Authors:
Xianyin Hu
;
Shangyin Zou
;
Yuki Ban
and
Shin’ichi Warisawa
Affiliation:
Development of Human and Engineered Environmental Studies, The Graduate School of Frontier Sciences, The University of Tokyo, Kashiwa, Chiba, Japan
Keyword(s):
Temporal Shift-invariance, Maxblur-pooling, Neural Network, Bio-signal Processing, Atrial Fibrillation (AF) Detection, Emotion Recognition.
Abstract:
Modern neural networks are widely employed in bio-signal processing due to their effectiveness. However, recent research showed that neural networks for image recognition is not shift-invariant as it was assumed, while it is an important property in bio-signal processing. Fortunately, a simple methodology was proposed referred to as Maxblur-pooling to improve the shift-invariance of neural networks for image recognition. However, the corresponding issue in the domain of bio-signal processing remains untouched. To verify the shift-invariance of neural networks when applied to bio-signal processing, we performed two experiments across different tasks and types of bio-signals. One is Atrial Fibrillation (AF) detection from R-R interval and the other is emotion recognition from multi-channel EEG. We were able to show that the lack of shift-invariance also happens in temporal bio-signal classification. In the AF detection task, we succeed to validate the effectiveness of Maxblur-pooling,
which demonstrating improvements in both accuracy (2%-13%) and consistency (8%-15%) compared to the baseline. While for the emotion recognition task, we did not observe any improvements using Maxblur-pooling. Our research provided empirical knowledge for developing real-time diagnose systems that is stable to temporal shifts.
(More)