Self-organizing map neural network

principle

• Clustering high dimensional visualization, data compression, feature extraction
• Self-organizing map neural network is essentially a two-layer neural network comprising an input layer and output layer (competitive layer) the number of output layer neurons is the number of clusters is usually
• A "competitive learning" training mode, each input node to find a match and most of it in the output layer, called active nodes. Decline followed by a stochastic gradient method to update the parameter of the active nodes, while activating adjacent node points are also appropriately updated according to their distance from the active nodes (update mode is "Mexican-hat"). This competition can be suppressed by a transverse connection between neurons (negative feedback path) to achieve.
• Output layer node topology is a (one-dimensional linear array, two-dimensional array, a three-dimensional grid array), and has the characteristics of order-preserving mapping
• learning process:
  • initialization
  • Competition: find active nodes
  • Collaboration: Critical Update node parameters
  • Adaptation: adjust related neuronal excitatory connection weights so that similar response winning neuron input model for subsequent use enhanced
  • Iteration: Back to competition

The difference between the average K

• The actual number of clusters does not need to determine the number of classes, because the clustering result may be less than the number of neurons
• K-means the most to find a similar class, the only class parameter update; self-organizing map neural network to update the neighboring nodes. So K-means susceptible to noise, while the latter accuracy may be lower
• Since visualization is better organization map neural network, with elegance topology diagram