Search for: All records

AI assistance is readily available to humans in a variety of decision-making applications. In order to fully understand the efficacy of such joint decision-making, it is important to first understand the human’s reliance on AI. However, there is a disconnect between how joint decision-making is studied and how it is practiced in the real world. More often than not, researchers ask humans to provide independent decisions before they are shown AI assistance. This is done to make explicit the influence of AI assistance on the human’s decision. We develop a cognitive model that allows us to infer thelatentreliance strategy of humans on AI assistance without asking the human to make an independent decision. We validate the model’s predictions through two behavioral experiments. The first experiment follows aconcurrentparadigm where humans are shown AI assistance alongside the decision problem. The second experiment follows asequentialparadigm where humans provide an independent judgment on a decision problem before AI assistance is made available. The model’s predicted reliance strategies closely track the strategies employed by humans in the two experimental paradigms. Our model provides a principled way to infer reliance on AI-assistance and may be used to expand the scope of investigation on human-AI collaboration.

 

The deployment of machine learning classifiers in high-stakes domains requires well-calibrated confidence scores for model predictions. In this paper we introduce the notion of variable-based calibration to characterize calibration properties of a model with respect to a variable of interest, generalizing traditional score-based metrics such as expected calibration error (ECE). In particular, we find that models with near-perfect ECE can exhibit significant miscalibration as a function of features of the data. We demonstrate this phenomenon both theoretically and in practice on multiple well-known datasets, and show that it can persist after the application of existing calibration methods. To mitigate this issue, we propose strategies for detection, visualization, and quantification of variable-based calibration error. We then examine the limitations of current score-based calibration methods and explore potential modifications. Finally, we discuss the implications of these findings, emphasizing that an understanding of calibration beyond simple aggregate measures is crucial for endeavors such as fairness and model interpretability. 

We expose the statistical foundations of deep learning with the goal of facilitating conversation between the deep learning and statistics communities. We highlight core themes at the intersection; summarize key neural models, such as feedforward neural networks, sequential neural networks, and neural latent variable models; and link these ideas to their roots in probability and statistics. We also highlight research directions in deep learning where there are opportunities for statistical contributions. 

Artificial intelligence (AI) and machine learning models are being increasingly deployed in real-world applications. In many of these applications, there is strong motivation to develop hybrid systems in which humans and AI algorithms can work together, leveraging their complementary strengths and weaknesses. We develop a Bayesian framework for combining the predictions and different types of confidence scores from humans and machines. The framework allows us to investigate the factors that influence complementarity, where a hybrid combination of human and machine predictions leads to better performance than combinations of human or machine predictions alone. We apply this framework to a large-scale dataset where humans and a variety of convolutional neural networks perform the same challenging image classification task. We show empirically and theoretically that complementarity can be achieved even if the human and machine classifiers perform at different accuracy levels as long as these accuracy differences fall within a bound determined by the latent correlation between human and machine classifier confidence scores. In addition, we demonstrate that hybrid human–machine performance can be improved by differentiating between the errors that humans and machine classifiers make across different class labels. Finally, our results show that eliciting and including human confidence ratings improve hybrid performance in the Bayesian combination model. Our approach is applicable to a wide variety of classification problems involving human and machine algorithms. 

Recent advances in machine learning have led to increased deployment of black-box classifiers across a wide variety of applications. In many such situations there is a critical need to both reliably assess the performance of these pre-trained models and to perform this assessment in a label-efficient manner (given that labels may be scarce and costly to collect). In this paper, we introduce an active Bayesian approach for assessment of classifier performance to satisfy the desiderata of both reliability and label-efficiency. We begin by developing inference strategies to quantify uncertainty for common assessment metrics such as accuracy, misclassification cost, and calibration error. We then propose a general framework for active Bayesian assessment using inferred uncertainty to guide efficient selection of instances for labeling, enabling better performance assessment with fewer labels. We demonstrate significant gains from our proposed active Bayesian approach via a series of systematic empirical experiments assessing the performance of modern neural classifiers (e.g., ResNet and BERT) on several standard image and text classification datasets. 

An increasingly common use case for machine learning models is augmenting the abilities of human decision makers. For classification tasks where neither the human nor model are perfectly accurate, a key step in obtaining high performance is combining their individual predictions in a manner that leverages their relative strengths. In this work, we develop a set of algorithms that combine the probabilistic output of a model with the class-level output of a human. We show theoretically that the accuracy of our combination model is driven not only by the individual human and model accuracies, but also by the model's confidence. Empirical results on image classification with CIFAR-10 and a subset of ImageNet demonstrate that such human-model combinations consistently have higher accuracies than the model or human alone, and that the parameters of the combination method can be estimated effectively with as few as ten labeled datapoints. 

Human-AI collaboration is an increasingly commonplace part of decision-making in real world applications. However, how humans behave when collaborating with AI is not well understood. We develop metacognitive bandits, a computational model of a human's advice-seeking behavior when working with an AI. The model describes a person's metacognitive process of deciding when to rely on their own judgment and when to solicit the advice of the AI. It also accounts for the difficulty of each trial in making the decision to solicit advice. We illustrate that the metacognitive bandit makes decisions similar to humans in a behavioral experiment. We also demonstrate that algorithm aversion, a widely reported bias, can be explained as the result of a quasi-optimal sequential decision-making process. Our model does not need to assume any prior biases towards AI to produce this behavior. 

Group fairness is measured via parity of quantitative metrics across different protected demographic groups. In this paper, we investigate the problem of reliably assessing group fairness metrics when labeled examples are few but unlabeled examples are plentiful. We propose a general Bayesian framework that can augment labeled data with unlabeled data to produce more accurate and lower-variance estimates compared to methods based on labeled data alone. Our approach estimates calibrated scores (for unlabeled examples) of each group using a hierarchical latent variable model conditioned on labeled examples. This in turn allows for inference of posterior distributions for an array of group fairness metrics with a notion of uncertainty. We demonstrate that our approach leads to significant and consistent reductions in estimation error across multiple well-known fairness datasets, sensitive attributes, and predictive models. The results clearly show the benefits of using both unlabeled data and Bayesian inference in assessing whether a prediction model is fair or not.