He argues that experimental
evidence supports the view that localist neurons are widespread throughout the
brain, in contrast with the connectionist brain model in which a pattern of
neuronal activity is needed to represent a concept.
The classic theory of the brain
is one of connections, in which the brain consists of a network of neurons that
interact with each other to allow us to think, see, interpret, and understand
the world around us. In this model, called distributed representation, an
individual neuron by itself has no inherent meaning, but only contributes to a
pattern of neuronal activity that has meaning. For example, a certain pattern
of many neurons fires when you think "dog" and another pattern for "cat."
"The belief in distributed
representation theory is that a concept or object is not represented by a
single neuron in the brain but
by a pattern of activations over a number of neurons," explains Asim Roy,
a professor of information
systems at Arizona State University, to Medical Xpress .
"Thus there is no single neuron in the brain representing a cat or a dog.
Proponents of this theory claim that a cat or a dog is represented by its
microfeatures such as legs, ears, body, tail, and so on. However, they think
that neurons have absolutely no meaning on a stand-alone basis. Therefore, they
go further and claim that these microfeatures are at the subsymbolic level,
which means that meaning arises only when you consider the pattern of
activations as a whole. Therefore, there are no neurons representing legs,
ears, body, tail, etc. The representation is at a much lower level."
Roy is among a number of
scientists working in the fields of neuroscience and artificial intelligence
(AI) who suspect that the brain may not be as connected as distributed
representation suggests. The basis of their alternative model, called localist
representation, is that a single neuron can represent a dog, a cat, or any
other object or concept. These neurons can be considered symbols since they
have meaning on a stand-alone basis. However, as Roy explains, this doesn't
necessarily mean only one neuron represents a dog; such "concept
cells" are high-level neurons, which fire in response to the firing of an
assortment of low-level neurons that represent the legs, ears, body, tail, etc.
"In localist representation,
there could be separate neurons for a dog and a cat, and also neurons for legs,
ears, body, tail, etc.," he said. "It's very similar to the model in
my paper for word recognition, which is an old model from James McClelland
[Chair of the Psychology Department at Stanford University] and [the late
pioneering neuroscientist] David Rumelhart. You have low-level neurons that
detect letters of the alphabet and then high-level neurons for individual
words. So letter neurons and word neurons, they both exist."
The origins of this dispute
between localist and distributed representation goes back to the early '80s, to
a dispute between the symbol processing hypothesis of artificial intelligence (AI)
and the subsymbolic paradigm of connectionists. In the past 30 years, the
debate has only intensified.
Not so different after all?
Staunchly on the side of the
symbol model, Roy has published a paper in a recent issue of Frontiers
in Cognitive Science in which he makes two main claims that he thinks
will ramp up support for localist representation. First, he proposes that
distributed representation and localist representation models are essentially
the same, with just one small but important difference: localist neurons have
meaning by themselves, and distributed neurons do not. Traditionally, the two
models have been thought to have inherent structural differences. Roy's second
claim is that localist representation and its symbolic, meaningful neurons are
widespread throughout the brain. Up to now, even the strongest proponents of
localist representation considered that the brain may use symbolic neurons only
in some areas at certain levels of processing.
In regards to his first point, he
explains that several misconceptions of the two models have led scientists to
assume that they differ more than they actually do.
"The first misconception is
that the property where 'each concept is represented by many units, and each
unit represents many different concepts' is exclusive to distributed
representation," he said. "I show that that property is actually a
property of the model that one builds, not of the units. A second misconception,
which is partly related to the first, is that a localist unit should respond to
one and only one concept. I show that that is not true either, that localist
units can indeed respond to many different higher-level concepts. All these
false notions haunt localist representation, and the first thing I did was show
that they are false notions. And you can show them to be false only if you
stick to the basic property of localist units, that they have 'meaning and
interpretation on a stand-alone basis.'"
If Roy is correct, it would mean
that many of the arguments used against localist representation – in
particular, against University of Bristol Psychology Professor Jeff Bowers'
"grandmother cell theory" – are invalid. (Put simply, grandmother
cells are high-level concept neurons.) But perhaps more importantly, Roy's
interpretation also means that any model built with distributed neurons can be
built with localist neurons, since there is no structural difference. In other
words, a model in which a neuron responds to multiple concepts can be either
distributed or localist.
A neuron for everyone and
everything
This interpretation clears the
path to Roy's second claim, that the brain processes information using symbols,
not subsymbolic connections. He explains that experimental support for
symbol-based localist representation is robust, with some of the earliest
evidence coming from studies of the visual system.
"There's more than four
decades of research on receptive fields in the primary visual cortex and even
in retinal ganglion cells that shows that the functionality of the cells in
those regions can be interpreted," Roy said. "Researchers have found
cells that detect orientation, edges, color, motion, and so on. David H. Hubel
and Torsten Wiesel won the Nobel Prize in physiology and medicine in 1981 for
breaking this 'secret code' of the brain."
The discovery of these vision
cells is just one piece of neurophysiological evidence suggesting that
individual neuron cells have meaning and interpretation. Roy also cites several
recent studies that have identified individual neurons in the hippocampus and
the medial temporal lobe that represent specific objects or concepts and do not
depend on the activity of other neurons. For example, in 2005, neuroscientists
discovered that an epilepsy patient had one neuron cell that fired whenever a
photo of Jennifer Aniston was presented. Various photos showing the blonde
actress in different poses and from different angles all elicited a response
from the same concept cell, a neuron in the hippocampus.
"Concept cells were also
found in different regions of the medial temporal lobe," Roy said.
"For example, a 'James Brolin cell' was found in the right hippocampus, a
'Venus Williams cell' was in the left hippocampus, a 'Marilyn Monroe cell' was
in the left parahippocampal cortex and a 'Michael Jackson cell' was in the
right amygdala."
Roy thinks that one of most
supportive studies of his argument is the Cerf experiment from 2010. In this
experiment, Moran Cerf, a neuroscientist at New York University and UCLA, asked
epilepsy patients to look at several different images on a screen while the
researchers attempted to identify one neuron in the medial temporal lobe that
independently fired for each of the different images. One of the images was
then randomly selected to become the target image, and patients were shown the
target image at 50% visibility and a distractor image at 50% visibility and
asked to focus their thoughts on the target image.
The visibility of the target
image increased when the firing rate of the previously identified target neuron
increased compared to the firing rate of the distractor neuron. By focusing on
the target images, the patients could increase the target neuron's firing rate,
with 69% of the patients succeeding in making the target image 100% visible.
In Roy's perspective, these
results suggest that the neuron the researchers originally identified as the
representative neuron for the target image was indeed a localist neuron. In
other words, when that neuron fired, it had one specific meaning: the patient
was thinking of the target image.
Roy emphasized that he did not
look exclusively for studies to support his claim and ignore studies that
contradicted it; he says he found no evidence that might contradict his claims.
"Although I have not
exhaustively searched this literature, from what I looked at, there was not
much to 'pick and choose' from," he said. "In the paper, I have cited
some recent studies. And although I have not covered the universe of single cell
studies on insects, animals, and humans, the ones I have looked at don't
contradict my broad claim.
"There are some studies that
show that a population of neurons has meaning," he acknowledged. "But
that doesn't contradict my theory. For example, one can read the outputs of
cells representing legs, ears, body, tail, and so on, and say that represents a
cat. However, that doesn't contradict the claim that all of these cells have
meaning and interpretation on a stand-alone basis, even though only when their
outputs are combined can you say that it's a cat."
Future developments
All this evidence further
solidifies Roy's impression that the brain is a system of symbols rather than a
network of connections. If he's correct, then it would have implications for our
understanding of the brain and future AI developments.
"The brain would need fewer
connections with localist representation than with distributed
representation," he said. "There is efficiency and filtering
associated with localist representation. We can quickly filter out aspects of a
scene without further processing. And that saves computations and energy
consumed. Our brains would be exhausted if they didn't filter out irrelevant
things quickly."
Applying the brain's symbolic
representation to create AI systems may sound more straightforward than
attempting to build AI systems using a subsymbolic mode, but it's far from
simple.
"Localist representation may
sound simplistic, but we are still struggling with the mathematics to replicate
those functionalities, even for the visual system," Roy said. "So
maybe it's not that simple."
Commentary on Roy's paper by
David Plaut
David Plaut, Psychology Professor
at Carnegie Mellon University, carries out research using the connectionist
framework for computational modeling of brain functions. He has found issues
with a few ideas in Roy's paper, starting with the fact that Roy frames the argument
on neural representation differently than how it's usually framed.
"Asim's main argument is
that what makes a neural representation localist is that the activation of a
single neuron has meaning and interpretation on a stand-alone basis,"
Plaut said. "This claim is about how scientists interpret neural activity.
It differs from the standard argument on neural representation, which is about
how the system actually works, not whether we as scientists can make sense of a
single neuron. These are two separate questions."
Plaut also thinks that Roy needs
to clearly define what he means when he says that a neuron has "meaning
and interpretation."
"My problem is that his
claim is a bit vacuous because he's never very clear about what a coherent
'meaning and interpretation' has to be like," he said. "He brings up
some examples that he claims are supportive of neurons having meaning and
interpretation, such as in the medial temporal lobe and hippocampal regions,
but never lays out what would count as evidence against his claim. On his view,
if we can't yet characterize the function of a neuron, it just means we haven't
figured it out yet. There's no way to prove him wrong."
In fact, Plaut thinks that much
of the experimental evidence that Roy cites as support for his view may not be
as supportive as Roy claims.
"If you look at what he says
'meaning and interpretation' is supposed to be coding for, if you look into the
examples he gives, they're not actually quite like that," Plaut said.
"If you look at the hippocampal cells (the Jennifer Aniston neuron), the
problem is that it's been demonstrated that the very same cell can respond to
something else that's pretty different. For example, the same Jennifer Aniston
cell responds to Lisa Kudrow, another actress on the TV show Friends with
Aniston. Are we to believe that Lisa Kudrow and Jennifer Aniston are the same
concept? Is this neuron a Friends TV show cell?"
He notes that there are other
examples; for instance, there is one neuron that fires for both spiders and snakes,
and another neuron that fires for both the Eiffel Tower and the Leaning Tower
of Piza – somewhat related concepts, perhaps, but still with quite distinct
meanings.
"Only a few experiments show
the degree of selectivity and interpretability that he's talking about,"
Plaut said. "For example, Young and Yamane published a study in 1992 in
which, out of 850 neurons, they found only one that had this high level of
selectivity, while the other cells had varying degrees of responses. If we
ignore what the vast majority of what neurons are doing, it's selection bias.
In some regions of the medial temporal lobe and hippocampus, there seem to be
fairly highly selective responses, but the notion that most cells respond to
one concept that is interpretable isn't supported by the data."
Commentary on Roy's paper by
James McClelland
As mentioned above, one of the
papers that Roy cites is coauthored by James McClelland, a psychology professor
at Stanford University whose work has played a pivotal role in developing the
connectionist framework. In response to Roy's paper, McClelland explained why
he still favors the distributed representation model:
"Roy's paper lays out his
claim that the brain uses localist representation – the view that individual
neurons in the brain have 'meaning and interpretation' on a stand-alone basis –
and contrasts this with the distributed representation view – the view that
each neuron participates in many representations, and that it is therefore not
possible to determine what concept is being represented by looking at the
activity of a single neuron. Although my collaboration with David Rumelhart
exploring neural networks began with the exploration of localist models
(McClelland & Rumelhart, 1981), we soon became convinced that the localist view
is unlikely to be correct (McClelland & Rumelhart, 1985). Here I briefly
explain why I still hold the distributed representation view.
"One problem with localist
representation is the question, when to start and when to stop using a localist
representation. Suppose I encounter a new kind of bread – one baked in thin
sheets with sesame and cardamom seeds. In order to understand that this new
kind of bread might smell or taste like, I would likely rely on representations
of other kinds of bread and of sesame and cardamom seeds, and also on my
knowledge of other kinds of foods in thin sheets that I may know about. I
already have a great deal of knowledge about this thin bread, having never
encountered it before. Did I already have a localist representation for it, or
did I compose my understanding of it out of knowledge I had previously acquired
for other things? If the latter, what basis do I have for thinking that the
representation I have for any concept – even a very familiar one – as
associated with a single neuron, or even a set of neurons dedicated only to
that concept?
"A further problem arises
when we note that I may have useful knowledge of many different instances of
every concept I know – for example, the particular type of chicken I purchased
yesterday evening at the supermarket, and the particular type of avocados I
found to put in my salad. Each of these is a class of objects, a class for
which we may need a representation if we were to encounter a member of the
class again. Is each such class represented by a localist representation in the
brain? The same problem arises with specific individuals, since we know each
individual in many different roles and phases. Do I have a localist
representation for each phase of every individual that I know? Given these
questions, my work since the 1985 paper has focused on understanding how the
brain may use what it has learned about many different and partially related
experiences, without relying exclusively on localist representations.
On this view, the knowledge
arising from an experience is the set of adjustments made to connection weights
among participating neurons – neurons that participate in representing many
different things.
"Roy lays out several lines
of argument in support of his point of view. Perhaps the central argument is
that recordings from neurons show that the neurons in some parts of the brain
have what some might consider to be surprisingly specific responses. Let us
discuss one such neuron – the neuron that fires substantially more when an
individual sees either the Eiffel Tower or the Leaning Tower of Pisa than when
he sees other objects. Does this neuron 'have meaning and interpretation
independent of other neurons'? It can have meaning for an external observer,
who knows the results of the experiment – but exactly what meaning should we
say it has? An even harder question is, what meaning does the neuron have for
the individual in whose brain it has been found? Let's take the simpler
question first.
"First, for the external
observer: it should be apparent that the full range of test stimuli used
affects what meaning we assign to such a neuron. The Japanese neuroscientist
Keiji Tanaka found neurons in monkeys' brains that others had called 'monkey
paw detectors' and others they might have called 'cheshire cat detectors,' but
he then constructed many special test stimuli to use in testing each neuron.
He found that the neurons
generally responded even better to schematic stimuli that were not recognizably
paws or cats but had features in common with them. Such neurons surely
participate in representing cats or paws but may also participate in
representing other objects with similar shape features. Critically, however,
the response of the neuron is difficult to pin down in simple verbal terms and
neighboring neurons have similar responses that shade continuously from one
combination of features to another. Is the same true of the Eiffel
Tower/Leaning Tower of Pisa neuron? In the context of these observations, the
Cerf experiment considered by Roy may not be as impressive. A neuron can
respond to one of four different things without really having a meaning and
interpretation equivalent to any one of these items.
"Second, to the individual
in whose brain the neuron has been found: Roy's analysis ignores the question
of how a neuron assigned to represent a concept is then used by the observer to
mediate use of the observer's knowledge of the concept. This is the issue my
colleagues and I have sought to explore with explicit models that rely on distributed
representations over populations of simulated neuron-like processing units.
While we sometimes (Kumeran & McClelland, 2012, as in McClelland &
Rumelhart, 1981) use localist units in our simulation models, it is not the
neurons, but their interconnections with other neurons, that gives them meaning
and interpretation. The sight of a picture of Saddam Hussein brings to mind
heinous crimes against the citizens of Iraq and Kuwait, not because a
particular neuron is activated but because it (and many other neurons)
participates in activating other neurons that are involved in the
representation of other heinous crimes and/or in verbal expressions and
imagined scenes involving such crimes. And it participates in activating these
other neurons because of its connections to these neurons. Again we come back
to the patterns of interconnections as the seat of knowledge, the basis on
which one or more neurons in the brain can have meaning and interpretation.
"In our work we have
proposed that different parts of the brain rely on representations that differ
in their relative specificity (McClelland et al, 1995; Goddard &
McClelland, 1996). The Medial Temporal Lobes are thought to represent items,
locations, events, and situations in terms of sparse patterns of activation,
but even here each neuron is thought of as participating in many
representations. Even here, the principles of distributed representation apply:
the same place cell can represent very different places in different
environments, for example, and two place cells that
represent overlapping places in one environment can represent completely
non-overlapping places in other environments. Other parts of the neocortex of
the brain are thought to rely on denser distributed representations, where a
somewhat larger overall fraction of the neurons are activated by a particular
item, location, etc. There is a lot more to understand about these
representations. Studies involving very small numbers of neurons may be
misleading in this regard. Progress will depend on recording from large numbers
of neurons, so that we can
more readily visualize the activity across the entire population."
Roy has responded to Plaut's and
McClelland's comments here.
More information:
Roy, A. "A theory of the
brain: localist representation is used widely in the brain." Frontiers
in Cognitive Science.
Kumaran, D. & McClelland, J.
L. (2012). "Generalization through the recurrent interaction of episodic
memories: A model of the hippocampal system." Psychological Review,
119, 573-616.DOI: 10.1037/a0028681
Rogers, T. T. & McClelland,
J. L. (2004). Semantic Cognition: A Parallel Distributed Processing Approach.
Cambridge, MA: MIT Press
McClelland, J. L., McNaughton, B.
L., & O'Reilly, R. C. (1995). "Why there are complementary learning
systems in the hippocampus and neocortex: Insights from the successes and
failures of connectionist models of learning and memory." Psychological
Review, 102, 419-457
McClelland, J. L. &
Rumelhart, D. E. (1985). "Distributed memory and the representation of
general and specific information." Journal of Experimental Psychology:
General, 114, 159-197
McClelland, J. L. &
Rumelhart, D. E. (1981). "An interactive activation model of context
effects in letter perception: Part 1. An account of Basic Findings." Psychological
Review, 88, 375-407
No comments:
Post a Comment