Neural Network Learning

  • Neural networks were first investigated by psychologists and neurobiologists to develop and test computational analogues of neurons 神经网络最初是由心理学家和神经生物学家研究的，目的是开发和测试神经元的计算模拟物
  • A neural network: a set of connected nodes where each node is an input (corresponding to an independent variable), output (corresponding to a dependent variable for numeric prediction or a class for classification; there may be several output nodes) or hidden and where each connection has a weight associated with it. 神经网络有输入、输出、和每个节点都有权重联系的隐藏节点。
  • During the learning phase, the network learns by adjusting the weights so as to be able to output the correct class label or predicted value for the input tuples. 在学习阶段，网络通过调整权值进行学习，以对应于输入，输出正确分类或预测值。
  • Also referred to as connectionist learning due to the connection between units
  • The classic training algorithm is called backpropagation. 经典的训练算法称为反向传播。
  • Convolutional Neural Networks (CNNs) are neural networks where there are specific roles and connection architectures for hidden nodes and these have been applied very sucessfuly to image recognition and other problems.
  • The term deep learning often refers to training CNNs. They are deep because there are many layers of hidden nodes.

**

Strength/Weakness of neural network as a classifier

Strength

  • High tolerance to noisy data 对噪点数据容易性高
  • Ability to classify untrained patterns
  • Well-suited for continuous-valued inputs and outputs 非常适合连续值输入和输出
  • Successful on an array of real-world data, e.g., hand-written letters, images, voice
  • Algorithms are inherently parallel
  • Techniques have recently been developed for the extraction of rules from trained neural networks最近已经发展出从训练过的神经网络中提取规则的技术

Weakness

  • _Long _training time
  • Require a number of parameters typically best determined empirically, e.g., network topology or structure 需要许多参数，通常最好的确定经验，例如，网络拓扑或结构
  • Poor interpretability: difficult to interpret the symbolic meaning behind the learned weights and of hidden units in network, that is, behaves as a “black box” model. 可解释性差