A new method selects optimal temperature for Bayesian Deep Learning.
problem Finding the optimal temperature for improving predictive performance in Bayesian Deep Learning.
method Data-driven approach to estimate temperature as a model parameter.
result Our method performs comparably to grid search but at a fraction of the cost.
Linear VAEs explain posterior collapse in VAEs via local maxima in log marginal likelihood.
problem Posterior collapse in VAEs where variational posterior matches prior for some latent variables.
method Analysis of linear VAEs and their relation to pPCA, proving ELBO does not introduce spurious local maxima.
result Linear VAEs have identifiable global maxima corresponding to principal component directions, explaining posterior collapse.
Transformers approximate Bayesian posteriors but not exactly.
problem Bayesian accounts of in-context learning face challenges due to task-preserving order changes in transformers.
method Showed that excess prequential code length is exactly cumulative predictive KL, decomposing expected regret into order-averaged predictor and order-averaging gain.
result Transformers approximate Bayesian posteriors but not exactly, priced by log loss.
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.
Bayesian models use marginal likelihood; non-Bayesian use cross-validation, shown equivalent.
problem Comparing Bayesian and non-Bayesian models for evaluation.
method Showed marginal likelihood is equivalent to leave-p-out cross-validation, with log posterior predictive as scoring rule.
result Marginal likelihood and cross-validation are formally equivalent under data exchangeability.
New method improves posterior sampling for complex data models.
problem Sampling from posterior distributions in high-dimensional data.
method Tilted transport technique combining denoising oracle and log-likelihood.
result Boosted posterior is strongly log-concave, facilitating easier sampling.
We compare various extensions of the Bradley-Terry model and a hierarchical Poisson log-linear model in terms of their performance in predicting the outcome of soccer matches (win, draw, or loss). The parameters of the Bradley-Terry extensions are estimated by maximizing the log-likelihood, or an appropriately penalize…
Sparse matrices simplify computation of GP variances and likelihoods.
problem Efficient computation of posterior variance and log-likelihood for additive Matérn GPs.
method Represented posterior mean, variance, log-likelihood, and gradient using sparse matrices.
result Efficient computation of posterior mean, variance, log-likelihood, and gradient in O(nlogn) time. Improved predictive uncertainties in Gaussian Process regression.
problem Substantially underestimated uncertainties in GP predictive distributions.
method Two methods for scalable GP regression: variational inference for FITC and direct posterior predictive distribution.
result Significantly better calibrated uncertainties and higher log likelihoods.
The paper integrates multiple Gaussian process predictions using Monte Carlo sampling.
problem Accurate prediction of variables using multiple models.
method Log-linear pooling of Gaussian process predictions, combined with Monte Carlo sampling.
result The log-linear pooling method improves prediction accuracy compared to linear pooling.
QLA improves Bayesian uncertainty estimation for DNNs without increasing computational cost.
problem Overconfident out-of-distribution predictions from DNNs.
method Proposes Quadratic Laplace Approximation (QLA) to improve Bayesian uncertainty quantification.
result QLA yields modest yet consistent uncertainty estimation improvements over Linearized Laplace Approximation (LLA) on five regression datasets.
A method to approximate posterior distributions using Monte Carlo and variational inference.
problem Lack of systematic understanding of how optimizing different objectives relates to approximating the posterior distribution.
method Divide and couple procedure to identify augmented proposal and target distributions.
result Maximizing the VI objective leads to an augmented variational distribution that approximates the posterior distribution.
PFN-TS uses Thompson sampling with PFNs to improve contextual bandit performance.
problem Improving contextual bandit performance using Thompson sampling with prior-data fitted networks.
method PFN-TS converts PFN posterior predictives into mean-reward samples using a subsampled predictive central limit theorem.
result PFN-TS achieves the best average rank across nonlinear synthetic and OpenML classification-to-bandit benchmarks.
New algorithm samples neural network posteriors efficiently.
problem Challenges of sampling multimodal Bayesian posteriors for neural networks.
method Greedy Bayes method using log-concave coupling of posterior and auxiliary random variable.
result Log-concave coupling facilitates efficient sampling of neuron weights.
Boosted decision trees typically yield good accuracy, precision, and ROC area. However, because the outputs from boosting are not well calibrated posterior probabilities, boosting yields poor squared error and cross-entropy. We empirically demonstrate why AdaBoost predicts distorted probabilities and examine three cali…
This paper treats prediction markets as Bayesian inverse problems to quantify uncertainty and identify event outcomes.
problem Uncertainty and identifiability in prediction market outcomes from price-volume histories.
method Formulates prediction markets as Bayesian inverse problems, introduces a log-odds observation model, and derives posterior uncertainty quantification and identifiability criteria.
result Explicit diagnostics for informative and stable inference regimes, and validation through synthetic data experiments.
A new method for Bayesian neural networks improves prediction uncertainty.
problem Estimating uncertainty in neural network predictions.
method Adversarial α-divergence minimization for approximate Bayesian inference.
result The method often gives better performance in terms of test log-likelihood and squared error in regression problems.
New method improves sampling for weakly log-concave posteriors.
problem Sampling from weakly log-concave posterior distributions.
method Stochastic Langevin Monte Carlo with over-damped diffusion.
result Simulation horizon is (dlog(n)2)(1+r)2 with Poisson subsampling. We consider the problem of transforming samples from one continuous source distribution into samples from another target distribution. We demonstrate with optimal transport theory that when the source distribution can be easily sampled from and the target distribution is log-concave, this can be tractably solved with c…
Study compares uncertainty estimation methods for Bayesian Neural Networks.
problem Quality of uncertainty quantification in Bayesian Neural Networks.
method Empirical comparison of 10 inference methods on regression and classification tasks.
result Common inference metrics can be misleading, and methods designed to capture posterior structure do not always produce high-quality approximations.
Generative AI can solve in-context learning problems using a martingale perspective.
problem Estimating when a conditional generative model can solve an in-context learning problem.
method Bayesian interpretation, ancestral sampling, generative predictive p-value.
result Developed a method to assess the suitability of CGMs for ICL problems using generative predictive p-values.
Bayesian neural networks approximate Gaussian, this method adapts to non-Gaussian posteriors.
problem Bayesian neural networks struggle with non-Gaussian posteriors, leading to poor performance.
method Proposes a Riemannian Laplace approximation to adapt to the shape of the true posterior.
result Consistently improves over conventional Laplace approximation across tasks.
LaPSRL achieves optimal regret for isoperimetric RL distributions.
problem Designing RL algorithms with sublinear regret for non-log-concave distributions.
method Posterior Sampling (PSRL) and Langevin sampling (LaPSRL) for isoperimetric distributions.
result LaPSRL achieves order-optimal regret and subquadratic complexity.
New method speeds up Bayesian inverse problem solving with neural operators.
problem Solving infinite-dimensional Bayesian inverse problems with high computational cost.
method Delayed-acceptance geometric MCMC driven by derivative-informed neural operator surrogates.
result Significant speedup in generating posterior samples (3-9 times faster).
Markov chain Monte Carlo (MCMC) algorithms have become powerful tools for Bayesian inference. However, they do not scale well to large-data problems. Divide-and-conquer strategies, which split the data into batches and, for each batch, run independent MCMC algorithms targeting the corresponding subposterior, can spread…
A simple pooling technique prevents posterior collapse in sequence VAEs.
problem Posterior collapse in sequence VAEs, causing models to ignore latent variables.
method Proposed a pooling technique to prevent posterior collapse.
result Significantly better data log-likelihood achieved compared to standard sequence VAEs.
Unified framework for causal inference with reliable uncertainty quantification.
problem Causal inference under unobserved confounding with unreliable uncertainty quantification.
method Deconditional Gaussian Process (DGP) framework for uncertainty-aware causal learning.
result Strong predictive performance and informative uncertainty quantification.
Normalizing flow regression approximates posterior distributions without additional sampling.
problem Bayesian inference with computationally expensive likelihood evaluations.
method Normalizing flow regression (NFR) for offline inference.
result NFR yields a tractable posterior approximation through regression on existing log-density evaluations.
Coherent uncertainty quantification is a key strength of Bayesian methods. But modern algorithms for approximate Bayesian posterior inference often sacrifice accurate posterior uncertainty estimation in the pursuit of scalability. This work shows that previous Bayesian coreset construction algorithms---which build a sm…
The variational autoencoder (VAE; Kingma, Welling (2014)) is a recently proposed generative model pairing a top-down generative network with a bottom-up recognition network which approximates posterior inference. It typically makes strong assumptions about posterior inference, for instance that the posterior distributi…
Variational dropout (VD) is a generalization of Gaussian dropout, which aims at inferring the posterior of network weights based on a log-uniform prior on them to learn these weights as well as dropout rate simultaneously. The log-uniform prior not only interprets the regularization capacity of Gaussian dropout in netw…
This work improves bounds on Bayesian coreset quality.
problem Limitations of existing theoretical analysis of Bayesian coresets.
method Develops general upper and lower bounds on KL divergence.
result Demonstrates flexibility of new theoretical bounds in various models.
The paper improves sparse Gaussian processes by optimizing predictive loss.
problem Optimizing predictive loss in sparse Gaussian processes.
method Direct loss minimization (DLM) for log-loss and square loss, with product sampling (uPS) and biased Monte Carlo (bMC) for non-conjugate cases.
result DLM shows significant performance improvement in both log-loss and square loss cases.
Bayesian models predict Collatz stopping times with high accuracy.
problem Predicting the total stopping time of Collatz sequences.
method Developed two complementary models: a hierarchical Negative Binomial regression and a mechanistic generative approximation.
result Bayesian models outperform generative approximations in predicting Collatz stopping times.
Paper estimates AI hallucinations in conditional generation tasks.
problem Estimating the frequency of AI-generated incorrect responses.
method Developed a method to estimate hallucination probability from generated responses and log probabilities.
result Method accurately estimates hallucination rate in natural language and synthetic tasks.
We introduce thermodynamic response functions for singular Bayesian models.
problem Singular Bayesian models violate regular asymptotics due to non-identifiability and degenerate Fisher geometry.
method Posterior tempering induces thermodynamic response functions, linking WAIC, WBIC, and singular fluctuation.
result WAIC, WBIC, and singular fluctuation are unified within a thermodynamic response framework.
In this paper, we propose a Double Thompson Sampling (D-TS) algorithm for dueling bandit problems. As indicated by its name, D-TS selects both the first and the second candidates according to Thompson Sampling. Specifically, D-TS maintains a posterior distribution for the preference matrix, and chooses the pair of arms…
PVI seeks a posterior that makes predictions closer to true data, not approximating the Bayesian posterior.
problem Finding meaningful posterior distributions under model misspecification.
method Predictive variational inference (PVI) seeks an optimal posterior density for close predictive matching to true data.
result PVI learns a posterior that is not the same as the Bayesian posterior, but is closer to the true data generating process.
Misspecification-Aware Simulation-Based Inference via Side-Channel Guidance
problem Simulation-based inference (SBI) of latent parameters is hindered by simulator misspecification.
method Misspecification-Aware Simulation-Based Inference (MA-SBI) turns side-channel text into a posterior correction.
result MA-SBI matches the oracle posterior across 10 seeds and two backbones.
Enhances SGLD for log-concave posteriors with asynchronous computation.
problem Sampling log-concave posterior distributions efficiently.
method Integrates asynchronous computation into SGLD with delayed gradients.
result Convergence in measure is not significantly affected by delayed gradient information.
It is well known that Markov chain Monte Carlo (MCMC) methods scale poorly with dataset size. A popular class of methods for solving this issue is stochastic gradient MCMC. These methods use a noisy estimate of the gradient of the log posterior, which reduces the per iteration computational cost of the algorithm. Despi…
Bayesian models' singular fluctuation is shown to be akin to specific heat, influencing model complexity and generalization.
problem Understanding the thermodynamic interpretation of singular fluctuation in Bayesian models.
method Showed singular fluctuation as the curvature of Bayesian free energy and variance of log-likelihood observable under a Gibbs posterior.
result Singular fluctuation is the statistical analogue of specific heat, controlling model complexity and generalization.
In this paper we study the probabilistic properties of the posteriors in a speech recognition system that uses a deep neural network (DNN) for acoustic modeling. We do this by reducing Kaldi's DNN shared pdf-id posteriors to phone likelihoods, and using test set forced alignments to evaluate these using a calibration s…
MFVI can overestimate predictive variance compared to the exact posterior
problem MFVI underestimates posterior variance
method Analyzing conjugate Bayesian Linear Regression
result MFVI can overestimate predictive variance compared to the exact posterior
Improved model-based estimation through tempered Bayes filter.
problem Improving predictive accuracy in partially-observable stochastic systems.
method Developed tempered Bayes filter combining likelihood and full posterior tempering.
result Tempered Bayes filter achieves improved predictive performance over the Bayes filter baseline.
Test log-likelihood comparisons can be misleading.
problem Misinterpretation of test log-likelihood in model comparison.
method Simple examples of model comparison and forecast accuracy.
result Test log-likelihood does not always correlate with model accuracy.
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.
Generative model improves sample quality on ImageNet32.
problem Improving sample quality in generative models.
method Iterative Gaussian posterior inference, treating generated sample as unknown variable.
result Improves sample quality on ImageNet32 over BFNs and VDMs.