Transcript PolandTorun

COST Action B27, WG1
Theoretical Study on Oscillation & Cognition,
Polish contributions
Reported by Włodzisław Duch
(Google: Duch)
Department of Informatics,
Nicolaus Copernicus University,
Torun, Poland
School of Computer Engineering,
Nanyang Technological University,
Singapore
Notes
Summer time ... not all responded on a short notice.
Most people do theory and applications and work in several
places ... I will talk about activities of 4 groups:
• Andrzej Cichocki, Warsaw Univ. Technology
& RIKEN Brain Science Institute, Wako-shi
• Rafał Bogacz, Bristol, Uni. Wrocław & Princeton
• Wiesław Kamiński, Maria Curie-Skłodowska University,
Lublin
• Włodzisław Duch, Nicolaus Copernicus University
& Nanyang Technological University, Singapore
Laboratory for
Advanced Brain Signal Processing
Andrzej Cichocki
http://www.bsp.brain.riken.jp/~cia/
RIKEN, Brain Science Institute, JAPAN
& Warsaw University of Technology, POLAND
Laboratory for Advanced Brain Signal Processing
Riken Brain Science Institute, Japan
Research mission and central research interest:
The laboratory for Advanced Brain Signal Processing is
focused on developing novel and state of the art
methods to:
• extract, detect, recognize,
• find functional connectivity
• classify brain signals
and to use the insights gained to build intelligent feature
extraction systems for Early Detection and Classification of
Dementia, especially Alzheimer Disease (AD), evaluation of
aging of the brain using Blind Signal Processing (BSP) and
Time Frequency Representation (TFR) of EEG and fMRI/PET.
Research Projects of the
Laboratory for Advanced Brain Signal Processing
Database
Experiments, collecting and preprocessing
EEG, EOG, EMG , PET, fMRI, MUR data
Modeling
Olfactory,
Auditory S
Diagnosis
of Dementia,
AD
Electronic Nose
Electronic Ears
BLIND SIGNAL
PROCESSING
MACHINE LEARNING
DATA MINING
Brain
Computer
Interface
Analysis of
EEG/ERP
Detection,
Enhancement
Classification,
Extraction. Functional
Connectivity
Analysis
MUR
Intelligent Communication
Human with machine
Spike Sorting
Information Retrieval
Clustering
Research Objectives
One of the main objective of the Laboratory is to develop and apply novel
blind signal processing (BSP) and Machine Learning (ML) algorithms and
methods including: Sparse Components Analysis (SCA), Time-
Frequency Component Analyzer (TFCA), Independent
Components Analysis (ICA), Blind Deconvolution - Equalization and
Hierarchical Clustering to analyze multi-sensory, multi-modal
biomedical signals, especially high density array EEG signals.
High Density Array EEG Recording/Analysis
Systems in LABSP, RIKEN BSI
LABSP Research Projects
1. Developments and Implementation of Novel Blind Signal
Processing and Machine Learning Techniques for Analysis,
Finding Functional Relationships and Modeling of Brain
Signals.
2. Intelligent Communication between Human Brain and Machine
- Development of Software/Hardware for Human/Brain
Computer Interface (H/BCI) and Classification of Various
Mental States.
3. Early Detection and Classification of Dementia, especially
Alzheimer Disease (AD) using Blind Signal Processing (BSP)
and Time-Frequency Representation (TFR) of EEG and Other
Neuroimaging Techniques.
4. Modeling Some Aspects of Auditory System and Olfactory
System: Contribution to Development of Electronic Ears and
Electronic Nose – Artificial Olfaction.
Unique Results
1.
Development of novel models for BSP (State space, Kalman
filter, multilayer, recurrent NN, BSE NN using linear predictability).
2.
Development, implementation, integration and
theoretical analysis of new associative learning
algorithms for ICA, SCA, BSE, MBD, NMF and SPCA.
3.
Applications of BSP algorithms to real-world problems
1.
2.
3.
4.
5.
Early detection of Alzheimer’s disease (Clinical Neurophysiology)
Analysis of high density array EEG data (extraction of unique
components and elimination of dependent artifacts, investigation
validity and reliability, improvements in source localization)
Reduction of artifacts in simultaneous recording EEG and fMRI
Speech separation, enhancement and modeling auditory cortex
Support clinical diagnosis of brain death using ICA.
Procedure for extracting markers of AD
Raw
EEG
Data
EEG unit
Clean
EEG
Data
Preprocessing:
Artifacts removal;
Denoising;
Filtering;
Model reduction
Components
BSS/BSE:
ICA, SCA, NMF,
TFCA
(W)
Signal
subspace
Ranking
and
clustering
of components
Significant
markers
Enhanced
EEG
Backprojection
+
(W )
Noise
subspace
Enhanced
EEG
AD/MCI
Wavelet TFR,
Sparse bump
modeling
Feature
Extraction
Classification
Diagnosis
Neural network
Normal
Processing flow of the developed method. The main novelty lies in ordering and
selection of only few significant AD markers (components), back-projecting
(deflation) of these components on the scalp level and processing them in the
time frequency domain using approximated sparsification. Advanced pattern
recognition and machine learning techniques are applied for classification and
analysis of the data.
Rafał Bogacz
Bristol/Princeton/Wrocław
Theory of Event Related Potentials (ERP)
• ERPs are computed by averaging EEG signals over many
trials, time locked to an event in psychological experiment
(e.g. stimulus presentation).
• Should ERP’s be regarded as uncorrelated with the
background EEG, or generated by the event-related
reorganization of this ongoing rhythmic activity?
• Detection of phase resetting in electroencephalogram;
paper with Nick Yeung, Clay Holroyd, Jonathan D. Cohen.
Theories of ERP origin
Phase resetting
and enhancement
Pure phase
resetting
…
…
Individual
EEG epochs
“Classical view”
(phasic peak)
…
Averaged
ERP
Evaluation of methods
• We evaluated a number of methods previously used to
support the phase-resetting theory of ERP origin.
• We generated artificial EEG signals by superimposing
phasic peak on noise (according to classical view).
• When applied to the simulated data, the methods in
question produced results that have been previously
interpreted as evidence of synchronized oscillations,
even though no such synchrony was present.
Wiesław Kamiński, Grzegorz Wójcik
Division of Complex Systems and Neurodynamics
Institute of Computer Science, Maria Curie-Sklodowska
University, Lublin, Poland
•
•
•
•
•
Neurocomputing
Brain and Visual System Modelling
Parallel Processing
Physical analysis
Software development and Visualisation
[email protected]
[email protected]
Brain and Visual System Modelling
• Modeling and investigation of large biological neural
networks
• Visual systems simulations and models of cortex
• Dynamical analysis and applications of Artificial Neural
Networks (ANN)
• Liquid State Machines (LSM), etc.
Analysis Based on Physics
• Thermodynamic and statistical physics methods in
network’s dynamics analysis
• Analysis based on informational theory
• Self Organising Criticality (SOC) investigations
• Chaos theory and applications
Parallel Processing/Visualization
• Grids and large clusters for simulations
• Adaptation of GENESIS/MPGENESIS simulators for MPI
environment
• Development of visualisation methods for the cortex
dynamics and comparison with experimental results.
• Participation in the CLUSTERIX project (National Linux
Cluster, more than 800 Itanium processors).
Włodzisław Duch & Co (Google: Duch)
Department of Informatics, Nicolaus Copernicus University,
Torun, Poland, and School of Computer Engineering, Nanyang
Technological University (NTU), Singapore
Done many smaller projects on:
• Hebbian associative memories with
chaotic itinerancy and large Lyapunow
exponents for mixed pattern separation
(P. Matykiewicz).
• Visualization of trajectories in such networks and stability
analysis of locally Hopfield nets with highly correlated patterns
(F. Piękniewski, L. Rybicki)
• Fuzzy symbolic dynamics for simplification of neurodynamics.
• A-life biots based on Boltzman machines (L. Rybicki)
• Global Brain Simulations – just starting ...
• Cognitive architectures, integration of perception with
cognition – just starting ...
Attention-Based Artificial Cognitive Control
Understanding System (ABACCUS)
First attempt: large EU integrated project, with 9 participants:
King’s College London (John G. Taylor, coordinator).
New version: BRAin as Complex System (BRACS), on a smaller
scale, more focused on simulations and understanding the
principles of complex brain-like information processing.
The time of large scale global brain simulations has come!
• Computer speeds have just reached brain power (about 1016
binop/s), but computers are far from brain’s complexity/style.
• Science: understand how high-level cognition arises from lowlevel interactions between neurons, build powerful research tool;
to understand complex systems is to be able to build them.
• Practical: humanized, cognitive computer applications require a
brain-like architecture (either software or hardware) to deal with
such problems efficiently; it is at the center of cognitive robotics.
Scheme of the brain ...
High-level sketch of the brain structures, with connections based
on different types of neurotransmiters marked in different colors.
BRACS Assumptions & Goals
• Assumption: gross neuroanatomical brain structure is critical for
its function, therefore it should be preserved.
• Should be founded on neuro-scientific understanding of
attention and the sensory and motor systems it controls,
development in children, simplified modeling, computer power.
• Fusion of the appropriate brain-based models, guided by the
overall architecture of the brain and developmental learning
stages should lead to high-level cognitive processing.
• Develop an attention control systems for focusing in sensory
surveillance tasks, and for image searching.
• Development of control structures for autonomous machines.
• Create its own goals in an autonomous fashion.
• Darwin VII small robot (G. Edelman) works with 53K mean firing
+phase neurons, 1.7 M synapses, modeling 28 brain areas and
achieving sensorimotor integration; our project is larger and
more structured, hopefully higher cognitive functions emerge ...
Sketch of the BRACS system
Computational Platform, Simulation Environment and Integration
Neuroscience and Development
Vision
Speech
Tactile
Memory
System
Motor
Control
Reasoning
System
Feedback
Attention Control
Value Maps
Drive and Intrinsic
reward system
Atomization
system
Learning of
PFC goals
Working
Memory
Action/Object
reward system
Rough sketch of the BRACS system, based on simplified spiking
neurons.