76 - Increasing In-Class Similarity by Retrofitting Embeddings with Demographics, with Dirk Hovy
NLP Highlights
00:00
How to Improve Linear Separability in Graph Convolutional Networks
In this case, we're giving a neural network or a classifier in general a leg up by making the classes more linearly separable and thereby basically infusing some outside information into the representations. Now you could do the same thing within a network in the class of graph convolutional networks. You're essentially learning this retrofitting matrix to the one step at a time as part of the training process. But it takes longer. It's more costly. So what I wonder about is linear separability because what you're doing in the end is a linear transformation on the same data space. Well, what it does is it increases the in class similarity and it should increase the separability or
Transcript
Play full episode
