Risk versus Uncertainty in Deep Learning: Bayes, Bootstrap and the Dangers of Dropout

By Ian Osband. 2016

136 citations recorded by the source corpus.

Subfields: Probabilistic & Bayesian Methods

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Probabilistic & Bayesian Methods

Citations in the corpus

A General Framework for Uncertainty Estimation in Deep Learning (2019, background): “A possible solution is to tune the dropout rates [16], however it is always possible to construct examples where this approach would generate erroneous predictions [4].”
Deep Reinforcement Learning in a Handful of Trials using Probabilistic Dynamics Models (2018, background): “Constructing good Bayesian NN models remains an open problem [35], [10], [36], [37], although recent promising work”
Lightweight Probabilistic Deep Networks (2018, background): “One limitation is the possibility to construct examples, where test-time dropout does not calibrate its predicted uncertainty to the observed variance in the data [43].”
Randomized Prior Functions for Deep Reinforcement Learning (2018, background): “This narrative has proved popular despite the fact that dropout distribution can be a poor approximation to most reasonable Bayesian posteriors [22, 46]: Lemma 1 (Dropout distribution does not concentrate with observed data).”
Pushing the bounds of dropout (2018, background): “Coming from another angle, Osband (2016) makes the point that in variational dropout the posterior over weights does not concentrate with more data, unlike for example in Graves (2011), which is unexpected behaviour from a Bayesian model.”
Hypermodels for Exploration (2020, background): “Our motivation is that intelligent ∗DeepMind (1)Although later work suggests that this dropout approximation can be of poor quality (Osband, 2016; Hron et al., 2017).”
Bayesian Prediction of Future Street Scenes using Synthetic Likelihoods (2018, background): “…is multimodal and the model set under consideration do not have explicit latent state/variables (as most popular deep deep neural network architectures), the approach of Gal & Ghahramani (2016b); Kendall & Gal (2017) is unable to recover the true model uncertainty (see Figure 1 and Osband (2016)).”
TRADI: Tracking deep neural network weight distributions (2019, background): “A potential drawback of MC Dropout concerns the fact that its uncertainty is not reducing with more training steps [41,42].”
Safe Imitation Learning via Fast Bayesian Reward Inference from Preferences (2020, background): “Recent work has shown that MC Dropout is not a principled Bayesian approximation since the distribution obtained from MC Dropout does not concentrate in the limit as the number of data samples goes to infinity and thus does not necessarily measure the kind of epistemic risk we are interested in (Osb”