Corpus ID: 225066839

PEP: Parameter Ensembling by Perturbation

@article{Mehrtash2020PEPPE,
  title={PEP: Parameter Ensembling by Perturbation},
  author={Alireza Mehrtash and P. Abolmaesumi and P. Golland and T. Kapur and D. Wassermann and W. M. Wells},
  journal={ArXiv},
  year={2020},
  volume={abs/2010.12721}
}
  • Alireza Mehrtash, P. Abolmaesumi, +3 authors W. M. Wells
  • Published 2020
  • Physics, Computer Science, Mathematics
  • ArXiv
  • Ensembling is now recognized as an effective approach for increasing the predictive performance and calibration of deep networks. We introduce a new approach, Parameter Ensembling by Perturbation (PEP), that constructs an ensemble of parameter values as random perturbations of the optimal parameter set from training by a Gaussian with a single variance parameter. The variance is chosen to maximize the log-likelihood of the ensemble average ($\mathbb{L}$) on the validation data set. Empirically… CONTINUE READING
    1 Citations

    Figures and Tables from this paper

    Encoding the latent posterior of Bayesian Neural Networks for uncertainty quantification
    • PDF

    References

    SHOWING 1-10 OF 43 REFERENCES
    Snapshot Ensembles: Train 1, get M for free
    • 334
    • PDF
    Adam: A Method for Stochastic Optimization
    • 56,546
    • PDF
    Bayesian Uncertainty Estimation for Batch Normalized Deep Networks
    • 97
    • PDF
    A Simple Baseline for Bayesian Uncertainty in Deep Learning
    • 149
    • PDF