Unsupervised Learning
Competitive Learning
Neurons compete against each other to be activated. Only a single output neuron active at a time (winner-takes-all neuron).
Self-Organizing Feature Map (SOFP)
Each sensory input is mapped into (associated with) a corresponding area of the cerebral cortex.
Single layer of computation neurons. Only the winner neuron gets activated. The winner neuron & its neighbors are allowed to learn, where the neighbors are neurons in close physical proximity to the winner.
- Connections
- Forward connection
- Lateral connection
- Produce excitatory or inhibitory effects, depending on the distance from the winning neuron
- Achieved by Mexican hat function
Mexican Hat Function
Algorithm
- Init weights
- Activation & finding winner
- $$minj |X - W_j| = min_j \sqrt{\sum{i=1}^n (xi - w{ij})^2}$$
- Learning by updating weights
- $$w'{ij} = w{ij} + \Delta w{ij} = w{ij} + \alpha (xi - w{ij}) \text{ if activated}$$
- Iteration
- Minimum-distance Euclidean distance satisfied