Current Trends inMachine Learning for Signal Processing
(MLSP)
Tulay AdalıChair, MLSP TC
Machine Learning for Signal Processing LabUniversity of Maryland Baltimore County
Baltimore, MD 21250
Tulay Adalı MLSP TC 1 of 8
First, a very brief history...
Started existence as the Technical Committee onNeural Networks for Signal Processing (NNSP) in 1990
First NNSP Workshop September 1991, in Princeton, NJ
First TC Chair, Fred B. H. Juang
Yearly workshops since 1991
Tulay Adalı MLSP TC 2 of 8
Yearly workshops since 1991
Workshops on Neural Network for Signal Processing
NNSP 1991, September 30–Oct. 2, Nassau Inn, Princeton, New Jersey, USA
NNSP 1992, August 31–September 2, Hotel Marienlyst, Helsingor, Denmark
NNSP 1993, September 6–9, Linthicum, Maryland, USA
NNSP 1994, September 6–8, Proto Hydra Resort Hotel, Ermioni, Greece
NNSP 1995, August 31–September 2, Royal Sonesta Hotel, Cambridge, Boston, USA
NNSP 1996, September 4–6, Keihanna, Seika, Kyoto, Japan
NNSP 1997, September 24–26, Amelia Island Plantation, Florida, USA
NNSP 1998, August 31–September 2, Newton Institute, Cambridge, England
NNSP 1999, August 23–25, Madison, Wisconsin, USA
NNSP 2000, December 11–13, Sydney, Australia
NNSP 2001, September 10–12, Falmouth, USA
NNSP 2002, September 4–6, 2002, Martigny, Valais, Switzerland
NNSP 2003, September 17–19, 2003, Toulouse, France
Workshops on Machine Learning for Signal Processing
MLSP 2004, September 29–October 1, 2004, Sao Luis, Brazil
MLSP 2005, September 28–30 2005, Mystic, USA
MLSP 2006, September 6–8, 2006, Maynooth, Ireland
MLSP 2007, August 27–29, 2007, Thessaloniki, Greece
MLSP 2008, October 16–19, 2008, Cancun, Mexico
MLSP 2009, September 2–4, 2009, Grenoble, France
MLSP 2010, August 29–September 1, 2010, Kittila, Finland
MLSP 2011, September 18–21, 2011, Beijing, China
Tulay Adalı MLSP TC 3 of 8
From NNSP to MLSP
“Neural Network for Signal Processing” was deemed to betoo narrow a scope by many
Working with the IEEE SPS President at the time, Fred Mintzer,the TC approved the name: Machine Learning for Signal Processingwhich became the TC’s new name after approval by the BoG
Submissions to ICASSP in MLSPand the MLSP Workshop since 2002
Tulay Adalı MLSP TC 4 of 8
MLSP: What is the scope?
The bridge between machine learning and signal processing
Learning is the key aspect
Signal processing defines the main applications of interestand the constraints
Attractive solutions for traditional signal processing applicationssuch as pattern recognition, speech, audio, and video processing
Primary candidates for emerging applications such as BCI,multimodal data fusion and processing, behavior and emotionrecognition, and learning in environments such as social networks
Tulay Adalı MLSP TC 5 of 8
Current EDICS
Applications of machine learning
Bayesian learning; Bayesian signal processing
Cognitive information processing
Graphical and kernel methods
Independent component analysis
Information-theoretic learning
Learning theory and algorithms
Neural network learning
Pattern recognition and classification
Bounds on performance
Sequential learning; sequential decision methods
Source separation
Tulay Adalı MLSP TC 6 of 8
Areas of activity, emerging trends
Methods
Sparsity-aware learningLearning in kernel spacesSemi-supervised learningDistributed learningSubspace and manifold learningSemi-blind data analysis, learning
Besides learning, integration of approaches has been a keyemphasis, making MLSP a natural home for
brain-computer interfacebehavior and emotion recognitionmultimodal data fusion and processingmultiple/joint data analysislearning in environments such as social networks
Tulay Adalı MLSP TC 7 of 8
Cognitive information processing represents a majorparadigm shift in learning
A dynamic system is called cognitive if it exhibitsall four cognitive properties:
Perception-action cycle, which produces information gainabout the environment, obtained from one cycle to the next
Memory, which predicts the consequences of actionon/in the environment
Attention, which is responsible for the allocation ofavailable resources
Finally, intelligence provides the basis for decision-makingwhereby intelligence choices are made in the face ofenvironmental uncertainties
Tulay Adalı MLSP TC 8 of 8
Trends in Machine Learning for Signal Processing
Kostas DiamantarasTEI of Thessaloniki GreeceTEI of Thessaloniki, Greece
Data collected from paper titles appearing in ICASSPs (MLSP track) and MLSP workshops between 2007 2011(MLSP track) and MLSP workshops between 2007‐2011
LEARNING CLASSIFICATION, RECOGNITION CLUSTERING
23.2
19.8
12.6 12.6
18.617.9
13.2 13.4
9.3 9.2
7.3
2.1 2.31.1
3.8
2007 2008 2009 2010 2011
Diamantaras, ICASSP 2011, Prague, Czechia3/6/2011 2
Data collected from paper titles appearing in ICASSPs (MLSP track) and MLSP workshops between 2007 2011(MLSP track) and MLSP workshops between 2007‐2011
NEURAL, ANN, RBF, MLP, … BRAIN COMPUTER INTERFACES
6.1
4.6
3 3
2.9
1.7
2.32.1
1.7
3.3
1.1 1.2
2007 2008 2009 2010 2011
Diamantaras, ICASSP 2011, Prague, Czechia3/6/2011 3
Data collected from paper titles appearing in ICASSPs (MLSP track) and MLSP workshops between 2007 2011(MLSP track) and MLSP workshops between 2007‐2011
BLIND, ICA, etc... SPARSE REPRESNTATION
15.6
18.4
15.915.0
15.6
10.4
8.5
3 64.0
4.65.5
3.6
2007 2008 2009 2010 2011
Diamantaras, ICASSP 2011, Prague, Czechia3/6/2011 4
Data collected from paper titles appearing in ICASSPs (MLSP track) and MLSP workshops between 2007 2011(MLSP track) and MLSP workshops between 2007‐2011
BAYESIAN SOCIAL
9.8
5 7 5 7
6.6
5.7
2.9
5.7
2.4
0.0 0.0 0.0 0.0
2007 2008 2009 2010 2011
Diamantaras, ICASSP 2011, Prague, Czechia3/6/2011 5
Data collected from paper titles appearing in ICASSPs (MLSP track) and MLSP workshops between 2007 2011(MLSP track) and MLSP workshops between 2007‐2011
MUSIC BIO, GENE EXPRESSION, ETC...
2.9 2.92.9
MUSIC BIO, GENE EXPRESSION, ETC...
2.3
1 7
1.2
1.7
1.1
0.50.6
0.0
2007 2008 2009 2010 2011
Diamantaras, ICASSP 2011, Prague, Czechia3/6/2011 6
Data collected from paper titles appearing in ICASSPs (MLSP track) and MLSP workshops between 2007 2011(MLSP track) and MLSP workshops between 2007‐2011
SVM, KERNEL METHODS
16.415.6
17.1
SVM, KERNEL METHODS
13.2
6.3
2007 2008 2009 2010 2011
Diamantaras, ICASSP 2011, Prague, Czechia3/6/2011 7
Data collected from paper titles appearing in ICASSPs (MLSP track) and MLSP workshops between 2007 2011(MLSP track) and MLSP workshops between 2007‐2011
COMMUNICATIONS, SENSORS, NETWORKS, ETC...
5.5
COMMUNICATIONS, SENSORS, NETWORKS, ETC...
MEDICAL, EEG, ECG, MRI, ETC...
3.53 3
3.7
5.0
4.0
5.2
2.1
3.3
2.4
0.6
2007 2008 2009 2010 2011
Diamantaras, ICASSP 2011, Prague, Czechia3/6/2011 8
Data collected from paper titles appearing in ICASSPs (MLSP track) and MLSP workshops between 2007 2011(MLSP track) and MLSP workshops between 2007‐2011
AUDIO, ACOUSTICSAUDIO, ACOUSTICS
SPEECH
MULTIMEDIA, IMAGING, ETC...
10.711.5
9.3
5.8 5.7
7.36.4
7.5
5.5
3 74.0
6.1
2.9
1.6
3.5 3.7
2007 2008 2009 2010 2011
Diamantaras, ICASSP 2011, Prague, Czechia3/6/2011 9
Data collected from paper titles appearing in ICASSPs (MLSP track) and MLSP workshops between 2007 2011(MLSP track) and MLSP workshops between 2007‐2011
COGNITIVE
2.3
COGNITIVE
0.6 0.5
0.0
0.5
0.0
2007 2008 2009 2010 2011
Diamantaras, ICASSP 2011, Prague, Czechia3/6/2011 10
Cognitive information processingCognitive information processing- an emerging trend for MLSP
Jan LarsenCognitive Systems SectionDept. of Informatics and Mathematical ModellingTechnical University of [email protected], www.imm.dtu.dk/[email protected], www.imm.dtu.dk/ jl
Why is it important? VISION
What should we do? MISSIONWhat should we do? MISSION
03/06/2011Jan Larsen2 DTU Informatics, Technical University of Denmark
The legacy
Allan Touring
Theory of
Norbert Wiener
CyberneticsTheory of computing, 1940’es
y
1948
processing adaption under-standing cognition
03/06/2011Jan Larsen3 DTU Informatics, Technical University of Denmark
VisionVisionCognition refers to the representations and processes involved in Cognition refers to the representations and processes involved in thinking and decision making. Cognitive information processing integrate information processing in brains and computers for collaborative problem solving in open-ended environmentsp g p
The vision is to design and implement profound g p pcognitive information processing systems for augmented human cognition in real-life environmentsenvironments
Disentanglement of confusing, ambiguous, conflicting, and vast
03/06/2011Jan Larsen4 DTU Informatics, Technical University of Denmark
g g, g , g,amounts of multimodal, multi-level data and information
Visnevski / Castillo-Effen tiered approachHow much is needed to qualify thesystem as being cognitive?
A i d h f l hi hrobustness
A tiered approach: from low to high-level capabilitiesadaptivity
efficiencynatural interaction
Ref: N A Visnevski and M Castillo Effen: A UAS capability description framework: Reactive
emergent properties
03/06/2011Jan Larsen5 DTU Informatics, Technical University of Denmark
Ref: N.A. Visnevski and M. Castillo-Effen: A UAS capability description framework: Reactive, adaptive, and cognitive capabilities in robotics, 2009 IEEE Aerospace Conference, pp. 1-7, 2009.
It takes cross-disciplinary effort to create a cognitive systemcognitive system
INFOEngineering and natural
sciences
BIOCOG CIP
BIONeuro and
life sciences
Cognitivepsychology,
social sciencies, linguisticslinguistics
03/06/2011Jan Larsen6 DTU Informatics, Technical University of Denmark
Ref: EC Cognitive System Unit http://cordis.europa.eu/ist/cognition/index.html
Revitalizing old visions through cognitive i f i i b f information processing systems by means of enabling technologies
Computationdistributed (grid cloud)
Connectivityinternet communication
Pervasive sensing and datadistributed (grid,cloud)
and ubiquitous computing
internet, communication technologies and social
networksdigital, accessible
information on all levels
New theories of the human brain
Neuroinformatics brain-
New business modelsFree tools paid by
advertisement, 99+1 Neuroinformatics, brain-computer interfaces,
mind reading
principle: 99% free, 1% buys, the revolution in
digital economy
03/06/2011Jan Larsen7 DTU Informatics, Technical University of Denmark
The unreasonable effectiveness of data
• E. Wigner 1960: The unreasonable efffectiveness of mathematics in the natural sciences.
• Simple linear classifiers based on many features from n-gram representations performs better than elaborate modelsrepresentations performs better than elaborate models.
• Unsupervised learning on unlabeled data which are abundant• The power of linking many different sources• Semantic interpretationSemantic interpretation
– The same meaning can be expressed in many ways – and the same expression can convey many different meanings
– Shared cognitive and cultural contexts helps the disambiguation of meaningmeaning
– Ontologies: a social construction among people with a common shared motive
– Classical handcrafted ontology building is infeasible – crowd computing / crowdsourcing are possible
Ref: A. Halevy, P. Norvig, F. Pereira: The unreasonbale effectiveness of data,
03/06/2011Jan Larsen8 DTU Informatics, Technical University of Denmark
y, g, ,
IEEE Intelligent Systems, March/April, pp. 8-12, 2009.
MissionMissionA cognitive information processing system should optimize itself according to:
The statistical model of the domain, the psycho-physical model of the users, the social context, and the computational resources in time and spacethe computational resources in time and space
03/06/2011Jan Larsen9 DTU Informatics, Technical University of Denmark
The cognitive information processing system and its world
R l/ i t l i t
Common sense knowledge Human
user HumanReal/virtual environment
Domian
user
HumanHuman
user
Multi
knowledge Humanuser
ACSMulti-modal
sensors ACS ACS
actionsACS
03/06/2011Jan Larsen10 DTU Informatics, Technical University of Denmark
Information processing and computing
Dynamical multi-level integration and learning ofDynamical, multi-level, integration and learning of– heterogeneous,– multi-modal,
multi representation (structured/unstructured)– multi-representation (structured/unstructured),– multi-quality (resolution, noise, validity)– data, information and human interaction streams
with the purpose of with the purpose of • achieving relevant specific goals for a set of users, • and ability to evaluate achievement of goals
iusing• new frameworks and architectures and• computation (platforms, technology, swarm intelligence,
id/ l d ti d ti )
03/06/2011Jan Larsen11 DTU Informatics, Technical University of Denmark
grid/cloud computing, crowd computing)
Examples of state of the art along diverse Examples of state of the art along diverse dimensions
• Cognitive radio networks• Cognitive radio networks• Cognitive radar• Cognitive components
(C iti ) i t k• (Cognitive) sensing networks• (Cognitive) social network models• (Cognitive) information retrieval and content management
engines
03/06/2011Jan Larsen12 DTU Informatics, Technical University of Denmark
What could the MLSP community contribute
• Bayesian learning as the fundamental learning and • Bayesian learning as the fundamental learning and information fusion principle
• Nonparametric Bayes• Signal representation and features• Signal representation and features• Sparse models for high-dimensinal data• Dedicated, efficient, robust on-line algorithms for large
l d tscale data• Engineering and demonstration of cognitive information
processing platforms
03/06/2011Jan Larsen13 DTU Informatics, Technical University of Denmark
We can only see a yshort distance
ahead, but we can h h i see that there is
much to be donemuch to be done
03/06/2011Jan Larsen14 DTU Informatics, Technical University of Denmark
Alan Turing, 1950