An increasingly common use case for machine learning models is augmenting the abilities of human decision makers. For classification tasks where neither the human nor model are perfectly accurate, a key step in obtaining high performance is combining their individual predictions in a manner that leverages their relative strengths. In this work, we develop a set of algorithms that combine the probabilistic output of a model with the class-level output of a human. We show theoretically that the accuracy of our combination model is driven not only by the individual human and model accuracies, but also by the model's confidence. Empirical results on image classification with CIFAR-10 and a subset of ImageNet demonstrate that such human-model combinations consistently have higher accuracies than the model or human alone, and that the parameters of the combination method can be estimated effectively with as few as ten labeled datapoints.

Amortized Bayesian inference (ABI) offers fast, scalable approximations to posterior densities by training neural surrogates on data simulated from the statistical model. However, ABI methods are highly sensitive to model misspecification: when observed data fall outside the training distribution (generative scope of the statistical models), neural surrogates can behave unpredictably. This makes it a challenge in a model comparison setting, where multiple statistical models are considered, of which at least some are misspecified. Recent work on self-consistency (SC) provides a promising remedy to this issue, accessible even for empirical data (without ground-truth labels). In this work, we investigate how SC can improve amortized model comparison conceptualized in four different ways. Across two synthetic and two real-world case studies, we find that approaches for model comparison that estimate marginal likelihoods through approximate parameter posteriors consistently outperform methods that directly approximate model evidence or posterior model probabilities. SC training improves robustness when the likelihood is available, even under severe model misspecification. The benefits of SC for methods without access of analytic likelihoods are more limited and inconsistent. Our results suggest practical guidance for reliable amortized Bayesian model comparison: prefer parameter posterior-based methods and augment them with SC training on empirical datasets to mitigate extrapolation bias under model misspecification.

We present a framework using variational inference with normalizing flows (VI-NFs) to generate proposals of reversible jump Markov chain Monte Carlo (RJMCMC) for efficient trans-dimensional Bayesian inference. Unlike transport reversible jump methods relying on forward KL minimization with pilot MCMC samples, our approach minimizes the reverse KL divergence which requires only samples from a base distribution, eliminating costly target sampling. The method employs RealNVP-based flows to learn model-specific transport maps, enabling construction of both between-model and within-model proposals. Our framework provides accurate marginal likelihood estimates from the variational approximation. This facilitates efficient model comparison and proposal adaptation in RJMCMC. Experiments on illustrative example, factor analysis and variable selection tasks in linear regression show that TRJ designed by VI-NFs achieves faster mixing and more efficient model space exploration compared to existing baselines. The proposed algorithm can be extended to conditional flows for amortized vairiational inference across models. Code is available at https://github.com/YinPingping111/TRJ_VINFs.

The expressiveness of flow-based models combined with stochastic variational inference (SVI) has, in recent years, expanded the application of optimization-based Bayesian inference to include problems with complex data relationships. However, until now, SVI using flow-based models has been limited to problems of fixed dimension. We introduce CoSMIC, normalizing flows (COntextually-Specified Masking for Identity-mapped Components), an extension to neural autoregressive conditional normalizing flow architectures that enables using a single amortized variational density for inference over a transdimensional target distribution. We propose a combined stochastic variational transdimensional inference (VTI) approach to training CoSMIC flows using techniques from Bayesian optimization and Monte Carlo gradient estimation. Numerical experiments demonstrate the performance of VTI on challenging problems that scale to high-cardinality model spaces.

An increasingly common use case for machine learning models is augmenting the abilities of human decision makers. For classification tasks where neither the human nor model are perfectly accurate, a key step in obtaining high performance is combining their individual predictions in a manner that leverages their relative strengths. In this work, we develop a set of algorithms that combine the probabilistic output of a model with the class-level output of a human. We show theoretically that the accuracy of our combination model is driven not only by the individual human and model accuracies, but also by the model's confidence. Empirical results on image classification with CIFAR-10 and a subset of ImageNet demonstrate that such human-model combinations consistently have higher accuracies than the model or human alone, and that the parameters of the combination method can be estimated effectively with as few as ten labeled datapoints.

Thunderstorms have significant social and economic impacts due to heavy precipitation, hail, lightning, and strong winds, necessitating reliable forecasts. Thunderstorm forecasts based on numerical weather prediction (NWP) often rely on single-level surrogate predictors, like convective available potential energy and precipitation rate, derived from vertical profiles of three-dimensional atmospheric variables. In this study, we develop SALAMA 1D, a deep neural network that directly infers the probability of thunderstorm occurrence from vertical profiles of ten atmospheric variables, bypassing single-level predictors. By training the model on convection-permitting NWP forecasts, we allow SALAMA 1D to flexibly identify convective patterns, with the goal of enhancing forecast accuracy. The model's architecture is physically motivated: sparse connections encourage interactions at similar height levels, while a shuffling mechanism prevents the model from learning non-physical patterns tied to the vertical grid. SALAMA 1D is trained over Central Europe with lightning observations as the ground truth. Comparative analysis against a baseline machine learning model that uses single-level predictors shows SALAMA 1D's superior skill across various metrics and lead times of up to at least 11 hours. Moreover, increasing the number of forecasts used to compile the training set improves skill, even when training set size is kept constant. Sensitivity analysis using saliency maps indicates that the model reconstructs environmental lapse rates and rediscovers patterns consistent with established theoretical understandings, such as positive buoyancy, convective inhibition, and ice particle formation near the tropopause, while ruling out thunderstorm occurrence based on the absence of mid-level graupel and cloud cover.

Thunderstorms pose a major hazard to society and economy, which calls for reliable thunderstorm forecasts. In this work, we introduce SALAMA, a feedforward neural network model for identifying thunderstorm occurrence in numerical weather prediction (NWP) data. The model is trained on convection-resolving ensemble forecasts over Central Europe and lightning observations. Given only a set of pixel-wise input parameters that are extracted from NWP data and related to thunderstorm development, SALAMA infers the probability of thunderstorm occurrence in a reliably calibrated manner. For lead times up to eleven hours, we find a forecast skill superior to classification based only on NWP reflectivity. Varying the spatiotemporal criteria by which we associate lightning observations with NWP data, we show that the time scale for skillful thunderstorm predictions increases linearly with the spatial scale of the forecast.

Bayesian model comparison (BMC) offers a principled approach for assessing the relative merits of competing computational models and propagating uncertainty into model selection decisions. However, BMC is often intractable for the popular class of hierarchical models due to their high-dimensional nested parameter structure. To address this intractability, we propose a deep learning method for performing BMC on any set of hierarchical models which can be instantiated as probabilistic programs. Since our method enables amortized inference, it allows efficient re-estimation of posterior model probabilities and fast performance validation prior to any real-data application. In a series of extensive validation studies, we benchmark the performance of our method against the state-of-the-art bridge sampling method and demonstrate excellent amortized inference across all BMC settings. We then showcase our method by comparing four hierarchical evidence accumulation models that have previously been deemed intractable for BMC due to partly implicit likelihoods. Additionally, we demonstrate how transfer learning can be leveraged to enhance training efficiency. We provide reproducible code for all analyses and an open-source implementation of our method.

To maintain user trust, large language models (LLMs) should signal low confidence on examples where they are incorrect, instead of misleading the user. The standard approach of estimating confidence is to use the softmax probabilities of these models, but as of November 2023, state-of-the-art LLMs such as GPT-4 and Claude-v1.3 We first study eliciting confidence linguistically -- asking an LLM for its confidence in its answer -- which performs reasonably (80.5% AUC on GPT-4 averaged across 12 question-answering datasets -- 7% above a random baseline) but leaves room for improvement. We then explore using a surrogate confidence model -- using a model where we do have probabilities to evaluate the original model's confidence in a given question. Surprisingly, even though these probabilities come from a different and often weaker model, this method leads to higher AUC than linguistic confidences on 9 out of 12 datasets. Our best method composing linguistic confidences and surrogate model probabilities gives state-of-the-art confidence estimates on all 12 datasets (84.6% average AUC on GPT-4). As large language models (LLMs) are increasingly deployed, it is important that they signal low confidence on examples where they are likely to make mistakes. This paper's goal is to produce good confidence estimates for state-of-the-art LLMs, which do not provide model probabilities or representations (such as GPT-4 and Claude-v1.3). We first examine a natural idea of eliciting linguistic confidence scores (Tian et al., 2023; Lin et al., 2022; Xiong et al., 2023) -- prompting the LLM to assess its confidence in its answer (Figure 1, GPT-4 Linguistic). We find that linguistic confidences work reasonably well for state-of-the-art models, and much better than a random guessing baseline, but still leave room for improvement (Section 3). Averaged across the datasets, GPT-4 achieves a selective classification AUC of 80.5%, which is 7% above a random guessing baseline. Our results hold across 12 standard datasets (8 MMLU datasets, TruthfulQA, CommonsenseQA, OpenbookQA, and MedQA), 5 models (GPT-4, Claude-v1.3,

Minimum Bayes Risk (MBR) decoding has been shown to be a powerful alternative to beam search decoding in a variety of text generation tasks. MBR decoding selects a hypothesis from a pool of hypotheses that has the least expected risk under a probability model according to a given utility function. Since it is impractical to compute the expected risk exactly over all possible hypotheses, two approximations are commonly used in MBR. First, it integrates over a sampled set of hypotheses rather than over all possible hypotheses. Second, it estimates the probability of each hypothesis using a Monte Carlo estimator. While the first approximation is necessary to make it computationally feasible, the second is not essential since we typically have access to the model probability at inference time. We propose Model-Based MBR (MBMBR), a variant of MBR that uses the model probability itself as the estimate of the probability distribution instead of the Monte Carlo estimate. We show analytically and empirically that the model-based estimate is more promising than the Monte Carlo estimate in text generation tasks. Our experiments show that MBMBR outperforms MBR in several text generation tasks, both with encoder-decoder models and with large language models.