gordon pipa institute of cognitive science university of
TRANSCRIPT
Identification of complex biological Motion patterns
Gordon Pipa
Institute of Cognitive Science University of Osnabrück
With:
Matthias Staib (UOS) Frank Jäckel (UOS) Robert Haslinger (MIT)
The Brain: Large, self-organized but mostly random Network
Giese 2003
RF V1
RF V4
RF IT
Hagmann, et al. (2008), ’Mapping the structural core of human
cerebral cortex’ PLoS Biol 6(7): e159
• Complex network with a high degree of randomness, and that is changing at all times
• Intrinsic activity dominates complex neuronal dynamics (Stimulus = Perturbation)
• Haslinger, R., Pipa, G., Lima, B., Singer, W., Brown, E. N., & Neuenschwander, S. (2012). Context Matters:
The Illusive Simplicity of Macaque V1 Receptive Fields. PLoS ONE, 7(7), e39699.
Biological Motion with Reservoir Computing
excitatory
inhibitory
Output
SORN:
• 25 bipolar cells (center surround RF), project randomly on excitatory neurons
• Reservoir: 100 exc. + 20 inh.
• Plasticity for Reservoir learning
• RF covers ~6% of visual field
• A. Lazar, G. Pipa, and J. Triesch, Frontiers Computational Neuroscience 2009
• R. Haslinger, G. Pipa, B. Lima, W. Singer, E.N. Brown, S. Neuenschwander (2012). PLoS ONE, 7(7), e39699
happy,
sad,
angry
Classification
Reservoir Computing
Incoming drive Threshold Input
task: Mood
Bipolar Cells
With center
surround
1
1 ( )N
i ij j i
j
h t f W t x t T t I t
• A. Lazar*, G. Pipa*, and J. Triesch (*authors contributed equally), Neural Networks, 20(3):312--322, 2007
• A. Lazar, G. Pipa, and J. Triesch, Frontiers Computational Neuroscience 2009
RF covers 6%
Network structure
Network activity
Plasticity
time
un
it
Information processing
Two types of neuronal plasticity:
• Spike timing dependent plasticity (STDP) for network structure formation
• Intrinsic plasticity for homeostasis of average activity of neurons
• Unsupervised learning in the recurrent network
Self-organization of the network and network activity
• V. Gomez, A. Kaltenbrunner, V. Lopez, H. Kappen, ’Self-organization using synaptic plasticity’, NIPS 2008
• Savin C, Joshi P, Triesch J (2010) Independent Component Analysis in Spiking Neurons. PLoS Comput Biol 6(4)
• A. Lazar*, G. Pipa*, and J. Triesch (*authors contributed equally), Fading memory and time series prediction in
recurrent networks with different forms of plasticity, Neural Networks, 20(3):312--322, 2007
• Binarized STDP:
• Intrinsic plasticity:
• Synaptic scaling:
SORN – 3 plasticity mechanisms
• A. Lazar*, G. Pipa*, and J. Triesch (*authors contributed equally), Neural Networks, 20(3):312--322, 2007
• A. Lazar, G. Pipa, and J. Triesch, Frontiers Computational Neuroscience 2009
bipolar cells
SORN:
• Bachelor Thesis in Cognitive Science, Osnabrück Germany by Matthias Staib (2013)
• A. Lazar, G. Pipa, and J. Triesch, Frontiers Computational Neuroscience 2009
Cycle:
16 Frames
PreTraining:
500x3 cycles
Training:
100x3 cycles
Testing:
100x3 cycles
from 35 subjects
Performance
• Similar performance for SORN and LSM with
shuffled weights and thresholds.
• RC properties, i.e. feature expansion and
memory are necessary
(Control Logistic classifier: ~55%)
• Why not just a random LSM ?
• What is the advantage of SORN that learns
temporal patterns based on plasticity ?
• A. Lazar*, G. Pipa*, and J. Triesch (*authors contributed equally), Neural Networks, 20(3):312--322, 2007
• A. Lazar, G. Pipa, and J. Triesch, Frontiers Computational Neuroscience 2009
Performance for occluder Task
• Initial fading memory the same for both the LSM and SORN
• After 15 frames memory maintains to be larger for SORN
• Distance between models (Hamming distance) larger for SORN
Generalisation for other viewing directions
Training orientations: 50°, 75°, 110°, 135°
Interpolation: Radom orientation in [55,° 70°] or [115° 130°]
Extrapolation: Radom orientation in [5,° 50°] or [140° 175°]
Conclusions
• Both the hierarchical and the RC based recognition of emotions work
well
• One to one comparison of performance is difficult since models are
rather different in structure
• Most of the activity even in early primary areas is not directly stimulus
related
• Classical approaches using feed forward structures explain this with
noise
• The RC approach provides an alternative that utilizes induced complex
temporal firing sequences to maintain memory, bridge lacking
stimulation, and optimize stimulus presentations.
• Processing a complex stimulus property such a mood can be
implemented with a single recuurent module, for example in V1.