Encoding features robust to unseen modes of variation with attentive long short-term memory
作者：
Highlights：
• We devise a novel LSTM adaptation called the attentive mode variational LSTM.
• Attention mechanism is proposed to separate the input signal into task-relevant, and task-irrelevant feature sequences.
• The proposed method encodes the mode of variation in the input sequence at test time using the task-irrelevant feature sequence.
• The proposed method encodes features robust to variation unseen during the training.
• The proposed attentive mode variational LSTM can be generalized to different applications.
摘要
•We devise a novel LSTM adaptation called the attentive mode variational LSTM.•Attention mechanism is proposed to separate the input signal into task-relevant, and task-irrelevant feature sequences.•The proposed method encodes the mode of variation in the input sequence at test time using the task-irrelevant feature sequence.•The proposed method encodes features robust to variation unseen during the training.•The proposed attentive mode variational LSTM can be generalized to different applications.
论文关键词：Long short-term memory,Recurrent neural networks,Attention,Robust features,Modes of variation,Facial expression recognition,Human action recognition
论文评审过程：Received 25 January 2019, Revised 10 December 2019, Accepted 11 December 2019, Available online 18 December 2019, Version of Record 2 January 2020.
论文官网地址：https://doi.org/10.1016/j.patcog.2019.107159
