|
Towards
a multiscale and multimodal model of the human
brain
Katrin Amunts
Jülich
Research Centre, Germany
Reference
brains are indispensable tools in human brain mapping enabling integration of
multimodal data into an anatomically defined standard space. Cytoarchitectonic maps as combined part of the JuBrain atlas, for example, inform about cortical areas
and nuclei, which are involved in a particular brain function as obtained in
a functional imaging study. Available reference brains, however, are
restricted to the macroscopical scale, and do not
provide information on the functionally important microscopical
dimension. They do not allow, e.g., to analyze findings on the level of
cortical layers and sublayers.
We
push the limits of current technology by creating the first ultra-high
resolution 3D- model of the human brain at nearly cellular resolution of 20
microns, based on 7,404 histological sections. The total volume of this
original histological data set was 1 TByte. Major
challenges of this human brain model comprise - among others - the highly
folded cerebral cortex, considerable inter-subject variability, and last but
not least, the pure size of the brain with its nearly 100 billion nerve cells
and the same number of glial cells. “BigBrain” is a freely available tool with unprecedented neuroanatomical insight. It allows extracting microscopic
data for modeling and simulation. BigBrain enables testing of hypotheses on optimal path
lengths between interconnected cortical regions, or spatial organization of
genetic patterning, redefining the traditional neuroanatomy
maps such as those of Brodmann and von Economo.
Back to Session I
|
|
Neuronal ensembles in human brain function
Nick
Ramsey
University of
Utrecht, Netherlands
With advances in imaging technologies for human research,
the attainable level of detail is starting to reveal the cortical
organization of brain functions. Thanks to the new 7 Tesla MRI scanners we
can now image humans at sub-mm resolution in vivo, and new insights into the
topographical representation of sensory and motor functions are starting to
emerge. Patches of several 100,000 neurons (approximately 1 mm of cortex), also called
functional units or neuronal ensembles, seem to perform specific functions
together, suggesting that they form the building blocks of sensory and motor
function.
However, the 7 Tesla MRI scanner measures bloodflow, a correlate of metabolism, and as such does
not reveal any of the neuronal processes underlying cortical function. Hence
we are now facing questions regarding the nature of the apparent coherent
activity within neuronal ensembles, and how information is represented and
transferred in terms of electrical and metabolic features. Addressing such
questions is starting to become possible with intracranial electrical recordings
in neurosurgery patients.
I will present some of the latest human functional
imaging results obtained with MRI and intracranial electrical recordings, as
well as potential clinical applications of recorded and decoded brain
signals. I will argue that we are on the verge of being able to address the
challenge of bridging what we know about single neuron dynamics from animal research and modeling, to what we know about dynamics of functional
units from neuroimaging and intracranial electrical
recordings in humans. I will then discuss what is required and what is
possible in terms of measurements in humans to address the challenge.
Back to Session I
|
|
Data-driven evaluation of functional
correlations in (massively) parallel spike trains
Sonja Grün
(1) Institute of Neuroscience
and Medicine (INM-6) and Institute for
Advanced Simulations (IAS-6) Forschungszentrum Jülich and
JARA, Germany
(2) Theoretical Systems Neurobiology, RWTH Aachen Univ., Germany
Cell assemblies, defined as groups of neurons
exhibiting precise spike coordination, were proposed as the units of network
processing in the cortex (Hebb, 1949). Fortunately,
in recent years considerable progress has been made in multi-electrode
recordings, which enable the simultaneous recording of massively parallel
spike trains of hundreds of neurons (e.g. Riehle et
al, 2013). However, statistical
methods developed for the identification of significant spike patterns in a
small number of parallel spike trains (e.g. Gr\"un
et al, 1999; 2002a,b; 2003; 2009) do not scale to
massively parallel spike data. Therefore, we develop statistical approaches
coping with the massive parallelism (Berger et al, 2007, 2010; Louis et al,
2010; Staude et al, 2010a,b; Picado-Muino
et al, 2013; Torre et al, subm). However, their
application to experimental data is still a challenge due to considerable
computational demands originating from a) large data sets and many recording
sessions, the requirement to b) perform the analyses in a time resolved
manner to follow the dynamics of neuronal processing and to c) evaluate the
significance of identified patterns based on surrogate data and Monte Carlo
approaches to account for the statistical properties of neuronal data and
avoid false positives. In this talk I will outline the reasonings
behind the above mentioned requirements and present first steps for deadling with the computational demands and the increased
complexity of the analysis workflows.
Back to Session I
|
|
Modeling Emergent Behavior
in Complex Dynamical Systems
Gyan Bhanot
Rutgers University
New Brunswick, NJ, U.S.A.
One of the outstanding problems in biology is to understand
how complexity and robustness arise as emergent phenomena, driven by
selection pressure in a world of interacting components. In this context, I
will first discuss a scenario for how and why complex systems such the brain
may have emerged in a world of RNA Replicators. Next, I will describe how
emergent stable behavior can be understood by modeling the dynamics of the cell cycle and predator-prey
systems. Finally, I will describe a genetic learning algorithm that may
explain the topological properties of neural systems that learn to perform
specific tasks.
Back to Session I
|
|
Integrating
brain structure and dynamics with spiking neuronal
network
models
Markus Diesmann
Jülich
Research Centre, Germany
(1) Institute
of Neuroscience and
Medicine (INM-6) and Institute for
Advanced Simulations (IAS-6) Forschungszentrum
J\"ulich and JARA,
Germany
(2) Medical Faculty, RWTH Aachen University, Germany
The cortical
microcircuit, the network comprising a square millimetre of brain tissue, has
been the subject of intense experimental and theoretical research. We
recently achieved full-scale simulations of this circuit at cellular and
synaptic resolution [1]: the model comprises about 100,000 neurons and one
billion local synapses connecting them.
The
purpose of the model is to investigate the effect of network structure on the
observed activity. To this end we incorporate cell-type specific connectivity
but use identical single neurons dynamics for all cell types. The emerging
network activity exhibits a number of the fundamental properties found in
nature: asynchronous irregular activity, layer specific spike rates, higher
spike rates of inhibitory neurons as compared to excitatory neurons, and a
characteristic response to transient input.
The importance of network structure is exemplified by the target
specificity of feedback connections which led us to the hypothesis of a
handshake mechanism between the cortical layers.
Despite
these successes, the explanatory power of such local models is limited as
half of the synapses of each excitatory nerve cell have non-local origins and
at the level of areas the brain constitutes a recurrent network of networks.
The model may be criticized as severly underconstrained. The second part of the talk therefore
argues for the need of brain-scale models to arrive at self-consistent
descriptions of the multi-scale architecture of the circuit. Such models will
enable us to relate the microscopic activity to mesoscopic
measures [2] and functional imaging data and to interpret those with respect
to brain structure.
www.nest-initiative.org
www.csn.fz-juelich.de
[1] Potjans TC, Diesmann M (2012)
Cerebral Cortex (online first)
doi:10.1093/cercor/bhs358
[2]
Linden H, Tetzlaff T, Potjans
TC, Pettersen KH, Grün S,
Diesmann M,
Einevoll GT
(2011) Neuron 72(5):859-872
Back to Session I
|
|
Scaling
up to brain-scale simulations with NEST
Abigail Morrison
Jülich
Research Centre, Germany
(1) Institute of Neuroscience
and Medicine (INM-6) and Institute for Advanced Simulations (IAS-6) Forschungszentrum Jülich and
JARA, Germany
(2) Institue for Cognitive Neuroscience, Faculty of
Psychology, Ruhr University Bochum,
Germany
As
the need for neuroscientific models integrating
multiple brain regions at cellular resolution grows, the availability of
corresponding high-performance simulation tools becomes ever more
critical. Scaling up the performance
of the software to exploit modern supercomputers, while important, is only
part of the story. In this talk, I will demonstrate the initiatives of the
NEST project to scale up not only its development practices, but also its
interaction with other neuroinformatics projects
and the wider neuroscience community.
[1] Zaytsev YV, Morrison A (2013) Front Neuroinform
6:31
[2] Helias M, Kunkel S, Masumoto G, Igarashi J, Eppler JM, Ishii S, Fukai T,
Morrison A, Diesmann M (2012) Front Neuroinform 6:26
[3]
Kunkel S, Potjans TC, Eppler
JM, Plesser HE, Morrison A, Diesmann
M (2012) Front Neuroinform 5:35
www.nest-initiative.org
www.fz-juelich.de/bfsd
Back to Session I
|
|
Vision for vision
Bart Ter Haar Romeny
Eindhoven University of Technology
Department of Biomedical Engineering, Netherlands
We aim to substantially improve computer vision
algorithms for image analysis in medical
imaging with an innovative
Lie group model, inspired by optical findings in early vision functional circuits,
and mathematical elegance. Current computer vision techniques often require
significant application-specific tuning and are often not generic, while
human vision is. Inspired by the multi-scale sampling in the visual front-end
we first discuss a multi-scale differential geometry engine model for early
vision, with a solid mathematical foundation, based on first principles and
proper regularization properties of the operators. The pinwheel structure in
the cortical columns have inspired us (Duits et al.)
to develop a novel multi-orientation framework, which combines (partial and
ordinary) differential equations on non-compact Lie groups(induced by
stochastic processes and sub-Riemannian geometric control) with wavelet
transforms. This leads to ’invertible orientation scores’, a
higher-dimensional transform enabling powerful contextual analysis.
The multi-orientation analysis can be regarded a
specific instance in a general higher dimensional Lie group theoretical model
for early vision. We aim to develop invertible scores for
multi-scale-and-orientation, multi-velocity, and multi-spatial frequency
analysis, at each spatial position.A key challenge
is to appropriately exploit these scores, their survey ofmultiple
features per position, their underlying group structure, and their invertibility.
In the invertible scores adaptive processes can be
exploited, such as geometry-driven diffusion with left-invariant evolutions
for contextual enhancement, and improvements in curveextractions
withleft-invariant sub-Riemannian optimal control.
For these challenges a junior ERC grant has recently been awarded to R. Duits in our group.
In medical image analysis, this approach is
exploited with excellent results on robust invariant features and shape
extraction in computer-aided diagnosis (e.g. in virtual colonoscopy), on
contextual enhancement of complex (i.e. highly curved, crossing) brain tractography from diffusion-weighted MRI(for
safenavigation in medio-temporal
epilepsy surgery), and on the tracking and analysis of complex retinal
micro-vessels (in a large screening program in China for early diabetes
detection).
These last applications prevent blindness, so
‘Vision for Vision’.
Short bio:
Prof. Bart M. ter Haar Romeny is professor at
Eindhoven University of Technology in the Netherlands, heading the
Biomedical Image Analysis (BMIA) group in the TU/e Department of Biomedical
Engineering. MSc in Applied Physics from Delft
University of Technology 1978 and
PhD from Utrecht
University in 1983. He
was Head of Physics at the Utrecht University Hospital Radiology Department
and associate professor at the Image Sciences Institute (ISI) of Utrecht University (1989-2001).
He is co-appointed professor at Maastricht University
and Northeastern
University (vice-dean
research). His research interests focus on biologically inspired image
analysis algorithms, multi-valued 3D visualization, especially brain
connectivity and computer-aided diagnosis (in particular for diabetes), and
image guided surgery, directed towards neurosurgery.
He is President of the Dutch Society for Pattern
Recognition and Image Processing, and has been President of the Dutch Society
for Biophysics & Biomedical Engineering (1998 – 2002) and the Dutch
Society of Clinical Physics (NVKF, 1990-1992).
He initiated the ‘Scale-Space’ conference series in
1997 (now SSVM). He is reviewer for many journals and conferences, and
organized several Summer Schools. He is an awarded teacher, and a frequent
keynote lecturer. Prof. Romeny is Senior Member of
IEEE, Board member of IAPR, Registered Clinical Physicist of NVKF, partner in
the Chinese Brainnetome consortium, and Board
member of the Dutch Foundation for Fundamental Research on Matter (FOM).
Back to Session II
|
|
Combining compositional shape hierarchy and
multi-class object taxonomy for efficient object categorisation
Ales Leonardis
University of Birmingham, United
Kingdom
Visual categorisation has been an area of intensive research
in the vision community for several decades.
Ultimately, the goal is to efficiently detect and
recognize an increasing number of object classes. The problem entangles three
highly interconnected issues: the internal object representation, which should
compactly capture the visual variability of objects and generalize well over
each class; a means for learning the representation from a set of input
images with as little supervision as possible; and an effective inference
algorithm that robustly matches the object representation against the image
and scales favorably with the number of objects. In
this talk I will present our novel approach which combines a learned
compositional hierarchy, representing (2D) shapes of multiple object classes,
and a coarse-to-fine matching scheme that exploits a taxonomy of objects to
perform efficient object detection.
Our framework for learning a hierarchical
compositional shape vocabulary for representing multiple object classes takes
simple contour fragments and learns their frequent spatial configurations.
These are recursively combined into increasingly more complex and
class-specific shape compositions, each exerting a high degree of shape
variability. At the top-level of the vocabulary, the compositions represent the
whole shapes of the objects. The vocabulary is learned layer after layer, by
gradually increasing the size of the window of analysis and reducing the
spatial resolution at which the shape configurations are learned. The lower
layers are learned jointly on images of all classes, whereas the higher
layers of the vocabulary are learned incrementally, by presenting the
algorithm with one object class after another.
However, in order for recognition systems to scale
to a larger number of object categories, and achieve running times
logarithmic in the number of classes, building visual class taxonomies
becomes necessary.
We propose an approach for speeding up recognition
times of multi-class part-based object representations. The main idea is to
construct a taxonomy of constellation models
cascaded from coarse-to-fine resolution and use it in recognition with an
efficient search strategy. The structure and the depth of the taxonomy is built automatically in a way that minimizes the number
of expected computations during recognition by optimizing the cost-to-power
ratio. The combination of the learned taxonomy with the compositional
hierarchy of object shape achieves efficiency both with respect to the
representation of the structure of objects and in terms of the number of modeled object classes. The experimental results show
that the learned multi-class object representation achieves a detection
performance comparable to the current state-of-the-art flat approaches with
both faster inference and shorter training times.
Back to Session II
|
|
2DSIL - A biologically plausible
computational model of shape representation
Antonio
Sanchez
University of Innsbruck, Austria
That shape is important for perception has been known
for almost a thousand years (thanks to Alhazen in
1083) and has been a subject of study ever since by scientists and phylosophers (such as Descartes, Helmholtz or the Gestalt
psychologists). Shapes are important object descriptors. If there was any remote
doubt regarding the importance of shape, recent experiments have shown that
intermediate areas of primate visual cortex such as V2, V4 and TEO are
involved in analyzing shape features such as corners and curvatures. The
primate brain appears to perform a wide variety of complex tasks by means of
simple operations. These operations are applied across several layers of
neurons, representing increasingly complex, abstract intermediate processing
stages. Recently, new models have attempted to emulate the human visual
system. However, the role of intermediate representations
in the visual cortex and their importance have not been adequately
studied in computational modeling.
In this talk I propose a model of shape-selective
neurons whose shape-selectivity is achieved through intermediate layers of
visual representation not previously fully explored. I hypothesize that hypercomplex - also known as endstopped
- neurons play a critical role to achieve shape selectivity and show how
shape-selective neurons may be modeled by
integrating endstopping and curvature computations.
This model - a representational and computational system for the detection of
2-dimensional object silhouettes that we term 2DSIL - provides a highly
accurate fit with neural data and replicates responses from neurons in area
V4 with an average of 83% accuracy. I successfully test a biologically
plausible hypothesis on how to connect early representations based on Gabor
or Difference of Gaussian filters and later representations closer to object
categories without the need of a learning phase as in most recent models.
References:
Pasupathy, A. and Connor, C.
(2002). Population coding of shape in area V4. Nature Neuroscience,
5(12):1332–1338.
Yamane, Y., Carlson, E., Bowman, K., Wang, Z., and Connor,
C. (2008). A neural code for three-dimensional object shape in macaque inferotemporal cortex. Nature Neuroscience,
11(11):1352–1360.
Serre, T., Wolf, L., Bileschi, S., and Riesenhuber,
M. (2007). Robust object recognition with cortex-like mechanisms. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 29(3):411–426.
Rodríguez-Sánchez, AJ and
Tsotsos, JK (2012), The roles of endstopped and curvature tuned computations in a
hierarchical representation of 2D shape. PLoS ONE 7
(8), pp. 1–13.
Back to Session II
|
|
Brain-Inspired Pattern Recognition
Nicolai
Petkov
Inst. Mathematics and Computer Science
University of Groningen, Netherlands
Insights into the function of the brain can provide clues
for designing effective computer algorithms for pattern recognition. This
thesis is illustrated by the design of feature detectors that is inspired by
the properties of shape-selective neurons in area V4 of visual cortex.
Such a filter is trainable as it is configured by
the automatic analysis of a feature specified by a user. Subsequently, it can
detect features that are similar to the training feature. By means of
training multiple such filters for different features of an object, we design
effective feature vector representation of that object that is analogous to
population coding in the brain. This approach shares two important aspects
with pattern recognition by the brain: the ability to learn and deal with
variability. It is illustrated by the detection of vascular bifurcations in
medical images, traffic sign recognition in complex scenes, optical character
recognition, word spotting, object recognition in a domestic environment and
the automatic motif and repetition detection in music.
Back to Session II
|
|
Towards the development of cognitive robots
Antonio
Bandera
University of Malaga, Spain
Cognitive neuroscience robotics integrates studies on
robotic, cognitive science and brain science. The aim is to translate those
models that explain the functionality of the human brain to the domain of
social robots. In order to engage people in social interactions, social
robots should be able to emanate responses at human interaction rates, and
exhibit a pro-active behaviour. This pro-active behaviour implies that the
internal architecture of these robots should not only be able to perceive and
act. It should also be able to perform off-line reasoning. Cognition is the
ability that allows us to internally deal with the information about the
external world and, hence, this ability is subject to the existence of an
internal representation of this information. Classical cognitive systems
posit an inner realm richly populated with internal tokens that stand for
external objects and states of affair. These internal representations,
however, are not valid to generate predictions or reasoning. Recent work
suggests that cognitive architectures cannot work on a passive, bottom-up
fashion, simply waiting to be activated by external stimuli. Instead, these
architectures must continuously use memory to interpret sensory information
and predict the immediate future.
These predictions about the outer world can be used
to actively drive the resources to relevant data in top-down modes of
behaviour, allowing an efficient and accurate interpretation of the
environment. This talk will describe two use cases where we are introducing
the use of socially interactive robots. These robots build the deliberative
planner over a mental model, a set of parameterised structures and procedures
that constitute a virtual representation of the reality. This model should
not be available only through interaction with the outer world, but it should
be used as virtual reality. This allows us to test the situation postulated
by Holland
(2004): at the heart of the mechanism is not just the body in the
environment, it is a model of the body in a model of the environment. This
situation allows emanating predictions from the model, which can be
correlated with real perceived information to drive attention, increase
efficiency and filter noisy perception, while the contents of the mental
model are also updated through experience.
Back to Session II
|
|
Brain Inspired Computing Structures, Scale,
and Function
Thomas
Sterling
Indiana
University, U.S.A.
Without exaggeration, the brain may be the most
complex system in the universe, and over the last half century or more has
inspired thinkers about its structure, shear scale (89 billion neurons with
10s of thousands of synaptic junctions), and its functionality yielding the
emergent behavioral properties of thought and
consciousness. Neural nets and algorithms as well as artificial intelligence
and cognitive science are among only a few of the many ways that researchers
are motivated and guided by understanding of the brain. This presentation
examines two issues inspired by brain form and function with practical
implications for future real-world systems and their operations. The first
objective discussed is to provide a lower bound on the resource requirements
to achieve a key property associated with the brain: intelligence. The CRIS
(Cognitive Real-time Interactive System) project at Indiana University (IU)
is dedicated to the development of an abstract architecture that embodies
many of the principal properties associated with intelligence and quantifying
the time and physical components required to achieve this level of operation
in real time. The second objective presented is to describe a simple cellular
automata structure capable of implementation of irregular dynamic graph
structures, the principal abstraction of intelligence. The CCA (Continuum
Computer Architecture) project, also at IU is dedicated to the creation of
the finest grain hardware components that in ensemble realize the emergent behavior of global general purpose parallel computing in
ways reminiscent to the way neural structures accomplish autonomic intelligent
operation but through virtualization of the interconnectivity via packet
switching (rather than physical topologies). While both projects are in their
inchoate phase, their respective goals are realizable and being undertaken.
This talk will describe both and tie them back to intrinsic characteristics
of the brain.
Back to Session III
|
|
Jesus Labarta
Barcelona Supercomputing Centre, Spain
Back
to Session III
|
|
Visual
Analysis of Human Brain Simulations -- An Overview
Torsten Kuhlen
Center
for Computing, Germany
While
in the past, visualization in Neuroscience focused mostly on medical imaging
techniques and microscopic data, the development of tools for a visual
analysis of simulated neuronal networks at scale is still in its infancy. My
talk will give a brief overview of related work in this field and will then
introduce some preliminary efforts resulting from a fruitful collaboration
between RWTH Aachen
University and the Institute of Neuroscience
and Medicine at the Research Center Jülich. Finally, I will present an outlook on future
work, particularly addressing the visualization requirements and challenges
in the Human Brain Project. To this end, I will briefly discuss the potential
of advanced methodology such as immersive Virtual Reality and multi-view
techniques.
Back to Session III
|
|
Analyzing
the brain: what makes visualization hard within the HBP
Luis Pastor
Polytechnical
University of Madrid, Spain
The
human brain is the most complex system Humanity has ever studied. The
complexity of each of the elements which compose it, as well as that of the
interconnections network that links them together is a barrier that has
hampered advances in research, despite the huge effort performed by a large
number of scientists during more than a century. The purpose of this talk is
to dig a little bit into the problem, analyzing the specific features that
make the visualization task difficult, and reviewing the main lines in which
research is strongly needed, from different points of view.
Back to Session III
|
|
Alessandro Curioni
Department Head, Mathematical &
Computational Sciences
IBM Research Division - Zurich Research Laboratory,
Switzerland
Back
to Session III
|
|
Visualization
towards the exascale in HBP
Vicente Martin
DLSIIS-
Facultad de Informatica, Univ. Politecnica and
Centro
de Supercomputacion y Visualizacion de Madrid, Spain
A key goal of the HBP is to develop a capability for
the analysis and visualisation of exascale data
sets and for the steering of simulations. With this capability,
supercomputers will act as an interactive scientific instrument, providing
researchers with visual feedback and allowing them to control simulations
while they are executing on a exascale computers.
A exascale computer will produce exabytes
of data. It is not expected that we will be able to store such an enormous
amount of data for later exploration. In fact, it is expected that the cost
of memory to floating point operation will grow, a trend that will make
necessary the in situ analysis of simulation data. In situ means essentially
that the analysis is done on the data as it is generated, since making a copy
to work on it in a completely asynchronous way is extremely costly. This
would allow to steer the simulation in close to real
time, turning a supercomputer in an interactive intrument
able to probe the structure and behaviour of the human brain. However, to
meet this challenge, current visualization paradigms have to be revisited and
the supporting hardware and software designed accordingly. The talk will
explore the challenges that have to be met to make this goal a reality.
Back to Session III
|
|
Non-Boolean Brain-like unconventional
computing
Youssry Botros
Intel, Aliso Viejo, CA, U.S.A.
Synchronized oscillations in disparate regions of
the brain have been observed in the visual cortex after a stimulus and it is
hypothesized that this synchronization may be correlated with the association
of memory fragments. This hypothesis is currently pursued by seeking to
organize hierarchies of associative memories.
An approach to realizing this associative memory hierarchy is to
utilize interacting nonlinear systems whose spontaneous synchronization
corresponds to a minimum in the energy space which is achieved by minimizing
the ‘distance’ between a memorized template and the test vectors; a point of
maximum entropy in information space.
A 15.5 G
pixel array has been tested and simulation results appear to offer signficant performance advantages (greater than 3 orders
of magnitude) in speed, power, and area, relative to the state of the art
micro-processor implementation. An
entropy-based associative memory implemented as coupled oscillators has been
conceptualized and several types of nano-oscillators
have been studied. An important task
ahead is to implement system prototypes and to identify classes of suitable
applications for this unconventional method of computing.
Back to Session IV
|
|
Bayesian
models of human multisensory perception
Ulrik Beierholm
University
of Birmingham, School of Psychology, United
Kingdom
A
human observer is constantly bombarded by a wealth of information from many
sensory sources, visual, auditory, haptics etc.. Given this complexity it quickly becomes neccessary, and even advantageous, to parse this
information and fuse together information from common sources. However infering the strucure
underlying the sources is a complex problem akin to performing causal
inference.
I
will present Bayesian statistical models for performing such inference and
will debate the experimental evidence for differentiating between them, with
a focus on human behaviour.
Overall,
evidence indicate that for perceptual tasks the
human brain performs close to an optimal observer.
Back to Session IV
|
|
Cognitive Vision Architectures - Fusing
Systems Engineering with Brain Science Insights
Ramesh Visvanathan
Frankfurt Institute for Advanced Studies
Frankfurt, Germany
The Frankfurt Vision initiative focuses on the
development of an integrated architecture for vision whose design is inspired
by bringing together insights from neuroscience, cognitive science/psychology
and computer vision systems. From a
systems perspective, we can view the human brain as an evolved system, with a
flexible learning architecture designed by nature to solve a range of
specific tasks in a class of environments that enhances the survival of
humans.
Model-driven systems engineering is a discipline
that formalizes application domain specification, i.e. task performance
requirements and contextual models, and translates these specifications into
system designs. Systems engineering in
the context of computer vision has its origins from the early 90’s and has
been refined over the years through practice.
At a coarse-level the architectures inspired from systems engineering
have parallels to models of brain function detailed in the brain and
cognitive sciences.
Our architectural design involves massively parallel
modules performing feed-forward decomposition of input visual signal into
constituent modalities (e.g. color, motion,
texture, shadow, reflection, contours, etc.) that allow for indexing into a
rich memory structure. Generated hypotheses are then refined via a dynamic,
recurrent process to converge to an interpretation. While both engineering
and brain science views of the architecture are in agreement at this higher
level, practical considerations present a multitude of options on module selection,
learning and inference approaches, and memory representation schemes. We present an overview of our ongoing
efforts in the construction of the cognitive vision framework and discuss
open research challenges.
Back to Session IV
|
|
Merging attention and segmentation: active foveal image representation
Rebecca Marfil
University of
Malaga, Spain
It is well established that the units of attention
on human vision are not merely spatial but closely related to perceptual
objects.
This implies a strong relationship between
segmentation and attention processes. This interaction is bi-directional: if
the segmentation process constraints attention, the way an image is segmented
may depend on the specific question asked to an observer, i.e. what she
'attend' in this sense. When the focus of attention is deployed from one
visual unit to another, the rest of the scene is perceived but at a lower
resolution that the focused object. The result is a multi-resolution visual
perception in which the fovea, a dimple on the central retina, provides the
highest resolution vision. While much work has recently been focused on
computational models for object-based attention, the design and development
of multi-resolution structures that can segment the input image according to
the focused perceptual unit is largely unexplored. This paper proposes a
novel structure for multi-resolution image segmentation that extends the
encoding provided by the Bounded Irregular Pyramid.
Bottom-up attention is enclosed in the same
structure, allowing to set the fovea over the most
salient image region. Preliminary results obtained from the segmentation of
natural images show that the performance of the approach is good in terms of
speed and accuracy.
Back to Session IV
|
|
Neuro-inspired Computational Engines
Murat Okandan
Microsystems Science and Technology
Sandia National Laboratories, U.S.A.
For neural processing functionality, a new hardware
approach, a new substrate is necessary. When deep neural approaches are
implemented on conventional von Neumann/Turing architecture with precise,
exact calculations only and large energy requirements for data movement, the
overhead penalty becomes too large. We
are developing a new substrate utilizing novel optoelectronic and electronic
devices, with the necessary functionality implemented at the lowest device
level possible to perform pattern recognition, abstraction, prediction and
model adjustment (learning), enabled by massive interconnectivity and reconfigurability (plasticity) driven by local rules.
Our proposed Grand Challenge Neuro-inspired
Computational Engines (NiCE) project is aimed at
demonstrating quantified performance improvement in specific applications
using the approach and systems outlined above. The goal for the project is to
show feasibility and benefits of using neuro-inspired
approaches, while developing the scientific and technological foundation for
this new hardware platform and information processing approach.
Back to Session IV
|
|
Computers like brains rather than brains on
computers
Karlheinz Meier
University of
Heidelberg, Germany
The lecture will discuss the state-of-the-art of
brain-inspired or neuromorphic computing systems.
The importance of configurability in view of our limited knowledge of brain
circuit architectures will be emphasized and technological solutions will be
presented. The plans for neuromorphic computing in
the Human Brain Project will be introduced with special emphasis on the
integration with the HPC infrastructure in the project and the operational
regimes of the two proposed large-scale neuromorphic systems.
Back to Session V
|
|
Challenges
and Opportunities for HPC in the Human Brain Project
Felix
Schurmann
Ecole Polytechnique
Federale de Lausanne, Switzerland
A
mouse brain is assumed have around 70 million and a human brain up to around
90 billion nerve cells. Each cell is a universe in itself, containing
possibly billions of proteins. In terms of scales, the physics and chemistry
involved for the functioning of the brain extends over 9 orders of magnitude
of spatial scales and up to 18 orders of magnitude of time scales considering
the lifespan of a human. Computationally speaking, modeling
such a system represents a formidable weak scaling matching today's trend of
parallelism in computing nicely yet requiring Moore's law to continue to work
at least for the next decade and a shift towards interactive supercomputing. At
the same time, it represents a strong
scaling challenge, which poses algorithmic challenges and opens up the
question of custom vs commodity in order to address
the rate limiting steps in hardware. Lastly, the brain is an information
processing device itself and thus might open up novel paths of computing
altogether. The lecture will give a detailed overview of the systematic and
architectural challenges of detailed brain simulations for High Performance
Computing.
Back to Session V
|
|
Thomas Lippert
Institute for Advanced Simulation
Jülich Supercomputing Centre and
University of Wuppertal, Computational
Theoretical Physics, and
John von Neumann Institute for
Computing (NIC), also
Europen PRACE IP Projects and of the
DEEP Exascale Project, Germany
Back
to Session V
|
|
Opportunities
for new developments in supercomputing in the HBP
Thomas Schulthess
Supercomputing
Center at Manno, Switzerland
Supercomputers
consume megawatts while the human brain runs at 30 watts. This comparison is
often used to motivate human brain studies as inspirations for future
computer systems development. But how do we move beyond the buzz? In this
lecture I will try to get us back on the ground of reality, and analyze the
real challenges for continued development of high-end computing. Hopefully
this will lead to a discussion on how neuroscience might inspire future
computing systems design.
Back to Session V
|
|
GOOSE:
Search on internet of connected sensors
George Azzopardi
TNO,
Netherlands
TNO is a non-profit research organization founded 80
years ago. The areas of expertise within TNO include technical sciences,
behavioural and societal sciences, as well as earth, environmental and life
sciences with applications in seven themes: a) healthy living, b) industrial
innovation, c) defence, safety and security, d) energy, e) transport and
mobility, f) built environment, and g) information society. Pattern
recognition and machine learning are active fields of research within
TNO. One major project is called GOOSE (GOOgle for SEnsors) which has the ambition to provide the capability
to search semantically for any relevant information within “all” (including
imaging) sensor streams, in near real time, in the entire internet of sensors.
The concept is similar to the capability provided by currently available
search engines which enable the retrieval of information on “all” pages on
the internet.
Back
to Session V
|
|
Brain Science and Computing - points of
encounter and chances for synergy
N. Petkov, T. Schulthess,
K. Meier, T. Sterling, K. Amunts
Back
to Session VI
|
|
Prototype-based learning and adaptive
distances for classification
Michael Biehl
University of Groningen, Netherlands
This tutorial talk gives an introduction to prototype
based systems in the context of supervised learning. The so called Learning
Vector Quantization (LVQ) serves as a particularly intuitive framework, in
which to discuss the basic ideas of distance or similarity based
classification. Problems from the bio-medical domain will be presented in
order to illustrate the concept.
An important issue is that of chosing
an appropriate distance measure for the task at hand. The elegant framework
of relevance learning aims at solving this problem means of adaptive distance
measures. These are determined together with the prototypes in the same
training process. Example problems illustrate how relevance learning provides
novel insight into the problem and how it can be used in the context of feature
selection.
Back to Session VII
|
