• why linear autoencoder gives PCA ? (see fig)
  • Why does P have to be orthonormal?
    • one simple explanation
    • find such that satisfies
    • to find the second direction, u need to have some imposition other wise it wil always find p1 ( the above condition comes from minimizing || x - Uz || = || x - U U .T z ||)
  • link to my GPTized handwritten notes

Explanation to below figure

  • the neural network is tasked to optimize this
  • if W_enc and W_dec are orthogonal and transpose of each other we get PCA. U and U.T
  • but the matrices are unidentifiable. U can multiple W_enc → A. W_enc and W_dec → W_dec A
  • u still get the same solution, so its not ultimately PCA. But PCA latent state multiplied by another matrix
  • Now if u put orthonormal constraint, u get rotation. If not, u get any general invertible matrix