Deep Ensembles: A Loss Landscape Perspective

Deep ensembles have been empirically shown to be a promising approach for improving accuracy, uncertainty and out-of-distribution robustness of deep learning models. While deep ensembles were theoretically motivated by the bootstrap, non-bootstrap ensembles trained with just random initialization also perform well in practice, which suggests that there could be other explanations for why deep ensembles work well. Bayesian neural networks, which learn distributions over the parameters of the network, are theoretically well-motivated by Bayesian principles, but do not perform as well as deep ensembles in practice, particularly under dataset shift. One possible explanation for this gap between theory and practice is that popular scalable variational Bayesian methods tend to focus on a single mode, whereas deep ensembles tend to explore diverse modes in function space. We investigate this hypothesis by building on recent work on understanding the loss landscape of neural networks and adding our own exploration to measure the similarity of functions in the space of predictions. Our results show that random initializations explore entirely different modes, while functions along an optimization trajectory or sampled from the subspace thereof cluster within a single mode predictions-wise, while often deviating significantly in the weight space. Developing the concept of the diversity--accuracy plane, we show that the decorrelation power of random initializations is unmatched by popular subspace sampling methods. Finally, we evaluate the relative effects of ensembling, subspace based methods and ensembles of subspace based methods, and the experimental results validate our hypothesis.

[1]  G. Brier VERIFICATION OF FORECASTS EXPRESSED IN TERMS OF PROBABILITY , 1950 .

[2]  Sebastian Nowozin,et al.  Can You Trust Your Model's Uncertainty? Evaluating Predictive Uncertainty Under Dataset Shift , 2019, NeurIPS.

[3]  Thomas B. Schön,et al.  Evaluating Scalable Bayesian Deep Learning Methods for Robust Computer Vision , 2019, 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW).

[4]  Aaron Klein,et al.  Bayesian Optimization with Robust Bayesian Neural Networks , 2016, NIPS.

[5]  D. Mackay,et al.  Bayesian methods for adaptive models , 1992 .

[6]  Andrew Gordon Wilson,et al.  A Simple Baseline for Bayesian Uncertainty in Deep Learning , 2019, NeurIPS.

[7]  Andrew Gordon Wilson,et al.  Loss Surfaces, Mode Connectivity, and Fast Ensembling of DNNs , 2018, NeurIPS.

[8]  Yee Whye Teh,et al.  Bayesian Learning via Stochastic Gradient Langevin Dynamics , 2011, ICML.

[9]  Andrew Y. Ng,et al.  Reading Digits in Natural Images with Unsupervised Feature Learning , 2011 .

[10]  Oriol Vinyals,et al.  Qualitatively characterizing neural network optimization problems , 2014, ICLR.

[11]  Andrew Gordon Wilson,et al.  Cyclical Stochastic Gradient MCMC for Bayesian Deep Learning , 2019, ICLR.

[12]  Leo Breiman,et al.  Bagging Predictors , 1996, Machine Learning.

[13]  Jason Yosinski,et al.  Measuring the Intrinsic Dimension of Objective Landscapes , 2018, ICLR.

[14]  Zoubin Ghahramani,et al.  Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning , 2015, ICML.

[15]  Geoffrey E. Hinton,et al.  Visualizing Data using t-SNE , 2008 .

[16]  Michael Cogswell,et al.  Why M Heads are Better than One: Training a Diverse Ensemble of Deep Networks , 2015, ArXiv.

[17]  Julien Cornebise,et al.  Weight Uncertainty in Neural Network , 2015, ICML.

[18]  Fred A. Hamprecht,et al.  Essentially No Barriers in Neural Network Energy Landscape , 2018, ICML.

[19]  Jian Sun,et al.  Identity Mappings in Deep Residual Networks , 2016, ECCV.

[20]  Fei-Fei Li,et al.  ImageNet: A large-scale hierarchical image database , 2009, 2009 IEEE Conference on Computer Vision and Pattern Recognition.

[21]  Alex Graves,et al.  Practical Variational Inference for Neural Networks , 2011, NIPS.

[22]  Alex Kendall,et al.  What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision? , 2017, NIPS.

[23]  Dustin Tran,et al.  Flipout: Efficient Pseudo-Independent Weight Perturbations on Mini-Batches , 2018, ICLR.

[24]  Geoffrey E. Hinton,et al.  Bayesian Learning for Neural Networks , 1995 .

[25]  David M. Blei,et al.  Stochastic Gradient Descent as Approximate Bayesian Inference , 2017, J. Mach. Learn. Res..

[26]  Andrew Gordon Wilson,et al.  Averaging Weights Leads to Wider Optima and Better Generalization , 2018, UAI.

[27]  Stanislav Fort,et al.  Large Scale Structure of Neural Network Loss Landscapes , 2019, NeurIPS.

[28]  Max Welling,et al.  Multiplicative Normalizing Flows for Variational Bayesian Neural Networks , 2017, ICML.

[29]  Thomas G. Dietterich,et al.  Benchmarking Neural Network Robustness to Common Corruptions and Perturbations , 2018, ICLR.

[30]  Julien Cornebise,et al.  Weight Uncertainty in Neural Networks , 2015, ArXiv.

[31]  Charles Blundell,et al.  Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles , 2016, NIPS.

[32]  Jian Sun,et al.  Deep Residual Learning for Image Recognition , 2015, 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).

[33]  Stanislav Fort,et al.  The Goldilocks zone: Towards better understanding of neural network loss landscapes , 2018, AAAI.

[34]  Jimmy Ba,et al.  Adam: A Method for Stochastic Optimization , 2014, ICLR.

[35]  Nitish Srivastava,et al.  Dropout: a simple way to prevent neural networks from overfitting , 2014, J. Mach. Learn. Res..

[36]  Alex Krizhevsky,et al.  Learning Multiple Layers of Features from Tiny Images , 2009 .