Abstract

Anomaly Detection with Variational Autoencoders

At this session you will be training a variational autoencoder to detect anomalies within data. Variational autoencoders are rooted in Bayesian inference and can outperform traditional techniques. There are use cases for anomaly detection in almost every industry including: cyber security, finance, healthcare, retail, telecom, ....


Speaker: Adam Henryk Grzywaczweski
Type:

DLI Instructor-Led Training 

Date: Thursday - October 18, 2018 01:00 PM - 03:00 PM
Location: Halls C4&C5
Topic: Intelligent Video Analytics

 

At this session you will be training a variational autoencoder to detect anomalies within data. Variational autoencoders are rooted in Bayesian inference and can outperform traditional techniques. There are use cases for anomaly detection in almost every industry including: cyber security, finance, healthcare, retail, telecom, and video surveillance. Anomalies are detected by assuming that the datapoints are coming from a distribution that we can model effectively with deep learning. A datapoint is considered an anomaly if the probability of it being generated from the model is below a threshold. Thresholds are set depending on the business/industry objective. During the training you will test how data processing, model architecture, and hyperparameter tuning affect the accuracy of the model. Prerequisites: Experience with CNNs