Kohonen Self-organizing Feature Maps Order out of randomness
Copyright Gene A. Taglairini, PhD Kohonen Network Model Neurons Inputs 11/26/2018 Copyright Gene A. Taglairini, PhD
Copyright Gene A. Taglairini, PhD Network Features Input nodes are connected to every neuron The “winner” neuron is the one whose weights are most “similar” to the input Neurons participate in a “winner-take-all” behavior The winner output is set to 1 and all others to 0 Only weights to the winner and its neighbors are adapted 11/26/2018 Copyright Gene A. Taglairini, PhD
Copyright Gene A. Taglairini, PhD Neighborhoods Symmetric with maximum value weight changes applied near the center Excitatory near the center, surrounded by an inhibitory band, and may be enclosed by an excitatory influence rapidly decreasing to zero 11/26/2018 Copyright Gene A. Taglairini, PhD
Network Equations: Similarity Similarity is measured using the Euclidean distance from an input pattern vector inputp The vector of weights w[i] represents all the weights to neuron i; hence, w[i][j] is the weight that joins input j to neuron i The winner has the smallest value of s 11/26/2018 Copyright Gene A. Taglairini, PhD
Network Equations: Weight Adaptation Suppose the winning neuron has index a For an input j and neuron i, the weight change Dw[i][j] and the new value of the weight w[i][j] are: Dw[i][j] = h (input[j] – w[i][j] ) NbdWt(a, i) w[i][j] = w[i][j] + Dw[i][j] Where h is the learning rate and NbdWt(a, i) is the neighborhood weighting function 11/26/2018 Copyright Gene A. Taglairini, PhD
Network Equations: Neighborhood Weighting Function The neighborhood weighting function may take forms such as: Where s is a scalar that sets the dilation of the weighting function 11/26/2018 Copyright Gene A. Taglairini, PhD
On-center Off-surround Neighborhood A field that: Reinforces stimuli “near” the center Attenuates effects in a region about the center Abates rapidly outside the attenuating region Contributes to noise control and localizes representation Complementary to off-center on-surround 11/26/2018 Copyright Gene A. Taglairini, PhD
On-center Off-Surround Neighborhood 11/26/2018 Copyright Gene A. Taglairini, PhD
Mexican Hat Neighborhood 11/26/2018 Copyright Gene A. Taglairini, PhD
Basic Training Algorithm Initialize weights from input nodes to neurons Choose a neighborhood function While (input patterns mismatch weights) Find a winner Adapt weights in the vicinity of the winner Develop an interpretation for the encoding—identify which neurons encode what patterns 11/26/2018 Copyright Gene A. Taglairini, PhD
Copyright Gene A. Taglairini, PhD Applications Speech encoding for a phonetic typewriter Geometric pattern coding 11/26/2018 Copyright Gene A. Taglairini, PhD