When finetuning multiple tasks altogether, it is important to carefully weigh them to get a good performance, but searching for good weights can be difficult and costly. Here, we propose to aid the search with fast previews to quickly get a rough idea of different reweighting options. We use model merging to create previews by simply reusing and averaging parameters of models trained on each task separately (no retraining required). To improve the quality of previews, we propose a Bayesian approach to design new merging strategies by using more flexible posteriors. We validate our findings on vision and natural-language transformers. Our work shows the benefits of model merging via Bayes to improve multitask finetuning.

Sampling from high dimensional, multimodal distributions remains a fundamental challenge across domains such as statistical Bayesian inference and physics based machine learning. In this paper, we propose Annealing Flow, a continuous normalizing flow based approach designed to sample from high dimensional and multimodal distributions. The key idea is to learn a continuous normalizing flow based transport map, guided by annealing, to transition samples from an easy to sample distribution to the target distribution, facilitating effective exploration of modes in high dimensional spaces. Unlike many existing methods, AF training does not rely on samples from the target distribution. AF ensures effective and balanced mode exploration, achieves linear complexity in sample size and dimensions, and circumvents inefficient mixing times. We demonstrate the superior performance of AF compared to state of the art methods through extensive experiments on various challenging distributions and real world datasets, particularly in high-dimensional and multimodal settings. We also highlight the potential of AF for sampling the least favorable distributions.

Estimating uncertainty in Large Language Models (LLMs) is important for properly evaluating LLMs, and ensuring safety for users. However, prior approaches to uncertainty estimation focus on the final answer in generated text, ignoring intermediate steps that might dramatically impact the outcome. We hypothesize that there exist key forking tokens, such that re-sampling the system at those specific tokens, but not others, leads to very different outcomes. To test this empirically, we develop a novel approach to representing uncertainty dynamics across individual tokens of text generation, and applying statistical models to test our hypothesis. Our approach is highly flexible: it can be applied to any dataset and any LLM, without fine tuning or accessing model weights. We use our method to analyze LLM responses on 7 different tasks across 4 domains, spanning a wide range of typical use cases. We find many examples of forking tokens, including surprising ones such as punctuation marks, suggesting that LLMs are often just a single token away from saying something very different.

Label distribution learning (LDL) provides a framework wherein a distribution over categories rather than a single category is predicted, with the aim of addressing ambiguity in labeled data. Existing research on LDL mainly focuses on the task of point estimation, i.e., pinpointing an optimal distribution in the probability simplex conditioned on the input sample. In this paper, we estimate a probability distribution of all possible label distributions over the simplex, by unleashing the expressive power of the recently introduced Squared Neural Family (SNEFY). With the modeled distribution, label distribution prediction can be achieved by performing the expectation operation to estimate the mean of the distribution of label distributions. Moreover, more information about the label distribution can be inferred, such as the prediction reliability and uncertainties. We conduct extensive experiments on the label distribution prediction task, showing that our distribution modeling based method can achieve very competitive label distribution prediction performance compared with the state-of-the-art baselines. Additional experiments on active learning and ensemble learning demonstrate that our probabilistic approach can effectively boost the performance in these settings, by accurately estimating the prediction reliability and uncertainties.

We propose to quantify dependence between two systems $X$ and $Y$ in a dataset $D$ based on the Bayesian comparison of two models: one, $H_0$, of statistical independence and another one, $H_1$, of dependence. In this framework, dependence between $X$ and $Y$ in $D$, denoted $B(X,Y|D)$, is quantified as $P(H_1|D)$, the posterior probability for the model of dependence given $D$, or any strictly increasing function thereof. It is therefore a measure of the evidence for dependence between $X$ and $Y$ as modeled by $H_1$ and observed in $D$. We review several statistical models and reconsider standard results in the light of $B(X,Y|D)$ as a measure of dependence. Using simulations, we focus on two specific issues: the effect of noise and the behavior of $B(X,Y|D)$ when $H_1$ has a parameter coding for the intensity of dependence. We then derive some general properties of $B(X,Y|D)$, showing that it quantifies the information contained in $D$ in favor of $H_1$ versus $H_0$. While some of these properties are typical of what is expected from a valid measure of dependence, others are novel and naturally appear as desired features for specific measures of dependence, which we call inferential. We finally put these results in perspective; in particular, we discuss the consequences of using the Bayesian framework as well as the similarities and differences between $B(X,Y|D)$ and mutual information.

Machine learning models have exhibited exceptional results in various domains. The most prevalent approach for learning is the empirical risk minimizer (ERM), which adapts the model's weights to reduce the loss on a training set and subsequently leverages these weights to predict the label for new test data. Nonetheless, ERM makes the assumption that the test distribution is similar to the training distribution, which may not always hold in real-world situations. In contrast, the predictive normalized maximum likelihood (pNML) was proposed as a min-max solution for the individual setting where no assumptions are made on the distribution of the tested input. This study investigates pNML's learnability for linear regression and neural networks, and demonstrates that pNML can improve the performance and robustness of these models on various tasks. Moreover, the pNML provides an accurate confidence measure for its output, showcasing state-of-the-art results for out-of-distribution detection, resistance to adversarial attacks, and active learning.

Pedestrian detection is a critical task in autonomous driving, aimed at enhancing safety and reducing risks on the road. Over recent years, significant advancements have been made in improving detection performance. However, these achievements still fall short of human perception, particularly in cases involving occluded pedestrians, especially entirely invisible ones. In this work, we present the Occlusion-Rich Road Scenes with Pedestrians (OccluRoads) dataset, which features a diverse collection of road scenes with partially and fully occluded pedestrians in both real and virtual environments. All scenes are meticulously labeled and enriched with contextual information that encapsulates human perception in such scenarios. Using this dataset, we developed a pipeline to predict the presence of occluded pedestrians, leveraging Knowledge Graph (KG), Knowledge Graph Embedding (KGE), and a Bayesian inference process. Our approach achieves a F1 score of 0.91, representing an improvement of up to 42% compared to traditional machine learning models.

Information Retrieval (IR) systems used in search and recommendation platforms frequently employ Learning-to-Rank (LTR) models to rank items in response to user queries. These models heavily rely on features derived from user interactions, such as clicks and engagement data. This dependence introduces cold start issues for items lacking user engagement and poses challenges in adapting to non-stationary shifts in user behavior over time. We address both challenges holistically as an online learning problem and propose BayesCNS, a Bayesian approach designed to handle cold start and non-stationary distribution shifts in search systems at scale. BayesCNS achieves this by estimating prior distributions for user-item interactions, which are continuously updated with new user interactions gathered online. This online learning procedure is guided by a ranker model, enabling efficient exploration of relevant items using contextual information provided by the ranker. We successfully deployed BayesCNS in a large-scale search system and demonstrated its efficacy through comprehensive offline and online experiments. Notably, an online A/B experiment showed a 10.60% increase in new item interactions and a 1.05% improvement in overall success metrics over the existing production baseline.

We study continuous-time mean--variance portfolio selection in markets where stock prices are diffusion processes driven by observable factors that are also diffusion processes yet the coefficients of these processes are unknown. Based on the recently developed reinforcement learning (RL) theory for diffusion processes, we present a general data-driven RL algorithm that learns the pre-committed investment strategy directly without attempting to learn or estimate the market coefficients. For multi-stock Black--Scholes markets without factors, we further devise a baseline algorithm and prove its performance guarantee by deriving a sublinear regret bound in terms of Sharpe ratio. For performance enhancement and practical implementation, we modify the baseline algorithm into four variants, and carry out an extensive empirical study to compare their performance, in terms of a host of common metrics, with a large number of widely used portfolio allocation strategies on S\&P 500 constituents. The results demonstrate that the continuous-time RL strategies are consistently among the best especially in a volatile bear market, and decisively outperform the model-based continuous-time counterparts by significant margins.

This work is about estimating when a conditional generative model (CGM) can solve an in-context learning (ICL) problem. An in-context learning (ICL) problem comprises a CGM, a dataset, and a prediction task. The CGM could be a multimodal foundation model; the dataset, a collection of patient histories, test results, and recorded diagnoses; and the prediction task to communicate a diagnosis to a new patient. A Bayesian interpretation of ICL assumes that the CGM computes a posterior predictive distribution over an unknown Bayesian model defining a joint distribution over latent explanations and observable data. From this perspective, Bayesian model criticism is a reasonable approach to assess the suitability of a given CGM for an ICL problem. However, such approaches--like posterior predictive checks (PPCs)--often assume that we can sample from the likelihood and posterior defined by the Bayesian model, which are not explicitly given for contemporary CGMs. To address this, we show when ancestral sampling from the predictive distribution of a CGM is equivalent to sampling datasets from the posterior predictive of the assumed Bayesian model. Then we develop the generative predictive p-value, which enables PPCs and their cousins for contemporary CGMs. The generative predictive p-value can then be used in a statistical decision procedure to determine when the model is appropriate for an ICL problem. Our method only requires generating queries and responses from a CGM and evaluating its response log probability. We empirically evaluate our method on synthetic tabular, imaging, and natural language ICL tasks using large language models. An in-context learning (ICL) problem comprises a conditional generative model (CGM), a dataset, and a prediction task (Brown et al., 2020; Dong et al., 2022).