Time Organized Maps – Learning cortical topography from spatiotemporal stimuli “ Learning cortical topography from spatiotemporal stimuli ”, J. Wiemer, F. Spengler, F. Joublin, P. Stagge, S. Wacquant, Biological Cybernetics, 2000 “The Time-Organized Map Algorithm: Extending the Self-Organizing Map to Spatiotemporal Signals”, Jan C.Wiemer, Neural Computation, 2003 Presented by: Mojtaba Solgi TexPoint fonts used in EMF. Read the TexPoint manual before you delete this box.: AAAA A
Outline 1. The purpose and biological motivation 2. The Model: TOM Algorithm Wave propagation Learning 3. Experiments and Results Gaussian stimuli Generic artificial stimuli Semi-natural stimuli 4. Discussion 5. z
Neurobiological experiments, Spengler et al., 1996, 1999
Terminology Integration Fusion of different stimuli into one representation Segregation: Process of Increasing representational distance of different stimuli z
2D Network Architecture Activation positional shift
One-dimensional model
Wave propagation
Integration and Segregation
Algorithm 1. Compute neurons activations and the position of the top winner neuron 2. Compute the neural position of the propagated wave from the last time step activation
Algorithm – Cont. 3. Shift the position of the top winner neuron due to interaction with propagated wave
Algorithm – Cont. 4. Again shift the position of the winner neuron this time due to noise 5. Update the winner neurons weights SOM Hebbian
Experiments with Gaussian stimuli & 2D neural layer 1. Simulation of ‘ontogenesis’ (Development)
Experiments with Gaussian stimuli & 2D neural layer 2. Simulation of post-ontogenetic plasticity
One-dimensional model
Experiments with generic artificial stimuli & 1D neural layer The input
Experiments with semi-natural stimuli & 1D neural layer
Discussion Importance of temporal stimulus for development of cortical topography Continuous mapping of related stimuli Inter-Stimulus-Interval-Dependant representations Hardly scalable No recognition performance on real-world problems Tested only on artificial input
Summary Utilizing temporal information in developing cortical topography Wave-like spread of cortical activity Experiments and results show compatibility of the model with neurobiological observations Biologically inspired and plausible, but no engineering performance
Thank you! Any thoughts/question?