I have a question about RNN's based on LSTM cells. Currently i'm trying to predict anomalies in time series data, based on the prediction error. Is it reasonable to run RNN's distributed in production? For example to scale it with Google Cloud ML Engine. I want to be able to scale the model in case it has to compute to many requests during inference.

But when I distribute the model what will happen to the memory cell? The data is split and distributed over multiple nodes, will it still recognize the pattern of the time series data?


  • $\begingroup$ I think it really depends on how the data is gathered. Because you're looking at time series data, I'm not sure what a distributed model would look like. Say you had sensor data from a hundred machines and you were trying to predict anomalies, then you'd need one instance of the model for each machine so that it would understand the individual stream of data points. If it were one model predicting on multiple data streams, the memory cells may often not be relevant to the sequence you're currently predicting on. $\endgroup$ – Jay Speidell Jul 28 '18 at 21:30
  • $\begingroup$ Okay very clear to me. Exactly what I was thinking. $\endgroup$ – Mike Evers Aug 24 '18 at 15:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.