Bayesian neural networks (BNNs) have recently gained popularity due to their ability to quantify model uncertainty. However, specifying a prior for BNNs that captures relevant domain knowledge is often extremely challenging. In this work, we propose a framework for integrating general forms of domain knowledge (i.e., any knowledge that can be represented by a loss function) into a BNN prior through variational inference, while enabling computationally efficient posterior inference and sampling. Specifically, our approach results in a prior over neural network weights that assigns high probability mass to models that better align with our domain knowledge, leading to posterior samples that also exhibit this behavior. We show that BNNs using our proposed domain knowledge priors outperform those with standard priors (e.g., isotropic Gaussian, Gaussian process), successfully incorporating diverse types of prior information such as fairness, physics rules, and healthcare knowledge and achieving better predictive performance. We also present techniques for transferring the learned priors across different model architectures, demonstrating their broad utility across various settings.

Covariate adjustment is a ubiquitous method used to estimate the average treatment effect (ATE) from observational data. Assuming a known graphical structure of the data generating model, recent results give graphical criteria for optimal adjustment, which enables efficient estimation of the ATE. However, graphical approaches are challenging for high-dimensional and complex data, and it is not straightforward to specify a meaningful graphical model of non-Euclidean data such as texts. We propose an general framework that accommodates adjustment for any subset of information expressed by the covariates. We generalize prior works and leverage these results to identify the optimal covariate information for efficient adjustment. This information is minimally sufficient for prediction of the outcome conditionally on treatment. Based on our theoretical results, we propose the Debiased Outcome-adapted Propensity Estimator (DOPE) for efficient estimation of the ATE, and we provide asymptotic results for the DOPE under general conditions. Compared to the augmented inverse propensity weighted (AIPW) estimator, the DOPE can retain its efficiency even when the covariates are highly predictive of treatment. We illustrate this with a single-index model, and with an implementation of the DOPE based on neural networks, we demonstrate its performance on simulated and real data. Our results show that the DOPE provides an efficient and robust methodology for ATE estimation in various observational settings.

The rise of the Internet of Things (IoT) and mobile internet applications has spurred interest in location-based services (LBS) for commercial, military, and social applications. While the global positioning system (GPS) dominates outdoor localization, its efficacy wanes indoors due to signal challenges. Indoor localization systems leverage wireless technologies like Wi-Fi, ZigBee, Bluetooth, UWB, selecting based on context. Received signal strength indicator (RSSI) technology, known for its accuracy and simplicity, is widely adopted. This study employs machine learning algorithms in three phases: supervised regressors, supervised classifiers, and ensemble methods for RSSI-based indoor localization. Additionally, it introduces a weighted least squares technique and pseudo-linear solution approach to address non-linear RSSI measurement equations by approximating them with linear equations. An experimental testbed, utilizing diverse wireless technologies and anchor nodes, is designed for data collection, employing IoT cloud architectures. Pre-processing involves investigating filters for data refinement before algorithm training. The study employs machine learning models like linear regression, polynomial regression, support vector regression, random forest regression, and decision tree regressor across various wireless technologies. These models estimate the geographical coordinates of a moving target node, and their performance is evaluated using metrics such as accuracy, root mean square errors, precision, recall, sensitivity, coefficient of determinant, and the f1-score. The experiment's outcomes provide insights into the effectiveness of different supervised machine learning techniques in terms of localization accuracy and robustness in indoor environments.

Neuro-Symbolic (NeSy) predictors that conform to symbolic knowledge - encoding, e.g., safety constraints - can be affected by Reasoning Shortcuts (RSs): They learn concepts consistent with the symbolic knowledge by exploiting unintended semantics. RSs compromise reliability and generalization and, as we show in this paper, they are linked to NeSy models being overconfident about the predicted concepts. Unfortunately, the only trustworthy mitigation strategy requires collecting costly dense supervision over the concepts. Rather than attempting to avoid RSs altogether, we propose to ensure NeSy models are aware of the semantic ambiguity of the concepts they learn, thus enabling their users to identify and distrust low-quality concepts. Starting from three simple desiderata, we derive bears (BE Aware of Reasoning Shortcuts), an ensembling technique that calibrates the model's concept-level confidence without compromising prediction accuracy, thus encouraging NeSy architectures to be uncertain about concepts affected by RSs. We show empirically that bears improves RS-awareness of several state-of-the-art NeSy models, and also facilitates acquiring informative dense annotations for mitigation purposes.

We propose ABCs (Adaptive Branching through Child stationarity), a best-of-both-worlds algorithm combining Boltzmann Q-learning (BQL), a classic reinforcement learning algorithm for single-agent domains, and counterfactual regret minimization (CFR), a central algorithm for learning in multi-agent domains. ABCs adaptively chooses what fraction of the environment to explore each iteration by measuring the stationarity of the environment's reward and transition dynamics. In Markov decision processes, ABCs converges to the optimal policy with at most an O(A) factor slowdown compared to BQL, where A is the number of actions in the environment. In two-player zero-sum games, ABCs is guaranteed to converge to a Nash equilibrium (assuming access to a perfect oracle for detecting stationarity), while BQL has no such guarantees. Empirically, ABCs demonstrates strong performance when benchmarked across environments drawn from the OpenSpiel game library and OpenAI Gym and exceeds all prior methods in environments which are neither fully stationary nor fully nonstationary.

Large language models (LLMs) exhibit a remarkable ability to perform in-context learning (ICL): learning from patterns in their input context (Brown et al., 2020; Dong et al., 2022). The ability of LLMs to adaptively learn from context is profoundly useful, yet the underlying mechanisms of this emergent capability are not fully understood. In an effort to better understand ICL, some recent works propose to study ICL in controlled synthetic settings--in particular, training transformers on mathematically defined tasks which require learning from the input context. For example, a recent line of works studies the ability of transformers to perform ICL of standard supervised learning problems such as linear regression (Garg et al., 2022; Akyürek et al., 2022; Li et al., 2023; Wu et al., 2023). Studying these well-understood synthetic learning tasks enables fine-grained control over the data distribution, allows for comparisons with established supervised learning algorithms, and facilitates the examination of the in-context "algorithm" implemented by the network. That said, these supervised settings are reflective specifically of few-shot learning, which is only a special case of the more general phenomenon of networks incorporating patterns from their context into their predictions. A few recent works (Bietti et al., 2023; Xie et al., 2022) go beyond the case of cleanly separated in-context inputs and outputs, studying in-context learning on distributions based on discrete stochastic processes. Work done while visiting Harvard University.

Abstract--Federated learning allows us to run machine learning algorithms on decentralized data when data sharing is not permitted due to privacy concerns. Ensemble-based learning works by training multiple (weak) classifiers whose output is aggregated. Federated ensembles are ensembles applied to a federated setting, where each classifier in the ensemble is trained on one data location. In this article, we explore the use of federated ensembles of Bayesian networks (FBNE) in a range of experiments and compare their performance with locally trained models and models trained with VertiBayes, a federated learning algorithm to train Bayesian networks from decentralized data. Our results show that FBNE outperforms local models and provides a significant increase in training speed compared with VertiBayes while maintaining a similar performance in most settings, among other advantages. We show that FBNE is a potentially useful tool within the federated learning toolbox, especially when local populations are heavily biased, or there is a strong imbalance in population size across parties. We discuss the advantages and disadvantages of this approach in terms of time complexity, model accuracy, privacy protection, and model interpretability.

In solving the non-myopic radar scheduling for multiple smart target tracking within an active and passive radar network, we need to consider both short-term enhanced tracking performance and a higher probability of target maneuvering in the future with active tracking. Acquiring the long-term tracking performance while scheduling the beam resources of active and passive radars poses a challenge. To address this challenge, we model this problem as a Markov decision process consisting of parallel restless bandit processes. Each bandit process is associated with a smart target, of which the estimation state evolves according to different discrete dynamic models for different actions - whether or not the target is being tracked. The discrete state is defined by the dynamic mode. The problem exhibits the curse of dimensionality, where optimal solutions are in general intractable. We resort to heuristics through the famous restless multi-armed bandit techniques. It follows with efficient scheduling policies based on the indices that are real numbers representing the marginal rewards of taking different actions. For the inevitable practical case with unknown transition matrices, we propose a new method that utilizes the forward Sarsa and backward Q-learning to approximate the indices through adapting the state-action value functions, or equivalently the Q-functions, and propose a new policy, namely ISQ, aiming to maximize the long-term tracking rewards. Numerical results demonstrate that the proposed ISQ policy outperforms conventional Q-learning-based methods and rapidly converges to the well-known Whittle index policy with revealed state transition models, which is considered the benchmark.

Bayesian networks model relationships between random variables under uncertainty and can be used to predict the likelihood of events and outcomes while incorporating observed evidence. From an eXplainable AI (XAI) perspective, such models are interesting as they tend to be compact. Moreover, captured relations can be directly inspected by domain experts. In practice, data is often real-valued. Unless assumptions of normality can be made, discretization is often required. The optimal discretization, however, depends on the relations modelled between the variables. This complicates learning Bayesian networks from data. For this reason, most literature focuses on learning conditional dependencies between sets of variables, called structure learning. In this work, we extend an existing state-of-the-art structure learning approach based on the Gene-pool Optimal Mixing Evolutionary Algorithm (GOMEA) to jointly learn variable discretizations. The proposed Discretized Bayesian Network GOMEA (DBN-GOMEA) obtains similar or better results than the current state-of-the-art when tasked to retrieve randomly generated ground-truth networks. Moreover, leveraging a key strength of evolutionary algorithms, we can straightforwardly perform DBN learning multi-objectively. We show how this enables incorporating expert knowledge in a uniquely insightful fashion, finding multiple DBNs that trade-off complexity, accuracy, and the difference with a pre-determined expert network.

Pre-trained language models (LMs) are capable of in-context learning (ICL): they can adapt to a task with only a few examples given in the prompt without any parameter update. However, it is unclear where this capability comes from as there is a stark distribution shift between pre-training text and ICL prompts. In this work, we study what patterns of the pre-training data contribute to ICL. We find that LMs' ICL ability depends on $\textit{parallel structures}$ in the pre-training data -- pairs of phrases following similar templates in the same context window. Specifically, we detect parallel structures by checking whether training on one phrase improves prediction of the other, and conduct ablation experiments to study their effect on ICL. We show that removing parallel structures in the pre-training data reduces LMs' ICL accuracy by 51% (vs 2% from random ablation). This drop persists even when excluding common patterns such as n-gram repetitions and long-range dependency, showing the diversity and generality of parallel structures. A closer look at the detected parallel structures indicates that they cover diverse linguistic tasks and span long distances in the data.