Jul 20, 2017 · 1 min read
Thank you for your comment. An embedding layer is used at the start of the network to transform your inputs into vectors. So, the first step is to question whether it makes sense for you data to be replaced by vectors. If you have categorical data or data that only consists of integers (with a fixed amount of values like a word dictionary) you can give it a try. However, if you have a time series of values like 18,388 I would not try embeddings.
If your data is multi-dimensional you can experiment with a shared embedding layer or a separate one for each input.
