Abstract :
[en] We know that color and form are processed in distinct areas of the human brain. This
information must somehow be brought together. How the brain might achieve these colorform
associations, as well as all other associations of this type, is one of the central themes of
this dissertation. When looking at a field of poppies on a sunny day, how can we correctly
associate the color red with the poppies, green with the grass, and blue with the sky, while
avoiding associating the color red with the grass and the color blue with poppies? How can
we associate the perception of red poppies with the name “red poppy,” and with its
superordinate category “flower?” A red poppy is composed of several features, like its
shape, color, texture, etc. How might a cognitive system bind these features to build a
coherent whole? If we see Louise picking a red poppy, how can we correctly associate
Louise with the picker and the red poppy with the picked object, without making the opposite
and incorrect association? These associations may seem easy to us, but how does the brain
achieve them? How a cognitive system binds a set of features together, associates a filler
with a role, a value with a variable, an attribute with a concept, ... is what we mean by “the
binding problem.”
This thesis focuses on the neurobiological processes that enable connectionist cognitive
systems to display binding abilities, on the constraints that affect the binding process, and on
the cognitive consequences of these constraints. To study these processes, we developed a
computer model of them. This method forces a detailed and unequivocal description of
processes used by the simulation. This method is also a powerful means of generating new
hypotheses. In this study we attempt to link psychological processes with the neuronal
constraints that act on brain functioning. The brain is composed of approximately 10 billion
highly interconnected neurons. To achieve binding it is necessary for neurons to
communicate with each other because it has been shown that different aspects of a perceived
object are not processed in the same cortical areas. Therefore, there must be a means for
binding neurons responding to each of these different aspects. The neurons responding to the color red, to the object’s shape, and to its name must be linked to produce a coherent
whole representing the red poppy.
Neurons are connected by synapses. The functioning of these connections is constrained
by the architecture of the brain and by the process of signal transmission. A particular neuron
is connected to a relatively small set of other neurons. Therefore, communication between
any two neurons generally requires a chain of transmission through intermediate neurons. A
pre-synaptic neuron has an effect on another neuron (called the post-synaptic neuron) only if
the pre-synaptic neuron emits an action potential (i.e., if it fires). As a consequence, this
brief polarization, which last a few milliseconds, results in a modification of other neurons'
firing potential. Transmission efficiency depends on the strength of the connecting synapses
and the state of the post-synaptic neuron. When a neuron emits an action potential, it is
completely insensitive to incoming signals for a short period, then its sensitivity slowly
increases. A single pre-synaptic cortical neuron cannot alone provoke the post-synaptic
neuron firing. This post-synaptic neuron must receive convergent and more or less
synchronized signals from many synapses in order to fire.
These neurobiological properties of neurons and neuronal firing constrain the way in
which the brain can achieve binding. Among the various hypotheses of how this could be
done, we chose synchronization of action potentials for our model. In the red poppy
example, neurons responding to the color red will fire in synchrony with those responding to
the shape of the flower and to the name “red poppy.” This particular synchronized cluster
corresponding to “red poppy” must be temporally distinguished from the cluster responding
to “green grass.” Numerous neurobiological studies seem to confirm this action-potential
synchrony hypothesis. They show that synchronization involves a particular timing precision
and occurs at a particular oscillation frequency. This oscillation requires participating
neurons to fire repeatedly and rhythmically for a particular period of time. These properties
of firing timing and duration have been implemented in a computer simulation called
INFERNET. This artificial neural network uses integrate-and-fire nodes (artificial neuronlike
elements). These nodes fire at a precise moment and transmit their activation, with a
particular strength and delay, to nodes connected to them. When the potential of the node
reaches a particular threshold, it emits a spike. Thereafter, the potential is reset to a resting
value. As with real neurons, this node will then be completely insensitive to incoming
signals for a short period, after which its sensitivity will slowly increase.
INFERNET solves the binding problem by means of oscillation synchrony. Symbols are
represented by clusters of nodes firing in synchrony. Fillers are also bound to their roles by
synchrony. This synchronous activity defines a window of synchrony i.e., an interval
during which the required nodes fire. This time interval takes neurally plausible values.
Object discrimination is achieved by a succession of windows of synchrony. Bindings are
maintained in memory by the use of particular oscillations. The rhythmic activity and the
synchrony precision constrain the number of distinct entities that the system is able to
maintain in memory. This represents the short term memory span of INFERNET. We show
that this span is comparable with human short term memory span.
The limited number of windows of synchrony also constrains predicate representations.
This prediction is tested on human participants. If there are too many windows of
synchrony, these will interfere with each other. In addition, binding strength decreases with
time. These two properties explain why the short-term memory of INFERNET displays
primacy and recency effects similar to those observed in humans. Bindings in INFERNET
are also constrained by the number of intermediate steps required for particular role nodes to
enter into synchrony with the filler nodes. This constraint is shown to provided a plausible
explanation of various differences human reasoning. The last INFERNET constraint
concerns multiple instantiation. This problem arises in connectionist networks as soon as a
symbol has to be simultaneously used twice in different ways. Since INFERNET’s short
term memory is the transient activation of parts of long term memory, it cannot make multiple
copies of a symbol, in the same way, for example, that a symbolic system does. The
INFERNET solution to the multiple instantiation problem involves superposition of different
node oscillations. This process is constrained by the refractory period of the nodes. A
number of simulations with INFERNET and experiments on humans show that this solution
is psychologically plausible. Multiple instantiation is also shown to be a plausible
explanation of certain similarity effects in short term memory. INFERNET is also shown to
be capable of symbolic processing with using neurologically and psychologically plausible
mechanisms that have the advantages of generalization and noise tolerance found in
connectionist networks. Finally, under certain circumstances, noise is shown to enhance
INFERNET’s processing capabilities.