Engineered and Artificial Systems

1
Engineered and Artificial Systems
2
Engineered and Artificial Systems

• Methodological and applied research
• Modelling of learning and perception
• Bayesian Object Recognition (Jouko Lampinen)
• Computational Neuroscience via Autonomous
  Robotics (Harri Valpola)
• Engineered nanosystems
• Biosensing systems (Jukka Tulkki)

3
Engineered and Artificial Systems
4
Modelling of learning and perception

• Learning and perception is central issue in many
  research topics in LCE
• Cohesive research in human and machine perception

5
Modelling of learning and perception
Basic paradigm perception is active prediction
process
Prediction
Action
Sensory input
State space model
Novelty
Update

• Generative models / Bayesian inference
• Task oriented data driven process
• Attention vs. dual control (Optimal control
  balancing control
• errors
  and estimation errors)

6
Bayesian Object Recognition
Perception as Bayesian Inference perception
prior knowledge sensory input

• Object matching
• Sequential Monte Carlo
• Clutter, occlusions etc
• View-point invariance
• 3D models / learnt views
• SMC, PMC, MCMC
• Segmentation
• Data-driven MCMC
• Multiple texture classes

7
Bayesian Object Recognition
Matching a face with occlusion by Sequential MC
8
Bayesian Object Recognition
Modeling of Feature Variation due to 3D Rotations
Sequential MC matching of 3D shape models
9
Bayesian Modelling of Perception

• Research goals
• Efficient and scalable algorithms
• Expectation propagation
• Particle Filters / MCMC with data driven
  proposals
• Computational models of the biological
  perception
• Hierarchical Bayesian inference in the visual
  cortex (following Mumford, Friston, etc)
• Modelling adaptation of auditory system as dual
  control process
• Hierarchy of learning and perception
• Category learning - from features to object
  classes
• Perceptual grouping processes

10
Computational Neuroscience
11
Goals
12
Methods
13
Timeline

• Adaptive motor control
• Learn abstract features serving behavioral goals
• Combine attention and learning
• Reward-based learning of orienting behaviors
• Imagination
• Navigation
• Planning
• Episodic memory
• Communication and language

2006
2007
2008
2009
2010
2011
14
Collaboration inside LCE

• Complex networks and agent-based models
• Competitive processes in networks
  (non-equilibrium dynamics)
• Adaptation of network weights and topology
• Cognitive systems
• Experimental research of attention and
  perceptual learning

15
External collaboration
Attention modelling Decos group in Barcelona
Robotics and comp. neuroscience EU
projects RobotCUB, ICEA
Machine learning Info-lab, TKK
16
Adaptive control
17
Abstractions and meaning
18
Engineered nanosystems

• Research group of Engineered Nanosystems studies
  properties of materials on an atomistic scale.
  Using computational approach we can both create
  and analyse structures which are hard to approach
  using experimental techniques.

Professors Kimmo Kaski, Jukka Tulkki Adjunct
Professor Adrian P. Sutton (Imperial
Collage) Researchers Sebastian von Alftan,
Fredrik Boxberg, Teppo Häyrynen, Jani Oksanen,
Sebastian Köhler
19
Cooling of radiative excitons by THz-radiation
Fredrik Boxberg , Jukka Tulkki, Go Yusa,
Hiroyuki Sakaki Institute of Industrial
Science, The University of Tokyo, Japan

• The free electron laser (FEL) of UCSB was used to
  modulate intra-band dynamics of electrons and
  holes.
• Cooling results from resonance transfer of holes
  from dark states in the piezominima to
  deformation potential minimum, where they can
  recombine radiatively.

20
Monte-Carlo model of dynamics
The spin-dependent Monte-Carlo model accounts for
governing relaxations mechanisms and makes use
of experimental time constants. Since the
system is not in quasi equilibrium there is no
temperature however the average energy of
subsystems can still change in analogy to
cooling and heating.
21
Cooling and transient effects
The steady state THz-excitation shows enhanced
emissin from exciton ground state
The transient THz-pulse induces luminescence even
seconds after carrier generation has been
switched off
The resonance energy transfer concept can be
generalized!
22
Biosensing systems

• Optical systems
• Quantum dot fluorescent labelling
• Autofluorescence
• Chemi-/bioluminescence
• Fourier transform IR spectroscopy
• Holographic sensors
• Surface plasmon resonance
• Surface-enhanced Raman spectroscopy

Sebastian Köhler and Jukka Tulkki, a new
project started 2006
23
Coherent optical flip-flop (COFF)

• A bistable system of phase locked lasers
• Nonlinear feedback obtained through interference
  of coherent signals
• Meets most of the critical requirements of an
  integratable flip-flop
• Logic gates with small modifications

Jani Oksanen and Jukka Tulkki, Apl. Phys. Lett.
(2006)
24
Operation principle (1/3)

• A phase locked laser amplifier

25
Operation principle (2/3)

• Adding a coherent bias signal makes the output
  nonlinear Pout(P1½ - Pbias½)2

26
Operation principle (3/3)

• Combining two laser amplifiers makes a bistable
  system with two stable states

27
A schematic of the COFF
28
Biomorphic networks
T. Häyrynen and J.Tulkki, collaboration with M.
Weckströn Univ. of Oulu and J. ahopelto and M.
Åberg VTT Microelectronics Centre
Research of SET-based neural circuits with
high parallelism and low dissipation
S
G1
G3
G2
G4
D
Multigate SET schematically