Gradient-based Bayesian Experimental Design for Implicit Models using Mutual Information Lower Bounds

Kleinegesse, Steven; Gutmann, Michael U.

doi:10.48550/arxiv.2105.04379

Cited by 7 publications

(20 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We can use this to help approximate the posterior over θ given the collected real data from the experiment. This means we can perform likelihood-free inference after training the critic, which extends previous results [28,29] from the static to the adaptive policy-based setting.…”

Section: Information Lower Bounds For Policy-based Experimental Desig...supporting

confidence: 72%

“…To establish a suitable likelihood-free training objective for the implicit setting, our high-level idea is to leverage recent advances in variational MI [see 42, for an overview], which have shown promise for static BOED [16,28,29]. While using these bounds in the traditional framework of (2) would not permit real-time experiments, one could consider a naive application of them to the policy objective of (3) by replacing each I ht−1 with a suitable variational lower bound that uses a 'critic' U t : H t−1 × Θ → R to avoid explicit likelihood evaluations, where H t−1 and Θ are the spaces of histories and parameters respectively.…”

Section: Information Lower Bounds For Policy-based Experimental Desig...mentioning

confidence: 99%

“…More recently, gradient-based approaches have been introduced [15,28], some of which allow the estimation and optimization simultaneously in a single stochastic-gradient scheme [16,23,29].…”

Section: Related Workmentioning

confidence: 99%

“…From a posterior estimation perspective, likelihood-free inference can be performed via approximate Bayesian computation [33,54], ratio estimation [56], conventional MCMC for methods that make tractable approximation to the likelihood [23,24], or as a byproduct of MI estimation [16,27,29,30].…”

Section: Related Workmentioning

confidence: 99%

See 3 more Smart Citations

Implicit Deep Adaptive Design: Policy-Based Experimental Design without Likelihoods

Ivanova¹,

Foster²,

Kleinegesse³

et al. 2021

Preprint

Self Cite

View full text Add to dashboard Cite

We introduce implicit Deep Adaptive Design (iDAD), a new method for performing adaptive experiments in real-time with implicit models. iDAD amortizes the cost of Bayesian optimal experimental design (BOED) by learning a design policy network upfront, which can then be deployed quickly at the time of the experiment. The iDAD network can be trained on any model which simulates differentiable samples, unlike previous design policy work that requires a closed form likelihood and conditionally independent experiments. At deployment, iDAD allows design decisions to be made in milliseconds, in contrast to traditional BOED approaches that require heavy computation during the experiment itself. We illustrate the applicability of iDAD on a number of experiments, and show that it provides a fast and effective mechanism for performing adaptive design with implicit models. Recently, Foster et al. [17] proposed an exciting alternative approach, called Deep Adaptive Design (DAD), that is based on learning design policies. DAD provides a way to avoid significant computation 35th Conference on Neural Information Processing Systems (NeurIPS 2021).

show abstract

Section: Information Lower Bounds For Policy-based Experimental Desig...supporting

confidence: 72%

Section: Information Lower Bounds For Policy-based Experimental Desig...mentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

See 2 more Smart Citations

Implicit Deep Adaptive Design: Policy-Based Experimental Design without Likelihoods

Ivanova¹,

Foster²,

Kleinegesse³

et al. 2021

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…which quantifies the extent to which observation of the datum δ(f) changes a posteriori belief; here KL denotes the Kullback-Leibler divergence. For related approaches and discussion see the recent survey in Kleinegesse and Gutmann (2021). However, in the setting where data are exactly observed, the two distributions in (3) will be mutually singular and the Kullback-Leibler divergence will not exist.…”

Section: Sequential Experimental Designmentioning

confidence: 99%

GaussED: A Probabilistic Programming Language for Sequential Experimental Design

Fisher,

Teymur,

Oates

2021

Preprint

View full text Add to dashboard Cite

Sequential algorithms are popular for experimental design, enabling emulation, optimisation and inference to be efficiently performed. For most of these applications bespoke software has been developed, but the approach is general and many of the actual computations performed in such software are identical. Motivated by the diverse problems that can in principle be solved with common code, this paper presents GaussED, a simple probabilistic programming language coupled to a powerful experimental design engine, which together automate sequential experimental design for approximating a (possibly nonlinear) quantity of interest in Gaussian processes models. Using a handful of commands, GaussED can be used to: solve linear partial differential equations, perform tomographic reconstruction from integral data and implement Bayesian optimisation with gradient data.

show abstract

Amortised Design Optimization for Item Response Theory

Keurulainen

Westerlund²,

Keurulainen³

et al. 2023

Communications in Computer and Information Science

View full text Add to dashboard Cite

I was fortunate to engage in numerous meetings, brainstorming sessions, and discussions with a variety of colleagues at Aalto University. I wish to express my gratitude to Jyri Kivinen, Vikas Verma, Mert Celikok, Tomi Peltola, and all the members of the Probabilistic Machine Learning (PML) group for their insightful and valuable contributions to these discussions. Furthermore, I had the pleasure of participating in many enriching conversations with Professor Antti Oulasvirta and the members of the Computational Behavior Lab at Aalto, led by Professor Oulasvirta.I am also grateful to Isak Westerlund, who has worked alongside me during my PhD project, taking on a significant amount of responsibility in coding the applications and training the various AI models. My thanks also go to Ariel Kwiatkowski, who contributed to our second paper.I am deeply grateful to the pre-examiners, Professor Anna Rafferty and Dr. Xiaoyang Wang, for their essential insights and constructive feedback on my thesis. In addition, I am thankful to Dr. John Williamson for willingly taking on the role of the opponent I want to thank Bitville Oy CEO Jouko Ranta, as well as all my colleagues at Bitville Oy, for making it possible for me to concentrate practically fulltime on the research projects during these years.I had the unique opportunity to share my research journey with my son, Oskar, who has now embarked on his own PhD journey at Aalto University as I write this preface. Thank you for the hundreds of hours of brainstorming, discussion, and reflection over nearly the last 10 years, and for your invaluable contribution as a member of the research team. Our discussions have been of the utmost importance to me, whether they were about mathematical details or the extensive influence of AI on society.Last but not least, I want to extend my heartfelt thanks to my daughter, Erika, and my wife, Sofia, for their encouragement and patience. I am aware that discussions about AI have prominently featured in our family conversations over the last few years. It is evident that this project could not have reached its conclusion without your unwavering support.

show abstract

Gradient-based Bayesian Experimental Design for Implicit Models using Mutual Information Lower Bounds

Cited by 7 publications

References 0 publications

Implicit Deep Adaptive Design: Policy-Based Experimental Design without Likelihoods

Implicit Deep Adaptive Design: Policy-Based Experimental Design without Likelihoods

GaussED: A Probabilistic Programming Language for Sequential Experimental Design

Amortised Design Optimization for Item Response Theory

Contact Info

Product

Resources

About