Autoencoder node saliency: Selecting relevant latent representations

作者：

Highlights：

• Novel node saliency methods are proposed for rankings autoencoder hidden nodes based on their capability of performing a learning task.

• The methods interpret what an autoencoder has learned during model training by identifying the specialty nodes that reveal explanatory input features.

• The methods suggest possible nodes that can be trimmed down for a more concise network structure.

摘要

•Novel node saliency methods are proposed for rankings autoencoder hidden nodes based on their capability of performing a learning task.•The methods interpret what an autoencoder has learned during model training by identifying the specialty nodes that reveal explanatory input features.•The methods suggest possible nodes that can be trimmed down for a more concise network structure.

论文关键词：Autoencoder,Latent representations,Unsupervised learning,Neural networks,Node selection,Model interpretation

论文评审过程：Received 8 March 2018, Revised 20 November 2018, Accepted 15 December 2018, Available online 17 December 2018, Version of Record 21 December 2018.

论文官网地址：https://doi.org/10.1016/j.patcog.2018.12.015