search

HHRClemson / Flood-stAge-predIction-thRough-Deep-Neural-Networks

This is the official code repository for the paper An End-to-End Flood Detection System Using Deep Neural Networks by Windheuser, Karanjit, Pally, Samadi, and Hubig.
MIT License
9 stars 0 forks source link

Supervised Variational Encoding Techniques #2

Closed VidyaSamadi closed 3 years ago

VidyaSamadi commented 3 years ago

How SVE works in encoder? other methods? uncertainty?

Rakshit-Pally26 commented 3 years ago

Attached below is a link to a paper explaining the same.

link: https://arxiv.org/abs/2010.02014

VidyaSamadi commented 3 years ago

another paper: https://www.sciencedirect.com/science/article/pii/S0020025518310077

Technically, SVE learns an encoder to induce the latent space for sampling from the data.

leonwind commented 3 years ago

Yes, I am aware of that. I just don't completely understand the difference between a SVE and a normal VAE, especially their NN architecture looks very similar, just that a SVE has an additional Bi-Interaction Pooling Layer, where this paper is also the only resource I found about Bi-Interaction Layers. However, this should not be the only difference.

VidyaSamadi commented 3 years ago

Bi-Interaction Pooling Layer seems to regulate the NN (deepen NN layers to learn meaningful information). It can be added to the network as a second-order interaction pooling layer. the interaction vectors are then input to a feed-forward neural network to model high-order feature interactions where the data are spare or the patterns are missing. It should provide a sum of interaction (apply to non-zero data) and can be applied on each hidden layer along with dropout. with Bi-interaction, NFM would be easier to train. This paper that Leon shared (Figure 2 clearly shows the bi-interaction structure) used Batch Normalization on bi-interaction for each hidden layer. These papers are also good to read -paper-1 paper-2 paper-3. More tomorrow.