I am interested in the role action has on learning to perceive. I believe that not only does the ability to interact with the environment help produce better structured perceptual representations, but that interaction is necessary for obtaining meaningful representations. Interaction offers a probe on the hidden geometry of the states of the environment. 

To prove my point, I explore how unsupervised learning augmented with an equivariance constraint leads to the disentanglement of the generative factors of the states space.

unsupervised learning geometry equivariance disentanglement