Skip to content

Latest commit

 

History

History

dae

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
Dockerfile
Dockerfile
 
 
LICENSE
LICENSE
 
 
README.md
README.md
 
 
algorithm.py
algorithm.py
 
 
manifest.json
manifest.json
 
 
model.py
model.py
 
 
requirements.txt
requirements.txt
 
 

README.md

De-Noising-AutoEncoder (AE)

citekey SakuradaYairi2014Anomaly
source code own
Learning type semi-supervised
input multivariate

Notes

  • DeNoising AutoEncoder is trained on all noisy non-anomaly data. Whenever it encounters an anomaly value, the reproduction error is quite higher than the error with non-anomaly instances.
  • Noise is inserted in randomly selected inputs and turning them to a value of zero. (salt and pepper noise). The De-Noising-AE learns to reproduce the input with noise. The reproduction error is again used to classify between anomalous and non-anomalous data.
  • An assumption is made that all errors are normally distributed with some mean and std. Any error value that follows mean + kstd > threshold or mean - kstd < thereshold is considered as an anomaly. The type of noise added is salt and pepper which usually refer to setting some proportion of inputs to zero.