Whenever a binary classifier is used to provide decision support, it typically provides both a label prediction and a confidence value. Then, the decision maker is supposed to use the confidence value to calibrate how much to trust the prediction. In this context, it has been often argued that the confidence value should correspond to a well calibrated estimate of the probability that the predicted label matches the ground truth label. However, multiple lines of empirical evidence suggest that decision makers have difficulties at developing a good sense on when to trust a prediction using these confidence values. In this paper, our goal is first to understand why and then investigate how to construct more useful confidence values. We first argue that, for a broad class of utility functions, there exist data distributions for which a rational decision maker is, in general, unlikely to discover the optimal decision policy using the above confidence values -- an optimal decision maker would need to sometimes place more (less) trust on predictions with lower (higher) confidence values. However, we then show that, if the confidence values satisfy a natural alignment property with respect to the decision maker's confidence on her own predictions, there always exists an optimal decision policy under which the level of trust the decision maker would need to place on predictions is monotone on the confidence values, facilitating its discoverability. Further, we show that multicalibration with respect to the decision maker's confidence on her own predictions is a sufficient condition for alignment. Experiments on four different AI-assisted decision making tasks where a classifier provides decision support to real human experts validate our theoretical results and suggest that alignment may lead to better decisions.
相關內容
In this work, we give a statistical characterization of the $\gamma$-regret for arbitrary structured bandit problems, the regret which arises when comparing against a benchmark that is $\gamma$ times the optimal solution. The $\gamma$-regret emerges in structured bandit problems over a function class $\mathcal{F}$ where finding an exact optimum of $f \in \mathcal{F}$ is intractable. Our characterization is given in terms of the $\gamma$-DEC, a statistical complexity parameter for the class $\mathcal{F}$, which is a modification of the constrained Decision-Estimation Coefficient (DEC) of Foster et al., 2023 (and closely related to the original offset DEC of Foster et al., 2021). Our lower bound shows that the $\gamma$-DEC is a fundamental limit for any model class $\mathcal{F}$: for any algorithm, there exists some $f \in \mathcal{F}$ for which the $\gamma$-regret of that algorithm scales (nearly) with the $\gamma$-DEC of $\mathcal{F}$. We provide an upper bound showing that there exists an algorithm attaining a nearly matching $\gamma$-regret. Due to significant challenges in applying the prior results on the DEC to the $\gamma$-regret case, both our lower and upper bounds require novel techniques and a new algorithm.
We propose a model for learning with bandit feedback while accounting for deterministically evolving and unobservable states that we call Bandits with Deterministically Evolving States. The workhorse applications of our model are learning for recommendation systems and learning for online ads. In both cases, the reward that the algorithm obtains at each round is a function of the short-term reward of the action chosen and how ``healthy'' the system is (i.e., as measured by its state). For example, in recommendation systems, the reward that the platform obtains from a user's engagement with a particular type of content depends not only on the inherent features of the specific content, but also on how the user's preferences have evolved as a result of interacting with other types of content on the platform. Our general model accounts for the different rate $\lambda \in [0,1]$ at which the state evolves (e.g., how fast a user's preferences shift as a result of previous content consumption) and encompasses standard multi-armed bandits as a special case. The goal of the algorithm is to minimize a notion of regret against the best-fixed sequence of arms pulled. We analyze online learning algorithms for any possible parametrization of the evolution rate $\lambda$. Specifically, the regret rates obtained are: for $\lambda \in [0, 1/T^2]$: $\widetilde O(\sqrt{KT})$; for $\lambda = T^{-a/b}$ with $b < a < 2b$: $\widetilde O (T^{b/a})$; for $\lambda \in (1/T, 1 - 1/\sqrt{T}): \widetilde O (K^{1/3}T^{2/3})$; and for $\lambda \in [1 - 1/\sqrt{T}, 1]: \widetilde O (K\sqrt{T})$.
Feature selection is popular for obtaining small, interpretable, yet highly accurate prediction models. Conventional feature-selection methods typically yield one feature set only, which might not suffice in some scenarios. For example, users might be interested in finding alternative feature sets with similar prediction quality, offering different explanations of the data. In this article, we introduce alternative feature selection and formalize it as an optimization problem. In particular, we define alternatives via constraints and enable users to control the number and dissimilarity of alternatives. Next, we analyze the complexity of this optimization problem and show NP-hardness. Further, we discuss how to integrate conventional feature-selection methods as objectives. Finally, we evaluate alternative feature selection with 30 classification datasets. We observe that alternative feature sets may indeed have high prediction quality, and we analyze several factors influencing this outcome.
Pre-trained language models (PLMs) serve as backbones for various real-world systems. For high-stake applications, it's equally essential to have reasonable confidence estimations in predictions. While the vanilla confidence scores of PLMs can already be effectively utilized, PLMs consistently become overconfident in their wrong predictions, which is not desirable in practice. Previous work shows that introducing an extra calibration task can mitigate this issue. The basic idea involves acquiring additional data to train models in predicting the confidence of their initial predictions. However, it only demonstrates the feasibility of this kind of method, assuming that there are abundant extra available samples for the introduced calibration task. In this work, we consider the practical scenario that we need to effectively utilize training samples to make PLMs both task-solvers and self-calibrators. Three challenges are presented, including limited training samples, data imbalance, and distribution shifts. We first conduct pilot experiments to quantify various decisive factors in the calibration task. Based on the empirical analysis results, we propose a training algorithm LM-TOAST to tackle the challenges. Experimental results show that LM-TOAST can effectively utilize the training data to make PLMs have reasonable confidence estimations while maintaining the original task performance. Further, we consider three downstream applications, namely selective classification, adversarial defense, and model cascading, to show the practical usefulness of LM-TOAST. The code will be made public at \url{//github.com/Yangyi-Chen/LM-TOAST}.
In settings where users are both time-pressured and need high accuracy, such as doctors working in Emergency Rooms, we want to provide AI assistance that both increases accuracy and reduces time. However, different types of AI assistance have different benefits: some reduce time taken while increasing overreliance on AI, while others do the opposite. We therefore want to adapt what AI assistance we show depending on various properties (of the question and of the user) in order to best tradeoff our two objectives. We introduce a study where users have to prescribe medicines to aliens, and use it to explore the potential for adapting AI assistance. We find evidence that it is beneficial to adapt our AI assistance depending on the question, leading to good tradeoffs between time taken and accuracy. Future work would consider machine-learning algorithms (such as reinforcement learning) to automatically adapt quickly.
Performance of a pre-trained semantic segmentation model is likely to substantially decrease on data from a new domain. We show a pre-trained model can be adapted to unlabelled target domain data by calculating soft-label prototypes under the domain shift and making predictions according to the prototype closest to the vector with predicted class probabilities. The proposed adaptation procedure is fast, comes almost for free in terms of computational resources and leads to considerable performance improvements. We demonstrate the benefits of such label calibration on the highly-practical synthetic-to-real semantic segmentation problem.
We consider the problem of group interactions in urban driving. State-of-the-art behavior planners for self-driving cars mostly consider each single agent-to-agent interaction separately in a cost function in order to find an optimal behavior for the ego agent, such as not colliding with any of the other agents. In this paper, we develop risk shadowing, a situation understanding method that allows us to go beyond single interactions by analyzing group interactions between three agents. Concretely, the presented method can find out which first other agent does not need to be considered in the behavior planner of an ego agent, because this first other agent cannot reach the ego agent due to a second other agent obstructing its way. In experiments, we show that using risk shadowing as an upstream filter module for a behavior planner allows to plan more decisive and comfortable driving strategies than state of the art, given that safety is ensured in these cases. The usability of the approach is demonstrated for different intersection scenarios and longitudinal driving.
A Bayesian Programming Approach to Car-following Model Calibration and Validation using Limited Data
Traffic simulation software is used by transportation researchers and engineers to design and evaluate changes to roadways. These simulators are driven by models of microscopic driver behavior from which macroscopic measures like flow and congestion can be derived. Many models are designed for a subset of possible traffic scenarios and roadway configurations, while others have no explicit constraints on their application. Work zones (WZs) are one scenario for which no model to date has reproduced realistic driving behavior. This makes it difficult to optimize for safety and other metrics when designing a WZ. The Federal Highway Administration commissioned the USDOT Volpe Center to develop a car-following (CF) model for use in microscopic simulators that can capture and reproduce driver behavior accurately within and outside of WZs. Volpe also performed a naturalistic driving study to collect telematics data from vehicles driven on roads with WZs for use in model calibration. During model development, Volpe researchers observed difficulties in calibrating their model, leaving them to question whether there existed flaws in their model, in the data, or in the procedure used to calibrate the model using the data. In this thesis, I use Bayesian methods for data analysis and parameter estimation to explore and, where possible, address these questions. First, I use Bayesian inference to measure the sufficiency of the size of the data set. Second, I compare the procedure and results of the genetic algorithm based calibration performed by the Volpe researchers with those of Bayesian calibration. Third, I explore the benefits of modeling CF hierarchically. Finally, I apply what was learned in the first three phases using an established CF model, Wiedemann 99, to the probabilistic modeling of the Volpe model. Validation is performed using information criteria as an estimate of predictive accuracy.
Prophet inequalities consist of many beautiful statements that establish tight performance ratios between online and offline allocation algorithms. Typically, tightness is established by constructing an algorithmic guarantee and a worst-case instance separately, whose bounds match as a result of some "ingenuity". In this paper, we instead formulate the construction of the worst-case instance as an optimization problem, which directly finds the tight ratio without needing to construct two bounds separately. Our analysis of this complex optimization problem involves identifying the structure in a new "Type Coverage" dual problem. It can be seen as akin to the celebrated Magician and OCRS problems, except more general in that it can also provide tight ratios relative to the optimal offline allocation, whereas the earlier problems only concerns the ex-ante relaxation of the offline problem. Through this analysis, our paper provides a unified framework that derives new prophet inequalities and recovers existing ones, including two important new results. First, we show that the "oblivious" method of setting a static threshold due to Chawla et al. (2020), surprisingly, is best-possible among all static threshold algorithms, under any number $k$ of units. We emphasize that this result is derived without needing to explicitly find any counterexample instances. This implies the tightness of the asymptotic convergence rate of $1-O(\sqrt{\log k/k})$ for static threshold algorithms from Hajiaghayi et al. (2007), is tight; this confirms for the first time a separation with the convergence rate of adaptive algorithms, which is $1-\Theta(\sqrt{1/k})$ due to Alaei (2014). Second, turning to the IID setting, our framework allows us to numerically illustrate the tight guarantee (of adaptive algorithms) under any number $k$ of starting units. Our guarantees for $k>1$ exceed the state-of-the-art.
Moderate calibration, the expected event probability among observations with predicted probability $\pi$ being equal to $\pi$, is a desired property of risk prediction models. Current graphical and numerical techniques for evaluating moderate calibration of clinical prediction models are mostly based on smoothing or grouping the data. As well, there is no widely accepted inferential method for the null hypothesis that a model is moderately calibrated. In this work, we discuss recently-developed, and propose novel, methods for the assessment of moderate calibration for binary responses. The methods are based on the limiting distributions of functions of standardized partial sums of prediction errors converging to the corresponding laws of Brownian motion. The novel method relies on well-known properties of the Brownian bridge which enables joint inference on mean and moderate calibration, leading to a unified 'bridge' test for detecting miscalibration. Simulation studies indicate that the bridge test is more powerful, often substantially, than the alternative test. As a case study we consider a prediction model for short-term mortality after a heart attack. Moderate calibration can be assessed without requiring arbitrary grouping of data or using methods that require tuning of parameters. We suggest graphical presentation of the partial sum curves and reporting the strength of evidence indicated by the proposed methods when examining model calibration.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191