|
|
|
Mesoscopic emergence and interaction of spatiotemporal
patterns of activity and connectivity
|
|
The foundational thesis of cognitive science is that the mind
relies on internal states, or representations, that correspond
to states of the external world. It operates by creating, assembling
and transforming these states, both under the influence of external
stimuli and the constraints of its own internal dynamics. The
nature and structure of these mental states is still an open
problem, in particular their embodiment in the neural code
(i.e., the laws of organization of the electrophysiological signals).
|
|
The cognitive challenge
|
|
Cognitive science is currently a federation of disciplines—psychology,
AI, linguistics, logic, neuroscience, neural modeling, robotics,
etc.—lacking a unified theory. Cognitive models are broadly divided
between a logical paradigm, or "cognitivism", and a dynamical
paradigm, or "connectionism." Logical systems define high-level
symbols and formal grammars, but do not have a microstructure needed
to explain the fuzzy complexity of perception, memory or learning.
Conversely, dynamical systems define functionality as emerging from
neural networks and low-level activation equations, but lack a
macroscopic level supporting the systematic symbol composition of
language and reasoning. Bridging the gap between both approaches
requires an intermediate or mesoscopic level of description, at a
finer granularity than symbols but larger structural complexity
than small artificial neural nets.
Today's machines, which surpass humans in computationally intensive
tasks, are still surpassed by children in scene recognition, story
understanding or interactive behavioral tasks. The reason for this
persistent hiatus is that most artificial systems are engineered as
symbolic machines and do not rely on the same type of "building
blocks" (Bienenstock 1995) that the mind uses
at the subsymbolic/supraneuronal level.
Yet, these blocks might be the key to a true representational
invariance, i.e., schemas (cognitive, perceptual or motor),
categories and constituents, which can be only addressed by complex,
biologically inspired engineered systems.
|
|
Mesoscopic computation
|
|
When DNA, RNA and proteins were discovered, evolution and genetics
became united in biology's Modern Synthesis. In other terms, by
elucidating the mesoscopic level of life's complex self-organization
(molecular biology), macroscopic phenomena (heredity, speciation)
could be explained on the basis of microscopic elements (atoms;
Fig. a). Likewise, to explain the laws of perception and language
on the basis of elementary neuronal activities, a new discipline
of "molecular cognition" is needed (Bienenstock 1996; Fig. b).
What could then be the candidate "molecules" of cognitive science's
new Mind-Brain Modern Synthesis?
|
|
|
|
Spatiotemporal patterns
|
|
In general terms, if xi(t) denotes the
time-varying potential of neuron i, the cognitive molecules
are dynamic cell assemblies made of large, coherent sets of
neuronal activities { x1(t), ...,
xn(t) }. In particular, such a set
can be described as a spatiotemporal pattern (STP), which
is a complex series of spike timings
{ t11, t12,
t13, ...,
tn1, tn2,
tn3, ... }
containing many high-order statistical moments
<xi(t)
xj(t - τij)
xk(t - τik) ...>.
These moments are typically combinations of synchronized groups
(delays τij = 0) and waves or "rhythms"
(delays τij > 0). They correspond to
reproducible temporal correlations among electric
signals, supported by underlying regular patterns of connectivity.
Similarly to proteins, STPs can interact in several ways and
assemble at several levels, forming a hierarchy of complex structures
from simpler ones in a modular fashion. Thus, by relying on temporal
coding, STPs might constitute the building blocks of intelligent
behavior. While individual firing rates
<xi(t)> have traditionally
dominated neuroscience, alternative theories
(von der Malsburg 1981, Abeles 1982) have long
proposed temporal code and higher-order correlations as the basic code
used by the brain to represent mental entities. Since the 1980's,
the correlation hypothesis launched a series of experiments and
models investigating synchronization and wave patterns among
oscillatory or otherwise excitable units. This new field of neural
dynamics is known today under the broad appellation of
spiking neural networks.
|
|
Compositionality
|
|
In summary, complex spatiotemporal phenomena in large-scale neural
populations have the potential to support the sought-after
mesostructure of symbolic and combinatorial systems. To continue the
chemical metaphor, temporal binding and synaptic plasticity together
play the role of elementary forces or "bonds": strong "covalent"
bonds maintain their cohesiveness and stability inside STPs; weaker
"ionic" or "hydrogen" coupling can quickly assemble and disassemble
STPs at a larger scale.
Different spiking neural models have focused on different
classes of neuronal dynamics at varying levels of biological
detail: conductance-based, integrate & fire, pulsed,
oscillatory, excitable, rate-coded, binary, etc. They have also
explored different forms of temporal order binding these neurons
together: synchronization, phase locking, delayed correlations,
waves, rhythms, induction, resonance, etc. In recent years, already
several theoretical proposals have populated the mesoscopic level
with STP-like molecular objects: synfire chains and braids
(Abeles 1982, Doursat 1991, Bienenstock 1995, Doursat &
Bienenstock 2006), polychronous groups (Izhikevich 2006), cortical
columns (Markram 2006), traveling waves (Doursat et al. 1995,
Doursat & Petitot 2005), subthreshold harmonics
(Doursat & Goodman 2006), etc.
|
|
Projects
|
The projects presented in this section each follow one of these
mesoscopic paradigms, addressing different topics and challenges in
robotics, machine vision, linguistic and pattern recognition:
- —
A neural network model of associative learning by "lock and key"
coherence induction between dynamic cell assemblies
- —
Parallel self-organization of connectivity and activity in an
initially random spiking neural network, which can support a
hierarchy of structured representations and be applied to visual,
auditory or linguistic tasks
- —
Bridging the gap between vision and language by
importing complex system modeling into linguistics
|
|
Selected publications
|
|
Doursat, R. & Bienenstock, E. (2006b)
Neocortical self-structuration as a basis for learning.
5th International Conference on Development and Learning
(ICDL 2006),
May 31-June 3, 2006, Indiana University, Bloomington, IN.
Doursat, R. & Petitot, J. (2005b)
Dynamical systems and cognitive linguistics: Toward an
active morphodynamical semantics.
18: 628-638.
Selected for this special issue among less than
10% of the papers accepted at the IJCNN 2005 conference.
Goodman, P. H., Doursat, R., Zou, Q., Zirpe, M. & Sessions, O. (2007)
RAIN brains: Mammalian neocortex as a hybrid analog-digital computer.
Unconventional Computation Conference
(),
March 21-23, 2007, Los Alamos National Laboratory (LANL) and
Santa Fe Institute (SFI), Santa Fe, NM.
|
|
References
|
|
Abeles, M. (1982)
Local cortical circuits: an electrophysiological study.
Studies of Brain Function (vol. 7), V. Braintenberg, ed.,
Springer-Verlag, Berlin.
Bienenstock, E. (1995)
A model of neocortex.
Network: Computation in Neural Systems, 6: 179-224.
Bienenstock, E. (1996)
Composition. In: Brain Theory - Biological Basis and Computational Theory of Vision,
Ad Aertsen and Valentino Braitenberg eds, Elsevier, pp 269-300.
Izhikevich, E. M. (2006)
Polychronization: Computation with spikes.
Neural Computation, 18: 245-282.
Markram, H. (2006)
The BlueBrain project.
Nature Reviews Neuroscience, 7(2): 153-160.
von der Malsburg, C. (1981)
The correlation theory of brain function.
Internal Report 81-2, Max Planck Institute for Biophysical Chemistry,
Department of Neurobiology, Göttingen, Germany.
|
|
