Time Series with convolution layers?

Course Q&A Machine Learning Specialization Advanced Learning Algorithms
1

Hi all,

I have dozens of process variable time series, and for one set of those process variable time series I have only one result, one number. Unfortunately, I don’t have a lot of examples to train my model. So lots of input data, and not a lot of output data. I am wondering what is the best approach to train a model in such cases? I the second week of the machine learning specialization I learned about convolution layers, and there I thought to myself, perhaps I could tackle such cases with convolutional layers, because it reduces dimensionality. I am looking just for some general answers, or way pointers. I appreciate your response.

Best Regards,
Rok Bohinc

2

Can you create some data that follow your certain pattern or maybe can you label the data you have so you can have inputs and outputs for them.

3

So I have about i=1-20 variables (floating points), and for each variable I have a time series with roughly j=1-500 time points. The variables are measured simultaneously, so the first time point j=1 is (Date Time) the same for all the variables the same, and they are of course correlated. So I have roughly 20*500=10000 data points per example. I have k=1-300 example points. The inputs are then x_ij^k. Per example, I have a floating point result ranging from 0 - 100, y^k. Does this help?

4

To be honest with you, you lost me halfway on it, sorry…if you dont have labels you have to come up with some labels I dont know what much other to tell you.