-1

If I look at my data with different look back steps, would the result be the same/similar if I use more LSTM layers? Or differently asked, would I also gain different time-resolutions or do more layers not mean different time resolutions but only complexity and nonlinearity of the patterns??

Is the LSTM fixed on the timestep for identifying patterns or does LSTM also find inbetween/shorter patterns?

The core idea is to feed the LSTM the longest possible timestep to find long, mid, and short term patterns? Is that a valid approach or should I feed a different timestep for each range?

Thanks for your help

Florida Man
  • 2,021
  • 3
  • 25
  • 43

1 Answers1

0

Generally, LSTMs can capture long- and short-term dependencies in a sequence. But Pengfei Liu et al. raised the concern that for long sequences the important information from the beginning of the sequence has to be dragged through the whole sequence.

They used groups of LSTMs that are activated at different times. Thereby, acting as short and long-term memory.

Any thoughts on that?

Florida Man
  • 2,021
  • 3
  • 25
  • 43