Most leading implementations of black-box variational inference (BBVI) are based on optimizing a stochastic evidence lower bound (ELBO). But such approaches to BBVI often converge slowly due to the high variance of their gradient estimates. In this work, we propose batch and match (BaM), an alternative approach to BBVI based on a score-based divergence. Notably, this score-based divergence can be optimized by a closed-form proximal update for Gaussian variational families with full covariance matrices. We analyze the convergence of BaM when the target distribution is Gaussian, and we prove that in the limit of infinite batch size the variational parameter updates converge exponentially quickly to the target mean and covariance. We also evaluate the performance of BaM on Gaussian and non-Gaussian target distributions that arise from posterior inference in hierarchical and deep generative models. In these experiments, we find that BaM typically converges in fewer (and sometimes significantly fewer) gradient evaluations than leading implementations of BBVI based on ELBO maximization.
Scientists often want to learn about cause and effect from hierarchical data, collected from subunits nested inside units. Consider students in schools, cells in patients, or cities in states. In such settings, unit-level variables (e.g. each school's budget) may affect subunit-level variables (e.g. the test scores of each student in each school) and vice versa. To address causal questions with hierarchical data, we propose hierarchical causal models, which extend structural causal models and causal graphical models by adding inner plates. We develop a general graphical identification technique for hierarchical causal models that extends do-calculus. We find many situations in which hierarchical data can enable causal identification even when it would be impossible with non-hierarchical data, that is, if we had only unit-level summaries of subunit-level variables (e.g. the school's average test score, rather than each student's score). We develop estimation techniques for hierarchical causal models, using methods including hierarchical Bayesian models. We illustrate our results in simulation and via a reanalysis of the classic "eight schools" study.
A recent line of work in natural language processing has aimed to combine language models and topic models. These topic-guided language models augment neural language models with topic models, unsupervised learning methods that can discover document-level patterns of word use. This paper compares the effectiveness of these methods in a standardized setting. We study four topic-guided language models and two baselines, evaluating the held-out predictive performance of each model on four corpora. Surprisingly, we find that none of these methods outperform a standard LSTM language model baseline, and most fail to learn good topics. Further, we train a probe of the neural language model that shows that the baseline's hidden states already encode topic information. We make public all code used for this study.
We develop a quantitative method to assess the style of American poems and to visualize a collection of poems in relation to one another. Qualitative poetry criticism helped guide our development of metrics that analyze various orthographic, syntactic, and phonemic features. These features are used to discover comprehensive stylistic information from a poem's multi-layered latent structure, and to compute distances between poems in this space. Visualizations provide ready access to the analytical components. We demonstrate our method on several collections of poetry, showing that it better delineates poetry style than the traditional word-occurrence features that are used in typical text analysis algorithms. Our method has potential applications to academic research of texts, to research of the intuitive personal response to poetry, and to making recommendations to readers based on their favorite poems.
This paper presents a case study on the design, administration, post-processing, and evaluation of surveys on large language models (LLMs). It comprises two components: (1) A statistical method for eliciting beliefs encoded in LLMs. We introduce statistical measures and evaluation metrics that quantify the probability of an LLM "making a choice", the associated uncertainty, and the consistency of that choice. (2) We apply this method to study what moral beliefs are encoded in different LLMs, especially in ambiguous cases where the right choice is not obvious. We design a large-scale survey comprising 680 high-ambiguity moral scenarios (e.g., "Should I tell a white lie?") and 687 low-ambiguity moral scenarios (e.g., "Should I stop for a pedestrian on the road?"). Each scenario includes a description, two possible actions, and auxiliary labels indicating violated rules (e.g., "do not kill"). We administer the survey to 28 open- and closed-source LLMs. We find that (a) in unambiguous scenarios, most models "choose" actions that align with commonsense. In ambiguous cases, most models express uncertainty. (b) Some models are uncertain about choosing the commonsense action because their responses are sensitive to the question-wording. (c) Some models reflect clear preferences in ambiguous scenarios. Specifically, closed-source models tend to agree with each other.
Amortized variational inference (A-VI) is a method for approximating the intractable posterior distributions that arise in probabilistic models. The defining feature of A-VI is that it learns a global inference function that maps each observation to its local latent variable's approximate posterior. This stands in contrast to the more classical factorized (or mean-field) variational inference (F-VI), which directly learns the parameters of the approximating distribution for each latent variable. In deep generative models, A-VI is used as a computational trick to speed up inference for local latent variables. In this paper, we study A-VI as a general alternative to F-VI for approximate posterior inference. A-VI cannot produce an approximation with a lower Kullback-Leibler divergence than F-VI's optimal solution, because the amortized family is a subset of the factorized family. Thus a central theoretical problem is to characterize when A-VI still attains F-VI's optimal solution. We derive conditions on both the model and the inference function under which A-VI can theoretically achieve F-VI's optimum. We show that for a broad class of hierarchical models, including deep generative models, it is possible to close the gap between A-VI and F-VI. Further, for an even broader class of models, we establish when and how to expand the domain of the inference function to make amortization a feasible strategy. Finally, we prove that for certain models -- including hidden Markov models and Gaussian processes -- A-VI cannot match F-VI's solution, no matter how expressive the inference function is. We also study A-VI empirically. On several examples, we corroborate our theoretical results and investigate the performance of A-VI when varying the complexity of the inference function. When the gap between A-VI and F-VI can be closed, we find that the required complexity of the function need not scale with the number of observations, and that A-VI often converges faster than F-VI.
Diffusion models have been successful on a range of conditional generation tasks including molecular design and text-to-image generation. However, these achievements have primarily depended on task-specific conditional training or error-prone heuristic approximations. Ideally, a conditional generation method should provide exact samples for a broad range of conditional distributions without requiring task-specific training. To this end, we introduce the Twisted Diffusion Sampler, or TDS. TDS is a sequential Monte Carlo (SMC) algorithm that targets the conditional distributions of diffusion models. The main idea is to use twisting, an SMC technique that enjoys good computational efficiency, to incorporate heuristic approximations without compromising asymptotic exactness. We first find in simulation and on MNIST image inpainting and class-conditional generation tasks that TDS provides a computational statistical trade-off, yielding more accurate approximations with many particles but with empirical improvements over heuristics with as few as two particles. We then turn to motif-scaffolding, a core task in protein design, using a TDS extension to Riemannian diffusion models. On benchmark test cases, TDS allows flexible conditioning criteria and often outperforms the state of the art.
Assessing the predictive uncertainty of deep neural networks is crucial for safety-related applications of deep learning. Although Bayesian deep learning offers a principled framework for estimating model uncertainty, the approaches that are commonly used to approximate the posterior often fail to deliver reliable estimates of predictive uncertainty. In this paper we propose a novel criterion for predictive uncertainty, that a model's predictive variance should be grounded in the empirical density of the input. It should produce higher uncertainty for inputs that are improbable in the training data and lower uncertainty for those inputs that are more probable. To operationalize this criterion, we develop the density uncertainty layer, an architectural element for a stochastic neural network that guarantees that the density uncertain criterion is satisfied. We study neural networks with density uncertainty layers on the CIFAR-10 and CIFAR-100 uncertainty benchmarks. Compared to existing approaches, we find that density uncertainty layers provide reliable uncertainty estimates and robust out-of-distribution detection performance.
We study causal representation learning, the task of inferring latent causal variables and their causal relations from high-dimensional functions ("mixtures") of the variables. Prior work relies on weak supervision, in the form of counterfactual pre- and post-intervention views or temporal structure; places restrictive assumptions, such as linearity, on the mixing function or latent causal model; or requires partial knowledge of the generative process, such as the causal graph or the intervention targets. We instead consider the general setting in which both the causal model and the mixing function are nonparametric. The learning signal takes the form of multiple datasets, or environments, arising from unknown interventions in the underlying causal model. Our goal is to identify both the ground truth latents and their causal graph up to a set of ambiguities which we show to be irresolvable from interventional data. We study the fundamental setting of two causal variables and prove that the observational distribution and one perfect intervention per node suffice for identifiability, subject to a genericity condition. This condition rules out spurious solutions that involve fine-tuning of the intervened and observational distributions, mirroring similar conditions for nonlinear cause-effect inference. For an arbitrary number of variables, we show that two distinct paired perfect interventions per node guarantee identifiability. Further, we demonstrate that the strengths of causal influences among the latent variables are preserved by all equivalent solutions, rendering the inferred representation appropriate for drawing causal conclusions from new data. Our study provides the first identifiability results for the general nonparametric setting with unknown interventions, and elucidates what is possible and impossible for causal representation learning without more direct supervision.
Controlled generation refers to the problem of creating text that contains stylistic or semantic attributes of interest. Many approaches reduce this problem to training a predictor of the desired attribute. For example, researchers hoping to deploy a large language model to produce non-toxic content may use a toxicity classifier to filter generated text. In practice, the generated text to classify, which is determined by user prompts, may come from a wide range of distributions. In this paper, we show that the performance of controlled generation may be poor if the distributions of text in response to user prompts differ from the distribution the predictor was trained on. To address this problem, we cast controlled generation under distribution shift as an invariant learning problem: the most effective predictor should be invariant across multiple text environments. We then discuss a natural solution that arises from this characterization and propose heuristics for selecting natural environments. We study this characterization and the proposed method empirically using both synthetic and real data. Experiments demonstrate both the challenge of distribution shift in controlled generation and the potential of invariance methods in this setting.