INTERNATIONAL WORKSHOP
ON
BRAIN-INSPIRED COMPUTING
Computational models, algorithms and applications
Cetraro –
Final
Programme
Updated
May 30th, 2017
Programme and Organisation Committee
|
Katrin Amunts |
(Germany) |
|
Lucio Grandinetti |
(Italy) |
|
Thomas Lippert |
( |
|
Nicolai
Petkov |
(The Netherlands) |
|
|
|
|
|
|
Katrin Amunts |
Lucio Grandinetti |
Thomas Lippert |
Nicolai Petkov |
Sponsors
|
|
|
|
CRAY |
|
|
|
|
|
IBM (t.b.c.) |
|
|
|
|
|
ICAR CNR |
|
|
|
|
|
INTEL (t.b.c.) |
|
|
|
|
|
JUELICH SUPERCOMPUTING CENTER, Germany |
|
|
|
|
|
Human Brain Project |
|
|
|
|
|
ParTec |
|
|
|
|
|
|
|
|
|
|
|
Dipartimento di Ingegneria
dell’Innovazione – Universitŕ del Salento |
|
|
Speakers 25 invited key-note speakers will
participate. So far the following speakers – among others – have confirmed
their participation: Thomas Lippert, Tianzi Jiang, Katrin Amunts, Thomas
Schulthess, Francesco Pavone, Nicolai Petkov, Thomas Sterling, Adrian Tate. |
|
Proceedings Refereed papers presented at the Workshop
will be published as a Proceedings Volume in the Springer series Lecture
Notes in Computer Science. Interested authors
are kindly invited to submit in advance to the PC, during the workshop, the
title and one page abstract of their intended contribution. |
Workshop Agenda
Monday, June 12th
|
Session |
Time |
Speaker/Activity |
|
|
9:45
– 10:00 |
Welcome Address |
|
|
Brain structure and function: a neuroscience
perspective I |
|
|
|
10:00
– 10:45 |
D. PLEITER |
|
|
10:45
– 11:30 |
T. Jiang The Brainnetome Atlas of Language and
its Inspiration for natural language processing |
|
|
11:30
– 12:00 |
COFFEE
BREAK |
|
|
12:00
– 12:30 |
DISCUSSION
AND CONCLUDING REMARKS |
|
|
Brain structure and
function: a neuroscience perspective II |
|
|
|
18:00
– 18:30 |
T. MANOS |
|
|
18:30
– 19:00 |
COFFEE
BREAK |
|
|
19:00
– 19:30 |
M. CANNATARO |
|
|
19:30
– 19:45 |
DISCUSSION
AND CONCLUDING REMARKS |
Tuesday, June 13th
|
Session |
Time |
Speaker/Activity |
|
|
Brain models, simulation
and brain inspired computing |
|
|
|
9:30
– 10:00 |
F. PAVONE |
|
|
10:00
– 10:30 |
J. JITSEV |
|
|
10:30
– 11:00 |
M. MIGLIORE |
|
|
11:00
– 11:30 |
COFFEE BREAK |
|
|
11:30
– 12:00 |
S. KUNKEL Extremely scalable simulation code for spiking neuronal
networks |
|
|
12:00
– 12:30 |
DISCUSSION AND CONCLUDING REMARKS |
|
|
Representation Learning,
Machine Learning |
|
|
|
17:00
– 17:30 |
P. CARLONI |
|
|
17:30
– 18:00 |
N. PETKOV |
|
|
18:00
– 18:30 |
M. BIEHL |
|
|
18:30
– 19:00 |
COFFEE BREAK |
|
|
19:00
– 19.30 |
N. STRISCIUGLIO |
|
|
19:30
- 19:45 |
DISCUSSION
AND CONCLUDING REMARKS |
Wednesday, June 14th
|
Session |
Time |
Speaker/Activity |
|
|
Building infrastructures
related for human brain research |
|
|
|
9:30
– 10:00 |
T. LIPPERT t.b.a. |
|
|
10:00
– 10:30 |
T. STERLING A Non von Neumann Architecture for General Neuromorphic
Computing |
|
|
10:30
– 11:00 |
S. YATES Modelling spiking multi-compartment neural networks at exascale |
|
|
11:00
– 11:30 |
COFFEE BREAK |
|
|
11:30
– 12:00 |
K. Amunts Decoding the multi-level brain organization - scientific and computational challenges |
|
|
12:00
– 12:30 |
DISCUSSION
AND CONCLUDING REMARKS |
|
|
Some Philosophical Issues |
|
|
|
17:30
– 18:00 |
D. MANDIC |
|
|
18:00
– 18:30 |
COFFEE BREAK |
|
|
18.30
- 19:00 |
G. GALLO and C. STANCATI The Future of the Mind: Some (Past and Present)
Philosophical Issues |
|
|
19:00
– 20:00 |
PANEL DISCUSSION Supercomputing the brain For the next 5 to 10
years, where do we want to go? What do we expect? Which ethical
problems will occur? Should humans be afraid of beyond-exascale
computers? Will future supercomputers simulate consciousness? The discussion aims to
stimulate debate and confrontation among researchers from
different areas of knowledge. |
Thursday, June 15th
|
Session |
Time |
Speaker/Activity |
|
|
TUTORIAL I |
|
|
|
9:30
– 11:00 |
D. MANDIC Hearables: In-ear EEG and vital signs monitoring of
the state of body of mind |
|
|
11:00
– 11:30 |
COFFEE BREAK |
|
|
TUTORIAL II |
|
|
|
11:30
– 13:00 |
M. BIEHL |
Friday, June 16th
|
Session |
Time |
Speaker/Activity |
|
|
GROUP OF INTEREST
MEETINGS |
|
|
|
9:30
– 11:00 |
N. PETKOV |
|
|
11:00
– 11:30 |
COFFEE BREAK |
|
11:30
– 13:00 |
GROUP OF
INTEREST MEETINGS |
Chairmen
|
SESSIONS I & 2 |
Nicolai Petkov
University of Groningen
THE NETHERLANDS
|
SESSION III |
Tianzi Jiang
Brainnetome Center, Institute of Automation
Chinese Academy of Sciences
Beijing
CHINA
|
SESSION IV |
Michele Migliore
Institute of Biophysics, National
Research Council, Palermo
ITALY
and
Department of Neuroscience
Yale University School of
Medicine
New Haven, CT
USA
|
SESSIONS V & VI |
Dirk Pleiter
Forschungszentrum Juelich
GERMANY
ABSTRACTS
|
Decoding the multi-level brain organization – scientific and
computational challenges Katrin Amunts Forschungszentrum Jülich, Germany The
human brain has a multi-level organisation and high complexity. New
approaches are necessary to decode the brain with its 86 billion nerve cells,
which form complex networks. E.g., 3D Polarized Light Imaging elucidates the
connectional architecture of the brain at the level of axons, while keeping
the topography of the whole organ; it results in data sets of several
petabytes per brain, which should be actively accessible while minimizing
their transport. Such ultra-high resolution models therefore pose massive
challenges in terms of data processing, visualisation and analysis. The Human
Brain Project creates a cutting-edge European infrastructure to address such
challenges including cloud-based collaboration and development platforms with
databases, workflow systems, petabyte storage, and supercomputers opening new
perspective to decode the human brain. |
|
Lifelong (machine) learning of drifting concepts Michael Biehl Johann Bernoulli Institute for Mathematics and Computer Science University of Groningen, The Netherlands Most
frequently, frameworks of machine learning comprise of two different stages:
First, in a training phase, a given set of example data is analysed,
information is extracted and a corresponding hypothesis is parameterized in
terms of, say, a classifier or regression system. In a subsequent working
phase, this hypothesis is then applied to novel data. For
many practical applications of machine learning this separation is convenient
and appears natural. A - by now - classical example
would be the automated classification of handwritten digits by means of a
neural network that has previously been trained from a large number of labeled input examples. Obviously,
the conceptual and temporal separation of training and working phase is not a
very plausible assumption for human and other biological learning processes. Moreover, it becomes inappropriate if the actual
task of learning, e.g. the target rule in a classification problem, changes
continuously in time. In such a situation, the learning system must be able
to detect and track the concept drift, i.e. forget irrelevant, older information
while continuously adapting to more recent inputs. In
this contribution we present a mathematical model of learning drifting
concepts in prototype-based classifiers, which are trained from high-dim.
data. Methods borrowed from statistical physics allow for the study of the
typical learning dynamics for different training strategies in the presence
of various drift scenarios. The mathematical framework is outlined and first
results are presented and discussed. |
|
Prototype-based systems in machine learning (tutorial) Michael Biehl Johann Bernoulli Institute for Mathematics and Computer Science University of Groningen, The Netherlands An
overview is given of prototype-based systems in machine learning. In this
framework, observations, i.e., data, are stored in terms of typical
representatives. Together
with a suitable measure of similarity, such systems can be employed in the
context of unsupervised and supervised analysis of potentially high
dimensional, complex datasets. We discuss basic schemes of unsupervised
Vector Quantization (VQ) as well as the so-called Neural Gas (NG) approach and Kohonen’
s topology-preserving Self-Organizing Map (SOM). Supervised
learning in prototype systems is exemplified in terms of Learning Vector
Quantization (LVQ). Most frequently, the familiar Euclidean distance serves
as a dissimilarity measure. We present extensions of the framework to
nonstandard measures and give an introduction to the use of adaptive
distances in so-called relevance learning schemes. To
a large extent, this tutorial will be along the lines of our review article: Michael
Biehl, Barbara Hammer, Thomas Villmann Prototype-based
methods in machine learning Advance
Review, WIREs Cognitive Science 2016 available
online: http://onlinelibrary.wiley.com/doi/10.1002/wcs.1378/abstract doi:
10.1002/wcs.1378 |
|
Methods and techniques for recognizing emotions: sentiment analysis
and biosignal analysis with applications in
neurosciences Chiara Zucco, Barbara
Calabrese, Mario Cannataro Department of Medical and Surgical Sciences Data Analytics Research Center University “Magna Grćcia” of Catanzaro 88100 Catanzaro, ITALY {czucco, calabreseb,
cannataro}@unicz.it Feelings
and emotions are related to biological, social and cognitive aspects of each
person’s life. With the advent of wearable devices and social networking
platforms, people began to monitor their lives on a daily basis not only by recording
physical signals (such as heart rate, steps, etc.), but also by expressing their
emotions through text, images, video, audio, emoticons and tags. Deriving
meaning from this vast amount of data is therefore a topic that, in recent
years, has received a growing interest in both industrial and academic
research. In particular, new
computational technologies such as sentiment analysis and affective computing
have found applications in many fields of knowledge such as marketing,
politics, social sciences, cognitive sciences, medical sciences etc. Such
technologies aim to automatically extract emotions from heterogeneous data
such as text, images, audio, video, and a pletora of
biosignals such as voice, facial expression, EEG,
near-infrared spectroscopy, etc. The
paper introduces main concepts of sentiment analysis and affective computing
and presents an overview of the main methodologies and techniques used to
recognize emotions from the analysis of various data sources such as text,
images, voice signals, EEG, near-infrared spectroscopy. Finally, the paper
discusses various applications of those techniques to neurosciences. |
|
Multiscale simulations of key molecular events in G-protein coupled receptors-based
neuronal cascades Paolo Carloni Forschungszentrum Jülich, Germany G-protein
coupled receptors (GPCRs) regulate fundamental brain processes, including
neurotransmission. Here I will
illustrate recent studies from our lab aimed at predicting structure and
energetics of agonists binding to GPCRs. We will present in particular hybrid
coarse-grain/molecular mechanics applications which may be particularly
useful for resolution models of these proteins. We will conclude with a brief
overview of our investigations of
GPCRs-based neuronal cascades within
the Human Brain Project. |
|
The Future of the Mind: Some (Past and Present)
Philosophical Issues Giusy Gallo e Claudia Stancati Dipartimento di Studi Umanistici, Universitŕ della Calabria Before
being investigated by the scientific research, the issues concerning our
knowledge of the world, the use of words and sentences, the acquisition of
language and the definition of the self are philosophical problems. Then,
philosophy is a sort of historical mirror of the issues mentioned above. We
will discuss the following open questions (see Gary Marcus, “The
Computational Brain”, in The future of the brain, pp. 212-214) showing for
each of them the ideas of two thinkers of past and contemporary philosophy: 1
If the brain is not a von Neumann stored program machine, what kind of
information processor is? How does the brain manage to be so coordinated in
the absence of a central clock? Is there a kind of neuronal algebra, a set of
operations that works on arbitrary values stored in synapses? Who denies that
the brain is a computer, has a reasonable alternative? 2
If in certain occasions the human brain behave like a von Neumann computer
(conscious and deliberate rule application) what kind of neural systems can
support the versatility of our cognition in other domains where knowledge and
instructions are both less explicit? It seems human brain is a hybrid system:
both digital and analogic. To understand how human mind works we should start
thinking about the principle of compositionality. 3
How does the brain implement variable binding? 4
Is there a single canonical form of computation or is there a wide range of
operations? 5
What format(s) does the brain use for encoding the information? Like ASCII,
JPEG or GIF. We know something about space and motor space but very little
about word, sentence, images, melody. 6
Why does the brain contain so much diversity and details in order to do
things which a wide but simple neural network can’t do? |
|
The Brainnetome Atlas of Language and its
inspiration for Natural Language Processing Tianzi Jiang1, 2, 3, 4, 5, 6 1 Brainnetome Center, Institute of Automation, Chinese Academy of
Sciences, Beijing 100190, China 2 National Laboratory of Pattern Recognition,
Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China 3 CAS Center for Excellence
in Brain Science and Intelligence Technology, Institute of Automation,
Chinese Academy of Sciences, Beijing 100190, China 4 Key Laboratory for NeuroInformation
of the Ministry of Education, School of Life Science and Technology,
University of Electronic Science and Technology of China, Chengdu 625014,
China 5 The Queensland Brain Institute, University of
Queensland, Brisbane, QLD 4072, Australia 6 University of Chinese Academy of Sciences, Beijing
100190, China The
human brain atlas plays a central role in neuroscience and clinical practice,
and is a prerequisite for studying brain networks and cognitive functions at
the macroscale. Using non-invasive multimodal neuroimaging techniques, we
have designed a connectivity-based parcellation framework
to build the human Brainnetome Atlas, which identifies
the subdivisions of the entire human brain, revealing the in vivo connectivity
profiles. This new brain atlas has the following four features: (A) It
establishes a fine-grained brain parcellation scheme
for 210 cortical and 36 subcortical regions with a coherent pattern of anatomical
connections; (B) It supplies a detailed map of anatomical and functional connections;
(C) it decodes brain functions using a meta-analytical approach; and (D) It
will be an open resource for researchers to use for the analysis of whole
brain parcellations, connections, and functions.
The human Brainnetome Atlas could constitute a
major breakthrough in the study of human brain atlas and provides the basis
for new lines of inquiry about the brain organization and functions. It could
be regarded as a start point, which will enable the generation of future
brain atlases that are more finely, defined and that will advance from single
anatomical descriptions to an integrated atlas that includes structure,
function, and connectivity, along with other potential sources of information.
In
this lecture, we first give an introduction on the human Brainnetome
atlas. Then we demonstrate what new knowledge on the brain regions of
language we can obtain with human Brainnetome atlas.
Firstly, we defined a convergent posterior anatomical border for Wernicke’s area
and indicated that the brain’s functional subregions
can be identified on the basis of its specific structural and functional
connectivity patterns. Secondly, we revealed a detailed parcellation
of Broca’s region on the basis of heterogeneity in
intrinsic brain activity, and investigated cross-cultural consistency and
diversity in intrinsic functional organization of Broca’s
Region. Finally, we give a brief introduction on the potential inspiration
for natural language processing. |
|
What computations are ‘brain-inspired’? - A view on neural information
processing, functionality and learning Jenia Jitsev Juelich Research Center, Germany There
has been a long tradition of casting models of information processing which
entail elementary generic computing units arranged in multiple stacked layers
to perform cascades of transformations of the incoming input as neurally inspired, or brain-like. However, what justifies
casting a particular model neurally inspired is
quite arbitrary and inconsistent. Often, a very limited set of properties of
biological neural networks, like their hierarchical processing organization
or spiking of single neurons, is taken to back up claim for neural
plausibility, while completely ignoring vast range of other presumable
relevant properties, e.g diversity of neuronal
single cell dynamics, short-term synaptic plasticity or signal processing in
active dendrites, to name only few. The
same concerns taking into account different architectural features of brain
networks, like different loops through subcortical structures, e.g thalamus or basal ganglia, or different fundamental
modes of brain operation like sleep. Moreover, adding clearly biological
implausible features into such network models to enforce a desired
functionality further obscures terminology of brain-inspired computation.
Furthermore, when arguing for certain brain-like functionality, the input and
the tasks the networks are demonstrating their capabilities on have often
very narrow and artificial character not plausible in a real world setting
where nervous systems are operating. Here,
to provide a perspective for a consistent framework of building neurally inspired information processing models, I would
like to put forward a view that in face of haunting neural diversity on the
one hand and still quite limited techniques to record large scale activities
across multiple spatial scales from the brain on the other hand, it is
necessary to establish solid foundation for functional and computational essense of brain phenomenology before attempting to
construct full scale neural network models. Working out functional and
computational essense means here specifying the
generic type of problems a brain has to solve in natural environment,
together with the type of computations and the type of complex real-world
input available to the brain. Error driven learning, defined within a closed
sensory-motor loop of forming and correcting predictions about sensory input
and hidden variables that are most likely causing it, is one potential
candidate framework to establish such a generic functional description for
brain-like information processing. Only then it will become indeed possible
to interpret properly different neurophysiological observations of biological
neural substrate and to come up with basic canonical models where both neural
and functional properties will reflect the principles of brain-like
information processing to satisfactory degree. |
|
Extremely scalable simulation code for spiking neuronal networks Susanne Kunkel KTH Royal Institute of Technology, Stockholm, Sweden Today’s
simulation code for spiking neuronal networks scales well from laptops to
supercomputers [1, 2] and it supports a wide range of models (e.g. [3, 4]).
In my talk, I will discuss the different requirements that such simulation
software needs to meet for different regimes of number of processes. Using
the NEST simulator [5] as an example, I will discuss the limiting factors
that prevent the current simulation technology [1] from scaling beyond the peta scale. I will introduce recent work on algorithms and
data structures that overcome these limitations, and I will also present optimizations
that ensure that the new simulation code is still efficient on small compute
clusters. Finally, I will show that the novel technology is not only expected
to scale well into the post-peta scale regime but
that it also decreases memory usage and run time of spiking neuronal network
simulations on current supercomputers. [1]
Kunkel S, Schmidt M, Eppler JM, Masumoto G,
Igarashi J, Ishii S, et al. (2014) Spiking network simulation code for petascale computers. Front.
Neuroinform. 8, 78. [2]
Ippen T, Eppler JM, Plesser HE and Diesmann M
(2017) Constructing Neuronal Network Models in Massively Parallel
Environments. Front. Neuroinform. 11:30. [3]
Hahne J, Helias M, Kunkel
S, Igarashi J, Bolten M, Frommer
A and Diesmann M (2015) A unified framework for
spiking and gap-junction interactions in distributed neuronal network
simulations. Front. Neuroinform.
9:22. [4]
Diaz-Pier S, Naveau M, Butz-Ostendorf
M and Morrison A (2016) Automatic Generation of Connectivity for Large-Scale
Neuronal Network Models
through Structural Plasticity. Front. Neuroanat. 10:57. [5]
Gewaltig MO and Diesmann M
(2007) NEST (NEural Simulation Tool). Scholarpedia 2, 1430. |
|
Complexity Science for Physiological Data Danilo Mandic Imperial College London, United Kingdom The
complexity loss theory states that structural complexity of the responses of
living organism decreases with constraints, such as ageing or illness. Yet,
methods to quantify the degree of this loss of structural complexity are few
and far between. This talk focuses on multivariate multiscale entropy, our
new tool for the quantification of the complexity of physiological data.
After introducing the method and validating its performance on synthetic
data, the utility of the method in general neuroscience is demonstrated over
example on multimodal body responses (EEG, heart rate, respiration) in both
clinical and out-of-clinic conditions. |
|
Hearables: In-ear EEG and vital signs monitoring of
the state of body of mind Danilo Mandic Imperial College London, United Kingdom This
tutorial brings together three main aspects of future wearable health
technology: (i) adequate signal processing
algorithms, (ii) miniaturised hardware for 24/7 continuous monitoring of the
mind and body, and (iii) development of applications for use in natural environments.
Based upon our 10 years of experience in human-computer interface, we will
bring together the latest advances in multiscale signal processing,
complexity science, and their application in real-world scenarios for
next-generation personalised healthcare, such as sleep, fatigue and stress.
Our particular emphasis will be on solutions to the challenges posed by the
imperfect but ultra-wearable, unobtrusive, and discreet sensors. To this end, insights into the biophysics
of the generation and acquisition of human physiological responses will be
used as a foundation, and indeed the motivation, for the multiscale signal
processing algorithms covered. We will also discuss opportunities in
multi-person behavioural science, enabled by our own wearable sensing
platforms, such as vital sign monitoring from inside the ear canal (ECG, EEG,
respiration, etc.) and our miniaturised biosignal
acquisition unit. Biography:
Danilo P. Mandic is a Professor in signal
processing with Imperial College London, UK, and has been working in the
areas of adaptive signal processing and bioengineering. He is a Fellow of the
IEEE, member of the Board of Governors of International Neural Networks
Society (INNS), member of the Big Data Chapter within INNS and member of the
IEEE SPS Technical Committee on Signal Processing Theory and Methods. He has
received five best paper awards in Brain Computer Interface, runs the Smart Environments
Lab at Imperial, and has more than 300 publications in journals and
conferences. Prof Mandic has received the President
Award for Excellence in Postgraduate Supervision at Imperial. |
|
Improving long lasting anti-kindling effects via coordinated reset
stimulation frequency mild modulation Thanos Manos1, Magteld
Zeitler1, Simon Eickhoff1,2, Peter A. Tass3 1Institute of Neuroscience and Medicine (INM-7),
Research Center Juelich, Juelich, Germany 2Institute for Systems Neuroscience, Medical Faculty,
Heinrich-Heine University Dusseldorf, Germany 3Department of Neurosurgery, Stanford University,
Stanford, CA, USA Keywords: coordinated reset, desynchronization, spike
time-dependent plasticity, anti-kindling, rapidly varying sequence, slowly
varying sequence, stimulation frequency, stimulation intensity Abstract Several
brain diseases are characterized by abnormally strong neuronal synchrony.
Coordinated Reset (CR) stimulation [1,2] was computationally designed to
specifically counteract abnormal neuronal synchronization processes by
desynchronization. In the presence of spike timing-dependent plasticity
(STDP) [3] this leads to a decrease of synaptic weights and ultimately to an
anti-kindling [4], i.e. unlearning of abnormal synaptic connectivity and
abnormal neuronal synchrony. The long-lasting desynchronizing impact of CR
stimulation has been verified in pre-clinical and clinical proof of concept
studies (e.g. [5]). However, as yet it is unclear how to optimally choose the
CR stimulation frequency, i.e. the repetition rate at which the CR stimuli
are delivered. This
work presents a first computational study on the dependence of the long-term
outcome on the CR stimulation frequency in neuronal networks with STDP while
a conductance-based Hodgkin-Huxley neuron model for the description of an
ensemble of spiking neurons was used. From a clinical standpoint, it is
desirable to achieve an anti-kindling already with stimulation durations as
small as possible. For this reason and due to CPU time constraints, we have
chosen a certain range of stimulation durations, where we were able to
achieve a reasonable success rate (i.e. anti-kindling) at least for suitable
stimulation frequencies. For a representative stimulation duration of this
kind, we have thoroughly varied the stimulation frequency while we have
preliminary evidence that even for longer stimulation durations the picture
does not change much. For this purpose, CR stimulation was applied with
Rapidly Varying Sequences (RVS) [4] in a wide range of stimulation
frequencies and intensities. A similar analysis was performed with a different
type of CR signal, the Slowly Varying Sequences (SVS) CR [6]. We
show that when comparing the two different CR signals, RVS CR turns out to be
more robust against variations of the stimulation frequency; however, SVS CR
can obtain stronger anti-kindling effects [7]. In cases where the initial
combination of CR intensity and frequency did not perform efficiently for the
majority of different network initializations, we implement three plausible
therapy-like stimulation protocols, which aim to ameliorate the long-lasting
effects. The first one prolongs the CR on period before ceasing it
completely, the second one consists of repetition of CR on and off
trial-periods with the same fixed CR frequency while the third one
incorporates a control mechanism monitoring the degree of synchronization at
the end of the CR off period and adjust the CR stimulation period for the
following trials via a mild modulation. When comparing these three
approaches, the last one not only manages to induce global desynchronization
(for all networks tested) but also shows pronounced robustness among
different signals and network dependent variations [8]. These findings can be
implemented into stimulation protocols for first in man and proof of concept
studies aiming at further improvement of CR stimulation. Extending
this work towards the integration of MRI-based neuroimaging and the analyses
of inter-individual variability of neuronal dynamics in larger populations,
together with Professor Simon Eickhoff we plan to
use large-scale mathematical models for the description of the brain-region
mean activity. For this purpose, and in order to investigate the dynamics of
resting states accompanied with appropriate preprocessing,
we will use The Virtual Brain platform
as well as Juelich’s Supercomputer Center facilities. References [1] Tass, P.A. (2003a). A model of desynchronizing deep brain
stimulation with a demand-controlled coordinated reset of neural
subpopulations. Biol. Cybern. 89:
81-88. [2] Tass, P.A. (2003b). Desynchronization by means of a
Coordinated Reset of neural sub-populations. Prog.
Theor.Phys. Suppl. 150: 281-296. [3]
Gerstner, W., Kempter, R., Van Hemmen,
J.L, and Wagner, H. (1996). A neuronal learning rule for sub-millisecond
temporal coding. Nature 383: 76-78. [4] Tass, P.A., and Majtanik, M.
(2006). Long-term anti-kindling effects of desynchronization brain
stimulation: a theoretical study. Biol. Cybern.
94: 58-66. [5] Adamchic, I., Hauptmann, C., Barnikol,
U.B., Pawelcyk, N., Popovych,
O.V., Barnikol, T., Silchenko,
A., Volkmann, J., Deuschl, G., Meissner, W., Maarouf, M., Sturm, V., Freund, H.-J., Tass, P.A. (2014). Coordinated Reset has lasting
aftereffects in patients with Parkinson’s Disease. Mov. Disord. 29: 1679-1684. [6] Zeitler, M. and Tass, P.A.
(2015) Augmented brain function by coordinated reset stimulation with slowly
varying sequences. Front. Syst. Neurosci. 9:
49. [7]
Manos, T., Zeitler, M., and Tass,
P.A. (2017). Effect of stimulation frequency and intensity on long-lasting
anti-kindling. To be submitted. [8] Manos, T., Zeitler,
M., and Tass, P.A. (2017). Improving long lasting
anti-kindling effects via coordinated reset stimulation frequency mild
modulation. To be
submitted. |
|
The implementation of brain-inspired cognitive architectures using
large-scale realistic computational models Michele Migliore Institute of Biophysics, National Research Council,
Palermo, Italy Department of Neuroscience, Yale University School
of Medicine, New Haven, CT, USA Understanding
the neural basis of brain functions and dysfunctions has a huge impact on a
number of scientific, technical, and social fields. Experimental findings
have given and continue to give important clues at different levels, from
subcellular biochemical pathways to behaviors involving
many brain regions. However, most of the multi-level mechanisms underlying
the cognitive architecture of the involved brain regions are still largely unknown
or poorly understood. This mainly depends on the practical impossibility to
obtain detailed simultaneous in vivo recordings from an appropriate set of cells,
making it nearly impossible to decipher and understand the emergent properties
and behavior of large neuronal networks. We are
addressing this problem using large-scale computational models of biologically
inspired cognitive architectures, which require substantial resources for
storage, computing, and scientific visualization that can be available only
through large international research infrastructures. In this talk, I will
present and discuss the main results and techniques, used in my lab and
within the Human Brain Project, to design and use realistic models of neurons
and networks implemented following their natural 3D structure. To illustrate
our approach and its relevance to understand computational and functional
processes, I will show the results obtained for the hippocampus and the
olfactory bulb. The main goal is to uncover the mechanisms underlying higher
brain functions, helping the development of innovative therapies to treat
brain diseases. Through movies and interactive simulations, I will show how
and why the dynamical interaction among neurons can predict new results and
account for a variety of puzzling experimental findings. Selected References Migliore
R, De Simone G, Leinekugel X, Migliore
M, (2016) The possible consequences for cognitive functions of external
electric fields at power line frequency on hippocampal CA1 pyramidal neurons,
Eur. J. Neurosci, doi:
10.1111/ejn.13325. Migliore
M, Cavarretta F, Marasco A,
Tulumello E, Hines ML, Shepherd GM. (2015) Synaptic
clusters function as odor operators in the
olfactory bulb, Proc Natl Acad Sci
U S A. 112(27):8499-504. Bianchi
D, De Michele PD, Marchetti C, Tirozzi
B, Cuomo S, Marie H, Migliore M (2014), Effects of
increasing CREB-dependent transcription on the storage and recall processes
in a hippocampal CA1 microcircuit, Hippocampus. 24:165-77. |
|
A multivariate analysis of a brain disease Francesco Pavone University of Florence, Physics Department, LENS, Italy Neuro-rehabilitative
therapy is the most effective treatment for recovering motor deficits in
stroke patients. Nevertheless, the neural bases of recovery associated with
rehabilitative intervention is debated. Here, we demonstrated how the
multivariate analysis of brain parameters, both from functional and
morphological point of view, is able to depict the damage and its
rehabilitation process on different perspectives IN
particular, we show the synergic action of robotic rehabilitation and
transient inhibition of the contralesional motor
cortex molded cortical plasticity at multiple
scales. By longitudinal imaging of cortical activity while training on a robotic
platform for mouse rehabilitation, we demonstrated progressive recovery of motor
map dedifferentiation and rise of a stronger and faster calcium response in
the peri-infarct area. The coupling of the spared
cortex to the injured hemisphere was reinforced, as demonstrated by our
all-optical approach. Alongside, a profound angiogenic
response accompanied the stabilization of peri-infarct
micro-circuitry at the synaptic level. The present work, by combining optical
tools of visualization and manipulation of neuronal activity, provides the first
in vivo evidence of the deep impact of rehabilitation on cortical plasticity. Finally,
the importance of deep leaning, and more in general machine learning, is
demonstrated in the analysis and process of informations
obtained. |
|
Representation learning with trainable COSFIRE filters Nicolai Petkov University of Groningen, Netherlands In
order to be effective, traditional pattern recognition methods typically
require a careful manual design of features, involving considerable domain
knowledge and effort by experts. The recent popularity of deep learning is
largely due to the automatic configuration of effective early and
intermediate representations of the data presented. The downside of this
approach is that it requires a huge number of training examples and a major
computational effort. Trainable
COSFIRE filters are an alternative to deep networks for the extraction of
effective representations of data. Such a filter is configured by the
automatic analysis of a single pattern. The highly non-linear filter response
is computed as a combination of the responses of simpler filters, such as Difference
of (color) Gaussians or Gabor filters, taken at different
positions of the concerned pattern. The identification of the parameters of
the simpler filters that are needed and the positions at which their
responses are taken is done automatically. We call this method Combination of
Shifted Filter Responses - COSFIRE. An advantage of this approach is its ease
of use as it requires no programming effort and little computation – the parameters
of a filter are derived automatically from a single training pattern. Hence,
a large number of such filters can be configured effortlessly and selected
responses can be arranged in feature vectors that are fed into a traditional
classifier. This
approach is illustrated by the automatic configuration of COSFIRE filters
that respond to randomly selected parts of many handwritten digits. We
configure automatically up to 5000 such filters and use their maximum
responses to a given image of a handwritten digit to form a feature vector
that is fed to a classifier. The COSFIRE approach is further illustrated by
the detection and identification of traffic signs and of sounds of interest
in audio signals. The
COSFIRE approach to representation learning and classification yields
performance results that are comparable to the best results obtained with
deep networks but at a much smaller computational effort. Notably, COSFIRE
representations can be obtained using numbers of training examples that are
many orders of magnitude smaller than those used by deep networks. About
the speaker: Nicolai
Petkov is professor of computer science with a
chair in intelligent systems at the University of Groningen since 1991. In
the period 1998-2009 he was scientific director of the Institute for
Mathematics and Computer Science. He applies machine learning and pattern
recognition to various problems. www.cs.rug.nl/is |
|
2D Gabor functions for modeling
simple and complex cells in visual cortex. Use in image processing and computer
vision Nicolai Petkov University of Groningen, Netherlands 2D
Gabor functions are introduced and their relation to the properties of simple
cells in the primary visual cortex is given. Their properties in the space and frequency
domain and the role of different parameters are discussed. Typical use of
Gabor functions in image processing and computer vision, such as edge
detection and texture characterization is considered. About
the speaker: Nicolai
Petkov is professor of computer science with a chair
in intelligent systems at the University of Groningen since 1991. In the
period 1998-2009 he was scientific director of the Institute for Mathematics
and Computer Science. He applies machine learning and pattern recognition to
various problems. www.cs.rug.nl/is. |
|
New HPC Architectures and Technologies for Brain Research Dirk Pleiter Forschungszentrum Juelich, Germany During
the early phase of the Human Brain Project a pre-commercial procurement (PCP)
had been launched for procuring research and development services. The goal
was to have commercial operators creating solutions that will augment their
HPC product roadmap and make these more suitable for computational
neuroscience applications. The project focussed on integration of dense
memory, scalable visualisation and dynamic resource management. In this talk
we will present and discuss the outcomes. Furthermore, we will introduce the
pilot systems that had been delivered by the PCP contractors for enabling
testing of their solutions. |
|
A Non von Neumann Architecture for General Neuromorphic Computing Thomas Sterling Indiana University, USA Brain
inspired computing refers both to a possible of means of achieving advanced
computing through methods and structures analogous to those of the human
brain and to computations intended to emulate or simulate operational properties
observed of (and in a sense, by) the human brain. There is the potential for significant
overlap of the two with computers made up of brain-like components employed
to model the human brain itself. The technical approach presented here
reflects an advanced cellular automata approach to neuromorphic computing
both as a means to achieve computational techniques like machine learning and
to use the same class of platforms possibly for brain simulation. As
previously reported, the ParalleX execution model is
a class of Asynchronous Multi-Tasking abstract architectures that improve
efficiency and scalability through dynamic adaptive computations. ParalleX has been embodied in the family of HPX runtime systems
(including work at LSU and IU) for proof of concept, first reduction to
practice, and prototypes. At the low-level hardware structure, cellular-like
organizations named Continuum Computer Architecture (CCA) can be efficiently
employed for time-varying irregular graph related computations. Unlike
classical cellular automata, CCA incorporates mechanisms that efficiently
support the ParalleX parallel computational model
and key functions in support of dynamic graph operations. A key property of
this approach to neuromorphic computing is that the communications among
cells are packet switched through worm-hole routing rather than line switched
as used by other methods. This non von Neumann approach should provide the
properties of general purpose software control while delivering hardware
enabled performance. It is anticipated that a single semiconductor die can
incorporate on the order of 214 primitive elements (i.e., “fontons”) with a peak operational performance of 1 exaops within 1 cubic meter. Throughout the presentation,
questions and comments from the audience will be welcomed. |
|
Bio-inspired representation learning in pattern
recognition Nicola Strisciuglio University of Groningen, The Netherlands Since
when very young, we can quickly learn new concepts and distinguish between
different kinds of object or sound. If
we see a single object or hear a particular sound, we are then able to
recognize such sample and even different versions of it in other scenarios. We
learn and store representations of the world and use them to detect and
understand it. Representation
learning is an important aspect of pattern recognition. In the recent years,
with the development of deep learning, it raised a large research interest.
The aim of techniques for representation learning is to construct effective
and reliable features directly from training samples instead of engineering
hand-crafted representations, which usually require extensive domain
knowledge. Some approaches to representation learning are based on machine
learning techniques, while other exploit the knowledge about biological and
natural systems. In
this presentation, I will discuss about the concept and techniques for
representation learning in pattern recognition and present two approaches,
COSFIRE and COPE, which take inspiration from some functions of the human
visual and auditory systems. I will present the basic idea of COSFIRE and
COPE features, how they are configured from training samples and the results
achieved by their use in several image and sound analysis applications. |
|
Modelling spiking multi-compartment neural networks at exascale Sam Yates CSCS, ETH Zurich, Switzerland The
simulation of increasingly large networks of highly detailed neuron models demands
in turn the use of large scale compute systems. The
ambition to exploit these systems — petascale today
and exascale when available — has implications for
the simulation software: implementations must make efficient use of diverse
hardware platforms; network construction costs and communication overheads
must remain constrained as models grow in scope and complexity. The
HPC landscape is changing rapidly, with the adoption of GPU accelerators and “many
core” processors such as Intel's Xeon Phi line. Achieving good utilization of
these diverse architectures is becoming increasingly difficult for the
developers and maintainers of simulator software. We
present our work on xmc, a HPC library for neural
network simulations that addresses these challenges. We describe the scalable
architecture of the library and show some initial benchmarks of simulations
based on this platform. |