https://arxiv.org/abs/2002.03704
Liberty or Depth: Deep Bayesian Neural Nets Do Not Need Complex Weight Posterior Approximations (Sebastian Farquhar, Lewis Smith, Yarin Gal)
Try Depth Instead of Weight Correlations: Mean-field is a Less Restrictive Assumption for Deeper Networks (Sebastian Farquhar, Lewis Smith, Yarin Gal)
Mean field approximation이 지나치게 강한 가정이라고 생각해왔지만 깊은 mean field approximate 모델이 shallow full covariance 모델을 근사할 수 있다는 주장. 깊은 모델을 고려하는 것이 중요.
#bayesian #mean_field