Any idea about sparse Matrix inputs? (for RNN, LSTM? )

So, imaging I’m building an RNN, where the input are discrete temporal events. The occurrences of those events are sparse, and the types of events are limited. (i.e, for a time scale of seconds, one only breaths every 20 seconds, and bats eye lids every 10 seconds, coughs at random intervals. )

I have tried to shove the data into an LSTM with very limited success.

Is there any way I can massage the data and or optimize the NN to make it work better?

Thanks before hand!

How much data do you have?

Have you tried condensing the input? Maybe you can squeeze the empty periods, like 60 seconds of nothing happening counts as 1.

Increasing the RNN sequence length is also worth a shot.

1 Like

Increased RNN sequence length, still not so good.
So, I’m parsing music(like) data, so turns out I just need to read up on magenta, a google deeplearning music package and how it works…