Hello everyone, Is there any approach to generate a fixed no. of feature vectors from the sequential data(similar to time series data...) and also express the same sequential data in terms of the generated feature vectors(each sequence as multinomial distribution of the feature vectors). Generating feature vectors is easy with neural networks(LsTMs, RNN...), but I have no idea about the decoding part. If anyone has worked on this or has an idea to approach, please answer this question. Thank you.
Considering the dimensions of a feature vector obtained by a CNN. It will be really complex because it has so many dimensions. However, using some kind of technique eg. PCA and reducing the dimensions can be away, but it depends upon the amount of information your need for further processing.
Khurram Hameed , I'm concerned more about representing the input data in terms of generated feature vectors. Let's consider we have generated feature vectors by CNN with lower dimensions, can we use the generated dense vector representations to describe the original sequence data?
You can have a look at the Eigenvectors of your data. Regarding the reconstruction, you want to get an autoencoder and also use the decoder part.
Otherwise, can you make your question more precise?
- Badges
- Science topic