Chap5

Chap5 Neuron models¶

Mcculloch-Pitts model¶

MCP network is a directed graph that contains a set of neurons.

Perceptron¶

XOR and identity can not be computed by a singel perceptron.

A perceptron can only compute linearly separable functions of n binary arguments.

Supervised learning¶

Unsupervised learning¶

strategy: winner-take-all only one of the neurons is selected for a weight update

Convergence to a good solution can be accelerated by distributing the initial weight vectors according to an adequate heuristic.

alternative: "learning with conscience" algorithm

The attraction from each cluster can be modeled as so-called energy function.

geometric interpretation: The dynamics

Competitive learning¶

评论区