grouding基础:概念应该能直接被人类理解。 为了实现以上内容,SENN学习概念的过程满足: h相当于自编码器,把原始输入映射到可解释的concepts空间; 通过稀疏性来满足diversity,使得一个输入尽可能用独立的concept来表示,而不是所有输入都与所有concepts有关; 通过prototyping来给概念提供可解释性,即提供一个最小训练集来最...
Apart from this framework, Self-Explaining Neural Network (SENN) has been proposed, which is a neural network version of the linear model aiming to achieve both explainability and model complexity. However, this model does not consider the relationship between the concepts inherent in the data and...
2.2.3 Self-explaining Neural Networks Alvarez-Melis & Jaakkola (2018) introduce self-explaining neural networks (SENN) – a class of intrinsically interpretable models motivated by explicitness, faithfulness, and stability properties. A SENN with a link function g(⋅) and interpretable basis con...
Apart from this framework, Self-Explaining Neural Network (SENN) has been proposed, which is a neural network version of the linear model aiming to achieve both explainability and model complexity. However, this model does not consider the relationship between the concepts inherent in the data and...