hu_ShiftInvarianceValidation_Biosignals2020

• Validation of the effect of maxblur pooling method for improving temporal shift-invariance of Convolutional Neural Network (CNN) in bio-signals classification tasks.
• About maxblur pooling, please refer to Making Convolutional Networks Shift-Invariant Again
• Paper
• Best Paper Award in Biosignals2020

Max Pooling	Maxblur Pooling

Task1: Validate on AF Detection from ECG

What is AF Detection

• We used MIT-BIH Atrial fibrillation (AF) Dataset
• The dataset contains 23 records of 10 hour ECG with heart beat annotation and AF annotation
• 100 RRI as a Segment
• Do 2-class classification Task (Normal / AF)

Results

Evaluate using Accuracy and Consistency

• Accuracy : classification accuracy on test data
• Consistency : how often the model predict the same label given 2 different shift to the same input

Data Augmentation	1-Layer	2-Layer	3-Layer
Yes
No

Evaluate using Accuracy and Robustness

• Accuracy : classification accuracy on test data
• Robustness : classification accuracy on crashed test data

Data Augmentation	1-Layer	2-Layer	3-Layer
Yes
No

Improvements of maxblur on non-augemented data in CNN with different number of pooling layers

filter size	baseline = max	baseline = avg
7