Similarity of Neural Networks with Gradients

  • 2020-03-25 17:04:10
  • Shuai Tang, Wesley J. Maddox, Charlie Dickens, Tom Diethe, Andreas Damianou
  • 8


A suitable similarity index for comparing learnt neural networks plays animportant role in understanding the behaviour of the highly-nonlinearfunctions, and can provide insights on further theoretical analysis andempirical studies. We define two key steps when comparing models: firstly, therepresentation abstracted from the learnt model, where we propose to leverageboth feature vectors and gradient ones (which are largely ignored in priorwork) into designing the representation of a neural network. Secondly, wedefine the employed similarity index which gives desired invariance properties,and we facilitate the chosen ones with sketching techniques for comparingvarious datasets efficiently. Empirically, we show that the proposed approachprovides a state-of-the-art method for computing similarity of neural networksthat are trained independently on different datasets and the tasks defined bythe datasets.


Quick Read (beta)

loading the full paper ...