Enhances robustness in experimental design through Generalised Bayesian inference.
problem Poor inference and estimates of information gain when statistical model is incorrectly specified.
method Generalised Bayesian (Gibbs) inference framework applied to experimental design.
result GBOED enhances robustness to outliers and incorrect assumptions about noise distribution.
Bayesian inference uses Stein discrepancy for robustness in intractable likelihoods.
problem Intractable likelihoods in Bayesian inference.
method Generalised Bayesian inference with Stein discrepancy as the loss function.
result Robust generalised posteriors with closed form or accessible using MCMC.
Develops a new Bayesian inference method for discrete data.
problem Computational challenges in discrete state spaces, especially intractable likelihoods.
method Uses a discrete Fisher divergence to update beliefs about model parameters, circumventing the intractable normalising constant.
result Establishes statistical properties of the generalised posterior and proposes a calibration approach.
Generalised Bayesian learning algorithms are increasingly popular in machine learning, due to their PAC generalisation properties and flexibility. The present paper aims at providing a self-contained survey on the resulting PAC-Bayes framework and some of its main theoretical and algorithmic developments.
Develops PAC-Bayesian framework for physics-informed machine learning.
problem Lack of statistical generalisation understanding for PIML models.
method PAC-Bayesian framework with multi-task perspective, incorporating physical structure.
result High-probability generalisation guarantees with unbounded losses.
Paper introduces a fast, robust, scalable method for detecting changes in data streams.
problem Detecting changes in data streams efficiently and reliably.
method Bayesian online changepoint detection with provable robustness and scalability.
result The proposed method is more than 10 times faster than previous approaches and provides provable robustness.
New bounds for shallow neural networks with deterministic parameters.
problem Developing generalisation bounds for shallow neural networks.
method PAC-Bayesian theory applied to shallow neural networks with deterministic parameters.
result Empirical non-vacuous bounds for shallow neural networks trained with vanilla SGD.
Contrastive unsupervised representation learning (CURL) is the state-of-the-art technique to learn representations (as a set of features) from unlabelled data. While CURL has collected several empirical successes recently, theoretical understanding of its performance was still missing. In a recent work, Arora et al. (2…
New bounds improve generalization in machine learning with high probability.
problem Erratic behavior of KL divergence limits practical applications.
method Replaced KL divergence with Wasserstein distance for better bounds.
result Proved high probability generalization bounds for i.i.d. and non-i.i.d. data.
New bounds for model generalization under deterministic gradient descent.
problem Establishing generalization bounds for models trained with gradient descent methods.
method PAC-Bayesian bounds for deterministic optimisation algorithms.
result Fully computable bounds that depend on initial distribution and Hessian.
ARC algorithm optimizes dynamic pricing with correlated observations.
problem Optimizing dynamic pricing with correlated and generally distributed observations.
method Extends ARC algorithm to batched bandits with generalised linear model.
result ARC algorithm outperforms alternative approaches in dynamic pricing.
Novel strategy for federated learning with privacy-preserving predictors and nonvacuous generalization bounds.
problem Privacy-preserving federated learning with nonvacuous generalization bounds.
method Randomized predictors, PAC-Bayesian generalization bound, synchronous and heterogeneous/homogenous cases.
result Achieves comparable predictive performance to batch approach while preserving privacy.
New bounds link flat minima to good generalisation in overparameterized models.
problem Understanding the relationship between flat minima and generalisation in overparameterized machine learning models.
method Combining PAC-Bayes, Poincaré, and Log-Sobolev inequalities to derive generalisation bounds involving gradient terms.
result Flat minima positively influence generalisation performance, highlighting the benefits of the optimisation phase.
Bayesian framework tackles measurement error in covariates.
problem Misleading inference due to corrupted covariates.
method Bayesian Nonparametric Learning framework robust to misspecification.
result General framework for Classical and Berkson error models.
New PAC-Bayes bounds for unbounded loss functions.
problem Generalization bounds for learning problems with unbounded loss functions.
method Introducing HYPE, a new notion for loss range, and deriving a novel PAC-Bayesian generalization bound.
result PAC-Bayes framework extended to unbounded loss functions.
Robust Kalman filtering method for outlier detection.
problem Outliers and misspecified measurement models in state-space models.
method Combines generalised Bayesian inference with Kalman filters for robustness and efficiency.
result Matches or outperforms other robust filtering methods at lower computational cost.
Mixability of a loss is known to characterise when constant regret bounds are achievable in games of prediction with expert advice through the use of Vovk's aggregating algorithm. We provide a new interpretation of mixability via convex analysis that highlights the role of the Kullback-Leibler divergence in its definit…
PAC-Bayesian bounds improve understanding of K-NN classifier performance.
problem Improving the understanding of K-NN classifier's generalization error.
method PAC-Bayesian analysis applied to K-NN classifier in kernel space.
result PAC-Bayesian bounds provide a function of the number of redundant training examples.
The use of L1 regularisation for sparse learning has generated immense research interest, with successful application in such diverse areas as signal acquisition, image coding, genomics and collaborative filtering. While existing work highlights the many advantages of L1 methods, in this paper we find that L1 regularis…
Bayesian GPR model predicts extreme stock market losses.
problem Forecasting rare but impactful extreme negative returns in equity markets.
method Developed a Bayesian Generalised Pareto Regression model linking scale parameter to market volatility.
result The Cauchy prior provides the best balance between predictive accuracy and model simplicity.
SLT reveals how grokking occurs via basin selection in training.
problem Understanding grokking in machine learning models.
method Singular Learning Theory (SLT) to analyze the loss landscape and local learning coefficient (LLC).
result LLC ranks basins by statistical preference, leading to grokking.
Improved PAC-Bayesian bounds by considering example difficulty.
problem Improving generalization bounds in machine learning.
method Introducing a modified excess risk that leverages example difficulty to reduce variance and tighten PAC-Bayesian bounds.
result Tighter PAC-Bayesian generalization bounds for machine learning models.
Online DEM improves tracking of latent states in dynamic systems.
problem Tracking latent states in dynamic systems with online updates.
method Specializes DEM for online data assimilation, separating temporal scales.
result ODEM can track latent states of a non-linear generative model.
Deep Gaussian processes (DGPs) are multi-layer hierarchical generalisations of Gaussian processes (GPs) and are formally equivalent to neural networks with multiple, infinitely wide hidden layers. DGPs are nonparametric probabilistic models and as such are arguably more flexible, have a greater capacity to generalise, …
Deep Gaussian processes (DGPs) are multi-layer hierarchical generalisations of Gaussian processes (GPs) and are formally equivalent to neural networks with multiple, infinitely wide hidden layers. DGPs are probabilistic and non-parametric and as such are arguably more flexible, have a greater capacity to generalise, an…
The class of chain event graph models is a generalisation of the class of discrete Bayesian networks, retaining most of the structural advantages of the Bayesian network for model interrogation, propagation and learning, while more naturally encoding asymmetric state spaces and the order in which events happen. In this…
We consider a Gaussian process formulation of the multiple kernel learning problem. The goal is to select the convex combination of kernel matrices that best explains the data and by doing so improve the generalisation on unseen data. Sparsity in the kernel weights is obtained by adopting a hierarchical Bayesian approa…
We prove a new minimax theorem connecting the worst-case Bayesian regret and minimax regret under partial monitoring with no assumptions on the space of signals or decisions of the adversary. We then generalise the information-theoretic tools of Russo and Van Roy (2016) for proving Bayesian regret bounds and combine th…
Wide stochastic networks show Gaussian behavior and improve training with PAC-Bayesian methods.
problem Analyzing and training over-parameterised neural networks with large width.
method Establishing Gaussian behavior for a stochastic architecture, applying PAC-Bayesian training.
result PAC-Bayesian training on large but finite-width networks outperforms standard methods.
New algorithm optimizes PAC-Bayes bound without surrogate loss.
problem Mismatch between optimisation objective and generalisation bound in stochastic neural networks.
method Proposes a novel training algorithm that optimizes the PAC-Bayesian bound directly.
result Empirical results show improved performance over existing PAC-Bayesian training methods.
Bayesian neural networks benefit from fully marginalizing over all modes to improve generalization.
problem Bayesian neural networks suffer from multimodal posterior distributions that can lead to suboptimal generalization.
method Use appropriate Bayesian sampling tools to fully marginalize over all posterior modes.
result Training with full marginalization improves the ability of the network to reason between multiple candidate solutions.
In this paper we propose a Bayesian method for estimating architectural parameters of neural networks, namely layer size and network depth. We do this by learning concrete distributions over these parameters. Our results show that regular networks with a learnt structure can generalise better on small datasets, while f…
Advocates for a new posterior that predicts better than classical and generalised Bayes.
problem Combining parameter inference and density estimation for better predictive models.
method Predictively Oriented (PrO) posterior using mean field Langevin dynamics.
result PrO posteriors converge to the predictively optimal model average, adapting to model misspecification.
Bayesian methods suffer from the problem of how to specify prior beliefs. One interesting idea is to consider worst-case priors. This requires solving a stochastic zero-sum game. In this paper, we extend well-known results from bandit theory in order to discover minimax-Bayes policies and discuss when they are practica…
SGD-trained deep nets often generalize well due to a strong inductive bias towards low-error, low-complexity functions.
problem Understanding why overparameterized deep nets generalize well despite fitting training data perfectly.
method Empirical investigation of PSGD(f∣S) and PB(f∣S) for various architectures and datasets. result The probability of SGD-converging on a function consistent with training data correlates well with the Bayesian posterior probability of expressing that function.
Researchers develop methods to calibrate ABMs using Bayesian techniques.
problem Challenges in calibrating ABMs due to likelihood intractability and non-differentiability.
method Generalised variational inference for differentiable ABMs.
result Accurate Bayesian parameter inferences for differentiable ABMs demonstrated.
A new method for incorporating preferences in multi-objective Bayesian optimization.
problem Incorporating preferences in computationally expensive multi-objective optimization problems.
method Building independent surrogate models on each objective function and using Generalised value distribution to approximate the scalarizing function.
result The proposed multi-surrogate approach outperforms the mono-surrogate approach on benchmark and real-world problems.
Efficiently marginalizes over Gaussian Process kernels for better model flexibility and uncertainty.
problem Inefficient marginalization over Gaussian Process kernels for large datasets.
method Bayesian Quadrature scheme with maximum mean discrepancies and invariances between Spectral Mixture kernels.
result Achieves more accurate predictions and better calibrated uncertainty than state-of-the-art baselines.
New algorithm improves Bayesian inference for complex models.
problem Poor performance of existing Bayesian approaches for simulators.
method Posterior bootstrap and maximum mean discrepancy estimators.
result Strong robustness and parallelizability of the new algorithm.
New methods combine model predictions to avoid linear mixtures' limitations.
problem Combining predictions from different models to avoid linear mixtures' limitations.
method Log-linear pooling (locking) and quantum superposition (quacking) to optimise model weights.
result Demonstrated locking method with illustrative example and practical application.
Paper tackles SBI under model misspecification, presenting robust strategies.
problem Challenges in SBI under model misspecification.
method Three key strategies: robust summary statistics, generalised Bayesian inference, and error modelling.
result Empirical results show vulnerabilities of SBI and effectiveness of misspecification-robust alternatives.
Paper introduces robust Gaussian process regression without sacrificing computational efficiency.
problem Violation of independent and identically distributed Gaussian observation noise assumption in Gaussian process regression.
method Proves robust and conjugate Gaussian process regression (RCGP) at no additional cost using generalised Bayesian inference.
result RCGP enables exact conjugate closed form updates in all settings where standard GPs admit them.
New framework improves robust inference in HMMs under model misspecification.
problem Inference in general state-space HMMs under likelihood misspecification.
method Generalized Bayesian Inference (GBI) and Sequential Monte Carlo (SMC) methods.
result Improved performance in object tracking and Gaussian process regression.
We formulate the problem of neural network optimization as Bayesian filtering, where the observations are the backpropagated gradients. While neural network optimization has previously been studied using natural gradient methods which are closely related to Bayesian inference, they were unable to recover standard optim…
Framework for Bayesian inference using GP emulated MH sampler for noisy likelihoods.
problem Approximate Bayesian inference with limited noisy log-likelihood evaluations.
method Gaussian process emulates MH sampler for log-likelihood evaluations; sequential experimental design selects evaluation points.
result Approximate sampler is sample-efficient and robust to GP assumptions.
We introduce the implicit processes (IPs), a stochastic process that places implicitly defined multivariate distributions over any finite collections of random variables. IPs are therefore highly flexible implicit priors over functions, with examples including data simulators, Bayesian neural networks and non-linear tr…
This paper proposes an online tree-based Bayesian approach for reinforcement learning. For inference, we employ a generalised context tree model. This defines a distribution on multivariate Gaussian piecewise-linear models, which can be updated in closed form. The tree structure itself is constructed using the cover tr…
FedGVI improves FL robustness to model misspecification.
problem Limited robustness in FL approaches to model misspecification.
method Probabilistic Federated Learning framework that generalizes previous methods.
result FedGVI provides robust and calibrated predictions under model misspecification.