Evaluating uncertainty estimation

bayesian deep learning vs. point estimate deep learning

• Bayesian deep learning methods

  • MC-Dropout
  • SWAG
  • DPN
  • JEM

• Datasets

  Train/In Distribution	Test/Out Distribution
  CIFAR10	{CIFAR100, SVHN, LSUN}
  FashionMNIST	{CIFAR100, CIFAR10, LSUN}
  CIFAR100	{CIFAR10, SVHN, LSUN}
  SVHN	{CIFAR100, CIFAR10, LSUN}

• Pretrained Models

  Pretrained WideResNet28x10 on CIFAR10
  DNN
  MC-Dropout
  SWAG
  DPN
  JEM

• Results

• Code Attribution

  • Stochastic Weight Averaging of Gaussian
  • Dirichlet Prior Networks
  • Joint Energy Models

• Cite

@inproceedings{mitros2020,
  author =        {Mitros, John and Pakrashi, Arjun and
                   Mac Namee, Brian},
  pages =         {71-87},
  series =        {European Conference on Computer Vision},
  title =         {Ramifications of Approximate Posterior Inference for
                   Bayesian Deep Learning in Adversarial and
                   Out-of-Distribution Settings},
  year =          {2020},
  publisher =     {Springer},
  doi =           {10.1007/978-3-030-66415-2_5},
  isbn =          {978-3-030-66414-5},
}