Black-box $\alpha$-divergence Minimization
By José Miguel Hernández-Lobato, Yingzhen Li, Mark Rowland, Daniel Hernández-Lobato, Thang D. Bui, Richard E. Turner. 2015
152 citations recorded by the source corpus.
Black-box alpha (BB-$\alpha$) is a new approximate inference method based on the minimization of $\alpha$-divergences. BB-$\alpha$ scales to large datasets because it can be implemented using stochastic gradient descent. BB-$\alpha$ can be applied to complex probabilistic models with little effort since it only requires as input the likelihood function and its gradients. These gradients can be easily obtained using automatic differentiation. By changing the divergence parameter $\alpha$, the method is able to interpolate between variational Bayes (VB) ($\alpha \rightarrow 0$) and an algorithm similar to expectation propagation (EP) ($\alpha = 1$). Experiments on probit regression and neural network regression and classification problems show that BB-$\alpha$ with non-standard settings of $\alpha$, such as $\alpha = 0.5$, usually produces better predictions than with $\alpha \rightarrow 0$ (VB) or $\alpha = 1$ (EP).
Subfields: Artificial Intelligence, Control and Systems Engineering, Optimization & Training, Information Theory
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Artificial Intelligence
- Sergey Levine
- Yoshua Bengio
- Yarin Gal
- Richard E. Turner
- Mohit Bansal
- Ruslan Salakhutdinov
- Pieter Abbeel
- Quoc V. Le
In Control and Systems Engineering
- Yarin Gal
- Lucy R. Chai
- Richard E. Turner
- Yingzhen Li
- José Miguel Hernández-Lobato
- Qiang Liu
- Finale Doshi-Velez
- Zoubin Ghahramani
In Optimization & Training
- Lucy R. Chai
- Qiang Liu
- José Miguel Hernández-Lobato
- Yarin Gal
- Yingzhen Li
- Finale Doshi-Velez
- Richard E. Turner
- Zoubin Ghahramani
In Information Theory
- Qiang Liu
- Lucy R. Chai
- Finale Doshi-Velez
- Yingzhen Li
- José Miguel Hernández-Lobato
- Richard E. Turner
- Cuong V Nguyen
- Steffen Udluft
Citations in the corpus
- Variational Continual Learning (2017, background): “…1992), variational inference (Hinton and Van Camp, 1993; Barber and Bishop, 1998; Graves, 2011; Blundell et al., 2015; Gal and Ghahramani, 2016), sequential Monte Carlo (de Freitas et al., 2000), EP (Hernández-Lobato and Adams, 2015), and approximate Power EP (Hernández-Lobato et al., 2016).”
- Deep Reinforcement Learning in a Handful of Trials using Probabilistic Dynamics Models (2018, background): “…Bayesian NN models remains an open problem [MacKay, 1992, Neal, 1995, Osband, 2016, Guo et al., 2017], although recent promising work exists on incorporating dropout [Gal et al., 2017], ensembles [Osband et al., 2016, Lakshminarayanan et al., 2017], and α-divergence [Hernández-Lobato et al., 2016].”
- Dropout Inference in Bayesian Neural Networks with Alpha-divergences (2017, methodology): “(3) One could verify that this is the same energy function as presented in (Hernández-Lobato et al., 2016) by considering q an exponential family distribution.”
- Learning and Policy Search in Stochastic Dynamical Systems with Bayesian Neural Networks (2016, methodology): “It can be shown that minimizing (6) when α→ 0 is equivalent to adjusting q by VB [10].”
- Task-Agnostic Continual Learning Using Online Variational Bayes With Fixed-Point Updates (2020, methodology): “Those methods include assumed density filtering (Soudry, Hubara, & Meir, 2014; Hernández-Lobato & Adams, 2015), approximate power expectation propagation (Hernández-Lobato et al., 2016), stochastic Langevin gradient descent (Welling & Teh, 2011; Balan, Rathod, Murphy, & Welling, 2015), incremental m”
- Gaussian Process Behaviour in Wide Deep Neural Networks (2018, background): “Fully connected Bayesian deep networks with finite variance priors on the weights have also been considered in other works (Graves, 2011; Hernández-Lobato et al., 2016; Blundell et al., 2015), though the specific details vary.”
- Structured Variational Learning of Bayesian Neural Networks with Horseshoe Priors (2018, methodology): “, 2014), black-box variational and alphadivergence minimization (Hernandez-Lobato et al., 2016; Ranganath et al., 2014), and probabilistic backpropagation (Hernández-Lobato & Adams, 2015) have reinvigorated interest in BNNs by allowing scalable inference.”
- Distributed Bayesian Learning with Stochastic Natural Gradient Expectation Propagation and the Posterior Server (2015, methodology): “Black-box methods have recently been developed for variational Bayes (Ranganath et al., 2014, Black-box Variational Inference or BBVI) and for power EP (Hernandez-Lobato et al., 2015, Black-box Alpha or BB-α).”
- A Unifying Framework for Gaussian Process Pseudo-Point Approximations using Power Expectation Propagation (2016, result): “Similar empirical conclusions have been made elsewhere (see e.g. Minka, 2005; Hernández-Lobato et al., 2016; Depeweg et al., 2016).”
- Partitioned Variational Inference: A unified framework encompassing federated and continual learning (2018, methodology): “Honkela et al. [2010] considered non-conjugate models, employed a Gaussian variational distribution and used natural gradients to update the variational distribution’s mean.”
- Decomposition of Uncertainty in Bayesian Deep Learning for Efficient and Risk-sensitive Learning (2017, unknown): “The reader is referred to the work of Hernández-Lobato et al. (2016); Depeweg et al. (2016) for more details on this.”
- Wild Variational Approximations (2016, methodology): “We present all these proposals in VI context (which is then named wild variational inference), although they can be extended to other methods such as Bethe free energy [14, 15] and α-divergence methods [16, 17].”
- Model Selection in Bayesian Neural Networks via Horseshoe Priors (2017, background): “These properties have resulted in BNNs being adopted in applications ranging from active learning [11, 8] and reinforcement learning [3].”
- Deep and Confident Prediction for Time Series at Uber (2017, background): “Several algorithms further extend the approximation framework to α-divergence optimization, including [19], [20].”
- Improving Confidence Estimates for Unfamiliar Examples (2018, background): “Bayesian approaches [1, 2, 17] estimate a distribution of network parameters and produce a Bayesian estimate for likelihood.”
- Uncertainty Estimation in Bayesian Neural Networks And Links to Interpretability (2018, methodology): “Secondly, BNNs can produce uncertainty estimates of varying quality depending on the divergence objective [31] and an approximation gap (the inability of the variational family to truly approximate the posterior distribution) [13].”
- Advances in Variational Inference (2017, methodology): “a-divergences have recently been used in variational inference [103], [104].”
- Model-Based Reinforcement Learning for Sepsis Treatment (2018, background): “) , and minimise the alpha-divergence between the true posterior and the variational approximation [5].”