上图是前馈网络基本结构,分别介绍几种前馈网络:
M-P model,single-layer perceptron network,multi-layer perceptron network, BP(back propagation)
(一)M-P model
Proposed by McCulloch & Pitts
Consists of fixed Structure & fixed Link-Weights
Restrain Convex-Touch Weight ⇒ output is 0
Exciting Convex-Touch Weight ⇒ output is 1 when
the accumulated value ≥θ
Advantages:
Simplest Feed-forward NN
Can implement some Logical Operational Func
Disadvantages:
I/O & weight must be discrete values: 0 or 1
Link-Weight can not be adjusted dynamically
(二)single-layer perceptron network
supervised method可调权重
这里y=f(s)也可以是其他函数,根据具体情况,因为这里是用来简单分类,所以sign函数就够了。
然后计算权重,使用delta study principle。
分类结果输出。
缺点:Only linear classifiable patterns can be classified by Single-Layer Perceptron
linear non-classifiable patterns cannot be classified by Single-Layer Perceptron