A classic kind of problem we addressed in practice: can we find signals of, or predictive signals of future occurrence? Powerful idea. How possible is this, very generally in streams of data? Connection to risk analysis?
Method Finds Hidden Warning Signals in Measurements Collected Over Time
MIT News Daniel Ackerman
A new technique developed by researchers at the Massachusetts Institute of Technology (MIT) and Spain's Rey Juan Carlos University uses a generative adversarial network (GAN) to identify anomalies in time series data. The TadGAN approach could be used to detect and respond to significant changes in a range of high-value systems. The researchers aimed to create a general framework for anomaly detection that could be applied across industries. TadGAN can distinguish between normal and anomalous data points by checking for discrepancies between the real-time series and a fake GAN-generated time series. Supplementing the GAN with an autoencoder algorithm helped the researchers strike a balance between being vigilant and raising too many false alarms. In anomaly detection tests on 11 datasets, TadGAN outperformed ARIMA, the traditional approach to time-series forecasting, for eight datasets. Said MIT's Sarah Alnegheimish, "We want to mitigate the stigma about artificial intelligence not being reproducible."
No comments:
Post a Comment