So my idea is to train an LSTM - autoencoder for anomaly detection by continual learning, i.e., I want to update the model after each 10 time steps. Firstly I will train it on source data, then re-train on first batch of new data and then re-train on second batch of new data and so on. My question is: how to preprocess these data? Should the metrics computed on source data be used for each new batch or should these metrics be updated somehow?
Asked
Active
Viewed 21 times
1
brazofuerte
- 737
- 4
- 19
pikachu
- 731
- 2
- 10
-
There is no fixed answer to this as it depends on your assumptions/knowledge about your incoming datastream (i.e. how it may differ in distribution from the source data) and also the CL architecture itself. – brazofuerte Feb 24 '21 at 11:21
-
@brazofuerte and is there any guideline how to deal with different scenarios? Maybe some hints from which one can bounce off. – pikachu Feb 24 '21 at 11:58
-
Mileage is going to worse than just a very. There are plenty of cases where the settings that make you successful in one area actually crater you in another area. You need to start with the business case, for your particular business. Then you can get down to drilling into the details and targeting it more appropriately. Another problem here is the “when your hammer every problem looks like a nail”. There are plenty of cases where LSTM is the wrong solution, and much simpler creatures, such as a mean, outperform it. – EngrStudent Feb 24 '21 at 12:07