search

flrngel / understanding-ai

personal repository
36 stars 6 forks source link

Subspace match probably does not accurately assess the similarity of learned representations #25

Open flrngel opened 5 years ago

flrngel commented 5 years ago

https://arxiv.org/pdf/1901.00884v1.pdf

image image

4. Conclusion

  1. It is easy to construct examples where the intermediate representations have no match, even when the outputs of the network are identical.
  2. In example 3.1, the activation vectors at layers with no subspace match may span isomorphic subspaces. The existence of an isomorphism between the subspaces suggests that we should be considering these representations as identical.