Timezone: »
The Bayesian paradigm has the potential to solve core issues of deep neural networks such as poor calibration and data inefficiency. Alas, scaling Bayesian inference to large weight spaces often requires restrictive approximations. In this work, we show that it suffices to perform inference over a small subset of model weights in order to obtain accurate predictive posteriors. The other weights are kept as point estimates. This subnetwork inference framework enables us to use expressive, otherwise intractable, posterior approximations over such subsets. In particular, we implement subnetwork linearized Laplace as a simple, scalable Bayesian deep learning method: We first obtain a MAP estimate of all weights and then infer a full-covariance Gaussian posterior over a subnetwork using the linearized Laplace approximation. We propose a subnetwork selection strategy that aims to maximally preserve the model’s predictive uncertainty. Empirically, our approach compares favorably to ensembles and less expressive posterior approximations over full networks.
Author Information
Erik Daxberger (University of Cambridge & MPI for Intelligent Systems, Tübingen)
Eric Nalisnick (University of Amsterdam)
James Allingham (University of Cambridge)
Javier Antorán (University of Cambridge)
I am a PhD student in Machine Learning at the University of Cambridge under the supervision of Dr. José Miguel Hernández-Lobato. I’m interested in Bayesian deep learning, representation learning, uncertainty in machine learning and information theory. I graduated from the University of Zaragoza in 2018 with an honorary distinction (“premio extraordinario”) in Telecommunications Engineering (EE/CS). I was awarded an MPhil in Machine Learning with distinction by the University of Cambridge in 2019. I also do freelance engineering consulting and am a co-founder of arisetech.es.
Jose Miguel Hernandez-Lobato (University of Cambridge)
Related Events (a corresponding poster, oral, or spotlight)
-
2021 Poster: Bayesian Deep Learning via Subnetwork Inference »
Tue. Jul 20th 04:00 -- 06:00 PM Room Virtual
More from the Same Authors
-
2021 : Bayesian Regression from Multiple Sources of Weak Supervision »
Putra Manggala · Holger Hoos · Eric Nalisnick · Putra Manggala -
2023 : Leveraging Task Structures for Improved Identifiability in Neural Network Representations »
Wenlin Chen · Julien Horwood · Juyeon Heo · Jose Miguel Hernandez-Lobato -
2023 : Minimal Random Code Learning with Mean-KL Parameterization »
Jihao Andreas Lin · Gergely Flamich · Jose Miguel Hernandez-Lobato -
2023 : Memory Maps to Understand Models »
Dharmesh Tailor · Paul Chang · Siddharth Swaroop · Eric Nalisnick · Arno Solin · Khan Emtiyaz -
2023 Poster: A Simple Zero-shot Prompt Weighting Technique to Improve Prompt Ensembling in Text-Image Models »
James Allingham · JIE REN · Michael Dusenberry · Xiuye Gu · Yin Cui · Dustin Tran · Jeremiah Liu · Balaji Lakshminarayanan -
2022 : A Product of Experts Approach to Early-Exit Ensembles »
James Allingham · Eric Nalisnick -
2022 Poster: Adapting the Linearised Laplace Model Evidence for Modern Deep Learning »
Javier Antorán · David Janz · James Allingham · Erik Daxberger · Riccardo Barbano · Eric Nalisnick · Jose Miguel Hernandez-Lobato -
2022 Spotlight: Adapting the Linearised Laplace Model Evidence for Modern Deep Learning »
Javier Antorán · David Janz · James Allingham · Erik Daxberger · Riccardo Barbano · Eric Nalisnick · Jose Miguel Hernandez-Lobato -
2022 Poster: Action-Sufficient State Representation Learning for Control with Structural Constraints »
Biwei Huang · Chaochao Lu · Liu Leqi · Jose Miguel Hernandez-Lobato · Clark Glymour · Bernhard Schölkopf · Kun Zhang -
2022 Spotlight: Action-Sufficient State Representation Learning for Control with Structural Constraints »
Biwei Huang · Chaochao Lu · Liu Leqi · Jose Miguel Hernandez-Lobato · Clark Glymour · Bernhard Schölkopf · Kun Zhang -
2022 Poster: Calibrated Learning to Defer with One-vs-All Classifiers »
Rajeev Verma · Eric Nalisnick -
2022 Poster: Fast Relative Entropy Coding with A* coding »
Gergely Flamich · Stratis Markou · Jose Miguel Hernandez-Lobato -
2022 Spotlight: Fast Relative Entropy Coding with A* coding »
Gergely Flamich · Stratis Markou · Jose Miguel Hernandez-Lobato -
2022 Spotlight: Calibrated Learning to Defer with One-vs-All Classifiers »
Rajeev Verma · Eric Nalisnick -
2021 Poster: Active Slices for Sliced Stein Discrepancy »
Wenbo Gong · Kaibo Zhang · Yingzhen Li · Jose Miguel Hernandez-Lobato -
2021 Spotlight: Active Slices for Sliced Stein Discrepancy »
Wenbo Gong · Kaibo Zhang · Yingzhen Li · Jose Miguel Hernandez-Lobato -
2021 Poster: A Gradient Based Strategy for Hamiltonian Monte Carlo Hyperparameter Optimization »
Andrew Campbell · Wenlong Chen · Vincent Stimper · Jose Miguel Hernandez-Lobato · Yichuan Zhang -
2021 Spotlight: A Gradient Based Strategy for Hamiltonian Monte Carlo Hyperparameter Optimization »
Andrew Campbell · Wenlong Chen · Vincent Stimper · Jose Miguel Hernandez-Lobato · Yichuan Zhang -
2020 : "Latent Space Optimization with Deep Generative Models" »
Jose Miguel Hernandez-Lobato -
2020 : Invited talk 2: Detecting Distribution Shift with Deep Generative Models »
Eric Nalisnick -
2020 : Spotlight Talk 3: Depth Uncertainty in Neural Networks »
Javier Antorán · James Allingham -
2020 : Invited Talk: Efficient Missing-value Acquisition with Variational Autoencoders »
Jose Miguel Hernandez-Lobato -
2020 Poster: Reinforcement Learning for Molecular Design Guided by Quantum Mechanics »
Gregor Simm · Robert Pinsler · Jose Miguel Hernandez-Lobato -
2020 Poster: A Generative Model for Molecular Distance Geometry »
Gregor Simm · Jose Miguel Hernandez-Lobato -
2019 Poster: Dropout as a Structured Shrinkage Prior »
Eric Nalisnick · Jose Miguel Hernandez-Lobato · Padhraic Smyth -
2019 Oral: Dropout as a Structured Shrinkage Prior »
Eric Nalisnick · Jose Miguel Hernandez-Lobato · Padhraic Smyth -
2019 Oral: Hybrid Models with Deep and Invertible Features »
Eric Nalisnick · Akihiro Matsukawa · Yee-Whye Teh · Dilan Gorur · Balaji Lakshminarayanan -
2019 Poster: EDDI: Efficient Dynamic Discovery of High-Value Information with Partial VAE »
Chao Ma · Sebastian Tschiatschek · Konstantina Palla · Jose Miguel Hernandez-Lobato · Sebastian Nowozin · Cheng Zhang -
2019 Poster: Variational Implicit Processes »
Chao Ma · Yingzhen Li · Jose Miguel Hernandez-Lobato -
2019 Poster: Hybrid Models with Deep and Invertible Features »
Eric Nalisnick · Akihiro Matsukawa · Yee-Whye Teh · Dilan Gorur · Balaji Lakshminarayanan -
2019 Oral: Variational Implicit Processes »
Chao Ma · Yingzhen Li · Jose Miguel Hernandez-Lobato -
2019 Oral: EDDI: Efficient Dynamic Discovery of High-Value Information with Partial VAE »
Chao Ma · Sebastian Tschiatschek · Konstantina Palla · Jose Miguel Hernandez-Lobato · Sebastian Nowozin · Cheng Zhang -
2018 Poster: Decomposition of Uncertainty in Bayesian Deep Learning for Efficient and Risk-sensitive Learning »
Stefan Depeweg · Jose Miguel Hernandez-Lobato · Finale Doshi-Velez · Steffen Udluft -
2018 Oral: Decomposition of Uncertainty in Bayesian Deep Learning for Efficient and Risk-sensitive Learning »
Stefan Depeweg · Jose Miguel Hernandez-Lobato · Finale Doshi-Velez · Steffen Udluft -
2017 Poster: Parallel and Distributed Thompson Sampling for Large-scale Accelerated Exploration of Chemical Space »
Jose Miguel Hernandez-Lobato · James Requeima · Edward Pyzer-Knapp · Alan Aspuru-Guzik -
2017 Poster: Grammar Variational Autoencoder »
Matt J. Kusner · Brooks Paige · Jose Miguel Hernandez-Lobato -
2017 Talk: Grammar Variational Autoencoder »
Matt J. Kusner · Brooks Paige · Jose Miguel Hernandez-Lobato -
2017 Talk: Parallel and Distributed Thompson Sampling for Large-scale Accelerated Exploration of Chemical Space »
Jose Miguel Hernandez-Lobato · James Requeima · Edward Pyzer-Knapp · Alan Aspuru-Guzik -
2017 Poster: Sequence Tutor: Conservative fine-tuning of sequence generation models with KL-control »
Natasha Jaques · Shixiang Gu · Dzmitry Bahdanau · Jose Miguel Hernandez-Lobato · Richard E Turner · Douglas Eck -
2017 Talk: Sequence Tutor: Conservative fine-tuning of sequence generation models with KL-control »
Natasha Jaques · Shixiang Gu · Dzmitry Bahdanau · Jose Miguel Hernandez-Lobato · Richard E Turner · Douglas Eck