We study Thompson sampling (TS) in online decision-making problems where the uncertain environment is sampled from a mixture distribution. This is relevant to multi-task settings, where a learning agent is faced with different classes of problems. We incorporate this structure in a natural way by initializing TS with a mixture prior -- dubbed MixTS -- and develop a novel, general technique for analyzing the regret of TS with such priors. We apply this technique to derive Bayes regret bounds for MixTS in both linear bandits and tabular Markov decision processes (MDPs). Our regret bounds reflect the structure of the problem and depend on the number of components and confidence width of each component of the prior. Finally, we demonstrate the empirical effectiveness of MixTS in both synthetic and real-world experiments.

We consider the quickest change detection problem where both the parameters of pre- and post- change distributions are unknown, which prevents the use of classical simple hypothesis testing. Without additional assumptions, optimal solutions are not tractable as they rely on some minimax and robust variant of the objective. As a consequence, change points might be detected too late for practical applications (in economics, health care or maintenance for instance). Available constant complexity techniques typically solve a relaxed version of the problem, deeply relying on very specific probability distributions and/or some very precise additional knowledge. We consider a totally different approach that leverages the theoretical asymptotic properties of optimal solutions to derive a new scalable approximate algorithm with near optimal performance that runs~in~$\mathcal{O}(1)$, adapted to even more complex Markovian settings.

Among different ML models, Deep Neural Networks (DNNs) [130] are well-known and widely used for their powerful representation learning from high-dimensional data such as images, texts, and speech. However, as ML algorithms enter sensitive real-world domains with trustworthiness, safety, and fairness prerequisites, the need for corresponding techniques and metrics for high-stake domains is more noticeable than before. Hence, researchers in different fields propose guidelines for Trustworthy AI [208], Safe AI [5], and Explainable AI [155] as stepping stones for next generation Responsible AI [6, 247]. Furthermore, government reports and regulations on AI accountability [75], trustworthiness [216], and safety [31] are gradually creating mandating laws to protect citizens' data privacy, fair data processing, and upholding safety for AI-based products. The development and deployment of ML algorithms for open-world tasks come with reliability and dependability limitations rooting from model performance, robustness, and uncertainty limitations [156]. Unlike traditional code-based software, ML models have fundamental safety drawbacks, including performance limitations on their training set and run-time robustness in their operational domain.

Communication is a cornerstone of social interactions, be it with human or artificial intelligence (AI). Yet it can be harmful, depending on the honesty of the exchanged information. To study this, an agent based sociological simulation framework is presented, the reputation game. This illustrates the impact of different communication strategies on the agents' reputation. The game focuses on the trustworthiness of the participating agents, their honesty as perceived by others. In the game, each agent exchanges statements with the others about their own and each other's honesty, which lets their judgments evolve. Various sender and receiver strategies are studied, like sycophant, egocentricity, pathological lying, and aggressiveness for senders as well as awareness and lack thereof for receivers. Minimalist malicious strategies are identified, like being manipulative, dominant, or destructive, which significantly increase reputation at others' costs. Phenomena such as echo chambers, self-deception, deception symbiosis, clique formation, freezing of group opinions emerge from the dynamics. This indicates that the reputation game can be studied for complex group phenomena, to test behavioral hypothesis, and to analyze AI influenced social media. With refined rules it may help to understand social interactions, and to safeguard the design of non-abusive AI systems.

Attention-based neural networks have achieved state-of-the-art results on a wide range of tasks. Most such models use deterministic attention while stochastic attention is less explored due to the optimization difficulties or complicated model design. This paper introduces Bayesian attention belief networks, which construct a decoder network by modeling unnormalized attention weights with a hierarchy of gamma distributions, and an encoder network by stacking Weibull distributions with a deterministic-upward-stochastic-downward structure to approximate the posterior. The resulting auto-encoding networks can be optimized in a differentiable way with a variational lower bound. It is simple to convert any models with deterministic attention, including pretrained ones, to the proposed Bayesian attention belief networks. On a variety of language understanding tasks, we show that our method outperforms deterministic attention and state-of-the-art stochastic attention in accuracy, uncertainty estimation, generalization across domains, and robustness to adversarial attacks. We further demonstrate the general applicability of our method on neural machine translation and visual question answering, showing great potential of incorporating our method into various attention-related tasks.

It is often remarked that neural networks fail to increase their uncertainty when predicting on data far from the training distribution. Yet naively using softmax confidence as a proxy for uncertainty achieves modest success in tasks exclusively testing for this, e.g., out-of-distribution (OOD) detection. This paper investigates this contradiction, identifying two implicit biases that do encourage softmax confidence to correlate with epistemic uncertainty: 1) Approximately optimal decision boundary structure, and 2) Filtering effects of deep networks. It describes why low-dimensional intuitions about softmax confidence are misleading. Diagnostic experiments quantify reasons softmax confidence can fail, finding that extrapolations are less to blame than overlap between training and OOD data in final-layer representations. Pre-trained/fine-tuned networks reduce this overlap.

Recent theoretical work studies sample-efficient reinforcement learning (RL) extensively in two settings: learning interactively in the environment (online RL), or learning from an offline dataset (offline RL). However, existing algorithms and theories for learning near-optimal policies in these two settings are rather different and disconnected. Towards bridging this gap, this paper initiates the theoretical study of policy finetuning, that is, online RL where the learner has additional access to a "reference policy" $\mu$ close to the optimal policy $\pi_\star$ in a certain sense. We consider the policy finetuning problem in episodic Markov Decision Processes (MDPs) with $S$ states, $A$ actions, and horizon length $H$. We first design a sharp offline reduction algorithm -- which simply executes $\mu$ and runs offline policy optimization on the collected dataset -- that finds an $\varepsilon$ near-optimal policy within $\widetilde{O}(H^3SC^\star/\varepsilon^2)$ episodes, where $C^\star$ is the single-policy concentrability coefficient between $\mu$ and $\pi_\star$. This offline result is the first that matches the sample complexity lower bound in this setting, and resolves a recent open question in offline RL. We then establish an $\Omega(H^3S\min\{C^\star, A\}/\varepsilon^2)$ sample complexity lower bound for any policy finetuning algorithm, including those that can adaptively explore the environment. This implies that -- perhaps surprisingly -- the optimal policy finetuning algorithm is either offline reduction or a purely online RL algorithm that does not use $\mu$. Finally, we design a new hybrid offline/online algorithm for policy finetuning that achieves better sample complexity than both vanilla offline reduction and purely online RL algorithms, in a relaxed setting where $\mu$ only satisfies concentrability partially up to a certain time step.

Probabilistic deep learning is deep learning that accounts for uncertainty, both model uncertainty and data uncertainty. It is based on the use of probabilistic models and deep neural networks. We distinguish two approaches to probabilistic deep learning: probabilistic neural networks and deep probabilistic models. The former employs deep neural networks that utilize probabilistic layers which can represent and process uncertainty; the latter uses probabilistic models that incorporate deep neural network components which capture complex non-linear stochastic relationships between the random variables. We discuss some major examples of each approach including Bayesian neural networks and mixture density networks (for probabilistic neural networks), and variational autoencoders, deep Gaussian processes and deep mixed effects models (for deep probabilistic models). TensorFlow Probability is a library for probabilistic modeling and inference which can be used for both approaches of probabilistic deep learning. We include its code examples for illustration.

This paper establishes a central limit theorem under the assumption that conditional variances can vary in a largely unstructured history-dependent way across experiments subject only to the restriction that they lie in a fixed interval. Limits take a novel and tractable form, and are expressed in terms of oscillating Brownian motion. A second contribution is application of this result to a class of multi-armed bandit problems where the decision-maker is loss averse.

Decision makers involved in the management of civil assets and systems usually take actions under constraints imposed by societal regulations. Some of these constraints are related to epistemic quantities, as the probability of failure events and the corresponding risks. Sensors and inspectors can provide useful information supporting the control process (e.g. the maintenance process of an asset), and decisions about collecting this information should rely on an analysis of its cost and value. When societal regulations encode an economic perspective that is not aligned with that of the decision makers, the Value of Information (VoI) can be negative (i.e., information sometimes hurts), and almost irrelevant information can even have a significant value (either positive or negative), for agents acting under these epistemic constraints. We refer to these phenomena as Information Avoidance (IA) and Information OverValuation (IOV). In this paper, we illustrate how to assess VoI in sequential decision making under epistemic constraints (as those imposed by societal regulations), by modeling a Partially Observable Markov Decision Processes (POMDP) and evaluating non optimal policies via Finite State Controllers (FSCs). We focus on the value of collecting information at current time, and on that of collecting sequential information, we illustrate how these values are related and we discuss how IA and IOV can occur in those settings.