Representation Learning

• The goal of representation learning is to extract sufficient but minimal information from data
• Representation learning builds on top of feature engineering by:
  • Being automated
  • Not requiring domain expertise.
  • Being unbiased
• The evaluation of a learned representation is related to the performance of downstream tasks that use this representation

Topics §

• Convolutional Neural Network - the convolutional kernels in a CNN are feature extractors.
• Encoder-Decoder Network - Encoders in particular can produce latent space representations.
• Transformer Model - The attention layers used by a transformer model provide a form of representation learning.
• Graph Neural Network - Learns representation for network based data.
  • Graph Embedding - non deep learning related techniques for embedding graphs
• Transfer Learning
• Belief Network

Links §

• Graph Neural Networks — Foundations, Frontiers and Applications by Wu et al.