We have successfully replaced thousands of complicated deep net time series based anomaly detectors at a FANG with statistical (nonparametric, semiparametric) process control ones.
They use 3 to 4 orders lower number of trained parameters and have just enough complexity that a team of 3 or four can handle several thousands of such streams.
The amount of baby sitting that deep net models needed was astronomical, debugging and understanding what has happened quite opaque.
For small teams, with limited resources I would still heavily recommend stats based models for time series anomaly detection.
May not be your best career move right now for political reasons. Those making massive bets do not like to confront that some of their bets might not have been well placed. They may try to make it difficult for contrary evidence to become too visible.
Rendered at 11:47:29 GMT+0000 (Coordinated Universal Time) with Vercel.
We have successfully replaced thousands of complicated deep net time series based anomaly detectors at a FANG with statistical (nonparametric, semiparametric) process control ones.
They use 3 to 4 orders lower number of trained parameters and have just enough complexity that a team of 3 or four can handle several thousands of such streams.
The amount of baby sitting that deep net models needed was astronomical, debugging and understanding what has happened quite opaque.
For small teams, with limited resources I would still heavily recommend stats based models for time series anomaly detection.
May not be your best career move right now for political reasons. Those making massive bets do not like to confront that some of their bets might not have been well placed. They may try to make it difficult for contrary evidence to become too visible.