Objective performance evaluation

For numerical prediction tasks, the methods used for evaluating linear regression may also be used for evaluating neural nets. 数值预测的神经网络可以使用线性回归的评估方法。
For classification tasks, any of the methods used for evaluating classification tasks may also be used for neural nets. 分类预测的神经网络可以使用任何其他分类器的评估方法。
In addition, as a neural net can be used as a probablistic classifier, methods for probablistic classifiers, such as ROC charts may be used. 此外，由于神经网络可以用作概率分类器，概率分类器的方法，如ROC图可以被使用。

Interpretability of neural network 神经网络的可解释性

“Neural networks are like a black box. How can I ‘understand’ what the backpropagation network has learned?”

  • A major disadvantage of neural networks lies in their knowledge representation.
  • Acquired knowledge in the form of a network of units connected by weighted links is difficult for humans to interpret.
  • Easier-to-interpret rules may be extracted from the network by pruning or by sensitivity analysis.

Rule extraction by network pruning for interpretability

  • Simplify the network structure by removing weighted links that have the least effect on the trained network 通过去除对训练网络影响最小的加权链接，简化网络结构
  • The set of input and activation values are studied to derive rules describing the relationship between the input and hidden unit layers 研究输入和激活值的集合，以导出描述输入和隐藏单元层之间关系的规则

**

Sensitivity analysis for interpretability

  • Assess the impact that a given input variable has on a network output. 评估给定的输入变量对网络输出的影响。
  • The input to the variable is varied while the remaining input variables are fixed at some value. 变量的输入是变化的，而其余的输入变量是固定在某个值。
  • The knowledge gained from this analysis can be represented in rules 从分析中获得的知识可以用规则表示
    • Such as “IF X decreases 5% THEN Y increases 8%.” 如“如果X减少5%，那么Y增加8%。”