thoughtidearepresentation-learning

motivation

  • each piece of information has its own representation
  • different tasks might require different amount of information
  • it is not practical to represent all information in one latent space
    • information could be different in various characteristics: modality, domain, complexity
    • represent everything in one giant space is not efficient

Hence, the representations need to be combined losslessly

  • from the combination, we should be able to get back each component
  • one simple way is concatenation
  • or using multiple modular modules

How to design an architecture that would be able to work with dynamic-sized representations ?question Do we still need, at some point, a unified latent space ?