Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations

Explore a critical analysis of unsupervised learning of disentangled representations in this informative video. Delve into the theoretical impossibility of unsupervised disentanglement without inductive biases, and examine the results of an extensive experimental study involving over 12,000 models. Discover the implications for future research in disentanglement learning, including the need for explicit consideration of inductive biases, investigation of concrete benefits, and reproducible experimental setups across multiple datasets.

Intro
AutoEncoder
Loss Function
Random Error
Theorem
Experiments