We propose a method for transfer learning for deep networks through Bayesian inference, where an approximate posterior distribution q(w|θ) of model parameters w is learned through variational approximation. Utilizing Bayes by Backprop we optimize the parameters θ associated with the approximate distribution. When performing transfer learning we consider two tasks; A and B. Firstly, an approximate posterior q_A(w|θ) is learned from task A which is afterwards transferred as a prior p(w) → q_A(w|θ) when learning the approximate posterior distribution q_B(w|θ) for task B. Initially, we consider a multivariate normal distribution q(w|θ) = N (µ, Σ), with diagonal covariance matrix Σ. Secondly, we consider the prospects of introducing more expressive approximate distributions - specifically those known as normalizing flows. By investigating these concepts on the MNIST data set we conclude that utilizing normalizing flows does not improve Bayesian inference in the context presented here. Further, we show that transfer learning is not feasible using our proposed architecture and our definition of task A and task B, but no general conclusion regarding rejecting a Bayesian approach to transfer learning can be made.
Comments: 6 Pages.
[v1] 2018-01-09 11:34:24
Unique-IP document downloads: 12 times
Vixra.org is a pre-print repository rather than a journal. Articles hosted may not yet have been verified by peer-review and should be treated as preliminary. In particular, anything that appears to include financial or legal advice or proposed medical treatments should be treated with due caution. Vixra.org will not be responsible for any consequences of actions that result from any form of use of any documents on this website.
Add your own feedback and questions here:
You are equally welcome to be positive or negative about any paper but please be polite. If you are being critical you must mention at least one specific error, otherwise your comment will be deleted as unhelpful.