Large language models (LLMs) frequently hallucinate on abstractive summarization tasks such as document-based question-answering, meeting summarization, and clinical report generation, even though all necessary information is included in context. However, optimizing to make LLMs hallucinate less is challenging, as hallucination is hard to efficiently, cheaply, and reliably evaluate at each optimization step. In this work, we show that reducing hallucination on a _synthetic task_ can also reduce hallucination on real-world downstream tasks. Our method, SynTra, first designs a synthetic task where hallucinations are easy to elicit and measure. It next optimizes the LLM's system message via prefix tuning on the synthetic task, then uses the system message on realistic, hard-to-optimize tasks. Across three realistic abstractive summarization tasks, we reduce hallucination for two 13B-parameter LLMs using supervision signal from only a synthetic retrieval task. We also find that optimizing the system message rather than the model weights can be critical; fine-tuning the entire model on the synthetic task can counterintuitively _increase_ hallucination. Overall, SynTra demonstrates that the extra flexibility of working with synthetic data can help mitigate undesired behaviors in practice.

We investigate the internal behavior of Transformer-based Large Language Models (LLMs) when they generate factually incorrect text. We propose modeling factual queries as constraint satisfaction problems and use this framework to investigate how the LLM interacts internally with factual constraints. We find a strong positive relationship between the LLM's attention to constraint tokens and the factual accuracy of generations. We curate a suite of 10 datasets containing over 40,000 prompts to study the task of predicting factual errors with the Llama-2 family across all scales (7B, 13B, 70B). We propose SAT Probe, a method probing attention patterns, that can predict factual errors and fine-grained constraint satisfaction, and allow early error identification. The approach and findings take another step towards using the mechanistic understanding of LLMs to enhance their reliability.

Large language models (LLMs) can be used to generate text data for training and evaluating other models.However, creating highquality datasets with LLMs can be challenging.In this work, we explore human-AI partnerships to facilitate high diversity and accuracy in LLM-based text data generation.We first examine two approaches to diversify text generation: 1) logit suppression, which minimizes the generation of languages that have already been frequently generated, and 2) temperature sampling, which flattens the token sampling probability.We found that diversification approaches can increase data diversity but often at the cost of data accuracy (i.e., text and labels being appropriate for the target domain).To address this issue, we examined two human interventions, 1) label replacement (LR), correcting misaligned labels, and 2) out-of-scope filtering (OOSF), removing instances that are out of the user's domain of interest or to which no considered label applies.With oracle studies, we found that LR increases the absolute accuracy of models trained with diversified datasets by 14.4%.Moreover, we found that some models trained with data generated with LR interventions outperformed LLM-based few-shot classification.In contrast, OOSF was not effective in increasing model accuracy, implying the need for future work in human-in-the-loop text data generation.

In AI-assisted decision-making, effective hybrid (human-AI) teamwork is not solely dependent on AI performance alone, but also on its impact on human decision-making. While prior work studies the effects of model accuracy on humans, we endeavour here to investigate the complex dynamics of how both a model's predictive performance and bias may transfer to humans in a recommendation-aided decision task. We consider the domain of ML-assisted hiring, where humans---operating in a constrained selection setting---can choose whether they wish to utilize a trained model's inferences to help select candidates from written biographies. We conduct a large-scale user study leveraging a re-created dataset of real bios from prior work, where humans predict the ground truth occupation of given candidates with and without the help of three different NLP classifiers (random, bag-of-words, and deep neural network). Our results demonstrate that while high-performance models significantly improve human performance in a hybrid setting, some models mitigate hybrid bias while others accentuate it. We examine these findings through the lens of decision conformity and observe that our model architecture choices have an impact on human-AI conformity and bias, motivating the explicit need to assess these complex dynamics prior to deployment.

Thomas Hartvigsen, Saadia Gabriel, Hamid Palangi, Maarten Sap, Dipankar Ray, Ece Kamar. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Autonomous systems that operate in the open world often use incomplete models of their environment. Model incompleteness is inevitable due to the practical limitations in precise model specification and data collection about open-world environments. Due to the limited fidelity of the model, agent actions may produce negative side effects (NSEs) when deployed. Negative side effects are undesirable, unmodeled effects of agent actions on the environment. NSEs are inherently challenging to identify at design time and may affect the reliability, usability and safety of the system. We present two complementary approaches to mitigate the NSE via: (1) learning from feedback, and (2) environment shaping. The solution approaches target settings with different assumptions and agent responsibilities. In learning from feedback, the agent learns a penalty function associated with a NSE. We investigate the efficiency of different feedback mechanisms, including human feedback and autonomous exploration. The problem is formulated as a multi-objective Markov decision process such that optimizing the agent’s assigned task is prioritized over mitigating NSE. A slack parameter denotes the maximum allowed deviation from the optimal expected reward for the agent’s task in order to mitigate NSE. In environment shaping, we examine how a human can assist an agent, beyond providing feedback, and utilize their broader scope of knowledge to mitigate the impacts of NSE. We formulate the problem as a human-agent collaboration with decoupled objectives. The agent optimizes its assigned task and may produce NSE during its operation. The human assists the agent by performing modest reconfigurations of the environment so as to mitigate the impacts of NSE, without affecting the agent’s ability to complete its assigned task. We present an algorithm for shaping and analyze its properties. Empirical evaluations demonstrate the trade-offs in the performance of different approaches in mitigating NSE in different settings.

AI practitioners typically strive to develop the most accurate systems, making an implicit assumption that the AI system will function autonomously. However, in practice, AI systems often are used to provide advice to people in domains ranging from criminal justice and finance to healthcare. In such AI-advised decision making, humans and machines form a team, where the human is responsible for making final decisions. But is the most accurate AI the best teammate? We argue "not necessarily" --- predictable performance may be worth a slight sacrifice in AI accuracy. Instead, we argue that AI systems should be trained in a human-centered manner, directly optimized for team performance. We study this proposal for a specific type of human-AI teaming, where the human overseer chooses to either accept the AI recommendation or solve the task themselves. To optimize the team performance for this setting we maximize the team's expected utility, expressed in terms of the quality of the final decision, cost of verifying, and individual accuracies of people and machines. Our experiments with linear and non-linear models on real-world, high-stakes datasets show that the most accuracy AI may not lead to highest team performance and show the benefit of modeling teamwork during training through improvements in expected team utility across datasets, considering parameters such as human skill and the cost of mistakes. We discuss the shortcoming of current optimization approaches beyond well-studied loss functions such as log-loss, and encourage future work on AI optimization problems motivated by human-AI collaboration.

Disaggregated evaluations of AI systems, in which system performance is assessed and reported separately for different groups of people, are conceptually simple. However, their design involves a variety of choices. Some of these choices influence the results that will be obtained, and thus the conclusions that can be drawn; others influence the impacts---both beneficial and harmful---that a disaggregated evaluation will have on people, including the people whose data is used to conduct the evaluation. We argue that a deeper understanding of these choices will enable researchers and practitioners to design careful and conclusive disaggregated evaluations. We also argue that better documentation of these choices, along with the underlying considerations and tradeoffs that have been made, will help others when interpreting an evaluation's results and conclusions.

A diverse representation of different demographic groups in AI training data sets is important in ensuring that the models will work for a large range of users. To this end, recent efforts in AI fairness and inclusion have advocated for creating AI data sets that are well-balanced across race, gender, socioeconomic status, and disability status. In this paper, we contribute to this line of work by focusing on the representation of age by asking whether older adults are represented proportionally to the population at large in AI data sets. We examine publicly-available information about 92 face data sets to understand how they codify age as a case study to investigate how the subjects' ages are recorded and whether older generations are represented. We find that older adults are very under-represented; five data sets in the study that explicitly documented the closed age intervals of their subjects included older adults (defined as older than 65 years), while only one included oldest-old adults (defined as older than 85 years). Additionally, we find that only 24 of the data sets include any age-related information in their documentation or metadata, and that there is no consistent method followed across these data sets to collect and record the subjects' ages. We recognize the unique difficulties in creating representative data sets in terms of age, but raise it as an important dimension that researchers and engineers interested in inclusive AI should consider.

Many researchers motivate explainable AI with studies showing that human-AI team performance on decision-making tasks improves when the AI explains its recommendations. However, prior studies observed improvements from explanations only when the AI, alone, outperformed both the human and the best team. Can explanations help lead to complementary performance, where team accuracy is higher than either the human or the AI working solo? We conduct mixed-method user studies on three datasets, where an AI with accuracy comparable to humans helps participants solve a task (explaining itself in some conditions). While we observed complementary improvements from AI augmentation, they were not increased by explanations. Rather, explanations increased the chance that humans will accept the AI’s recommendation, regardless of its correctness. Our result poses new challenges for human-centered AI: Can we develop explanatory approaches that encourage appropriate trust in AI, and therefore help generate (or improve) complementary performance?

Traditional evaluation metrics for learned models that report aggregate scores over a test set are insufficient for surfacing important and informative patterns of failure over features and instances. We introduce and study a method aimed at characterizing and explaining failures by identifying visual attributes whose presence or absence results in poor performance. In distinction to previous work that relies upon crowdsourced labels for visual attributes, we leverage the representation of a separate robust model to extract interpretable features and then harness these features to identify failure modes. We further propose a visualization method aimed at enabling humans to understand the meaning encoded in such features and we test the comprehensibility of the features. An evaluation of the methods on the ImageNet dataset demonstrates that: (i) the proposed workflow is effective for discovering important failure modes, (ii) the visualization techniques help humans to understand the extracted features, and (iii) the extracted insights can assist engineers with error analysis and debugging.

Given a fixed hypothesis space, defined to model class structure in a particular domain of application, unknown unknowns (u.u.s) are data examples that form classes in the feature space whose structure is not represented in a trained model. Accordingly, this leads to incorrect class prediction with high confidence, which represents one of the major sources of blind spots in machine learning. Our method seeks to reduce the structural mismatch between the training model and that of the target space in a supervised way. We illuminate further structure through cross-validation on a modified training model, set up to mine and trap u.u.s in a marginal training class, created from examples of a random sample of the test set. Contrary to previous approaches, our method simplifies the solution, as it does not rely on budgeted queries to an Oracle whose outcomes inform adjustments to training. In addition, our empirically results exhibit consistent performance improvements over baselines, on both synthetic and real-world data sets.

Understanding how racial information impacts human decision making in online systems is critical in today's world. Prior work revealed that race information of criminal defendants, when presented as a text field, had no significant impact on users' judgements of recidivism. We replicated and extended this work to explore how and when race information influences users' judgements, with respect to the saliency of presentation. Our results showed that adding photos to the race labels had a significant impact on recidivism predictions for users who identified as female, but not for those who identified as male. The race of the defendant also impacted these results, with black defendants being less likely to be predicted to recidivate compared to white defendants. These results have strong implications for how system-designers choose to display race information, and cautions researchers to be aware of gender and race effects when using Amazon Mechanical Turk workers.

Existing VQA datasets contain questions with varying levels of complexity. While the majority of questions in these datasets require perception for recognizing existence, properties, and spatial relationships of entities, a significant portion of questions pose challenges that correspond to reasoning tasks - tasks that can only be answered through a synthesis of perception and knowledge about the world, logic and / or reasoning. Analyzing performance across this distinction allows us to notice when existing VQA models have consistency issues - they answer the reasoning questions correctly but fail on associated low-level perception questions. For example, in Figure 1, models answer the complex reasoning question “Is the banana ripe enough to eat?” correctly, but fail on the associated perception question “Are the bananas mostly green or yellow?” indicating that the model likely answered the reasoning question correctly but for the wrong reason. We quantify the extent to which this phenomenon occurs by creating a new Reasoning split of the VQA dataset and collecting VQAintrospect, a new dataset1 which currently consists of 200K new perception questions which serve as sub questions corresponding to the set of perceptual tasks needed to effectively answer the complex reasoning questions in the Reasoning split. Our evaluation shows that state-of-the-art VQA models have comparable performance in answering perception and reasoning questions, but suffer from consistency problems. To address this shortcoming, we propose an approach called Sub-Question Importance-aware Network Tuning (SQuINT), which encourages the model to attend to the same parts of the image when answering the reasoning question and the perception sub question. We show that SQuINT improves model consistency by ~7%, also marginally improving performance on the Reasoning questions in VQA, while also displaying better attention maps.

Agents operating in unstructured environments often create negative side effects (NSE) that may not be easy to identify at design time. We examine how various forms of human feedback or autonomous exploration can be used to learn a penalty function associated with NSE during system deployment. We formulate the problem of mitigating the impact of NSE as a multi-objective Markov decision process with lexicographic reward preferences and slack. The slack denotes the maximum deviation from an optimal policy with respect to the agent's primary objective allowed in order to mitigate NSE as a secondary objective. Empirical evaluation of our approach shows that the proposed framework can successfully mitigate NSE and that different feedback mechanisms introduce different biases, which influence the identification of NSE.

A rising vision for AI in the open world centers on the development of systems that can complement humans for perceptual, diagnostic, and reasoning tasks. To date, systems aimed at complementing the skills of people have employed models trained to be as accurate as possible in isolation. We demonstrate how an end-to-end learning strategy can be harnessed to optimize the combined performance of human-machine teams by considering the distinct abilities of people and machines. The goal is to focus machine learning on problem instances that are difficult for humans, while recognizing instances that are difficult for the machine and seeking human input on them. We demonstrate in two real-world domains (scientific discovery and medical diagnosis) that human-machine teams built via these methods outperform the individual performance of machines and people. We then analyze conditions under which this complementarity is strongest, and which training methods amplify it. Taken together, our work provides the first systematic investigation of how machine learning systems can be trained to complement human reasoning.

In many applications of machine learning (ML), updates are performed with the goal of enhancing model performance. However, current practices for updating models rely solely on isolated, aggregate performance analyses, overlooking important dependencies, expectations, and needs in real-world deployments. We consider how updates, intended to improve ML models, can introduce new errors that can significantly affect downstream systems and users. For example, updates in models used in cloud-based classification services, such as image recognition, can cause unexpected erroneous behavior in systems that make calls to the services. Prior work has shown the importance of "backward compatibility" for maintaining human trust. We study challenges with backward compatibility across different ML architectures and datasets, focusing on common settings including data shifts with structured noise and ML employed in inferential pipelines. Our results show that (i) compatibility issues arise even without data shift due to optimization stochasticity, (ii) training on large-scale noisy datasets often results in significant decreases in backward compatibility even when model accuracy increases, and (iii) distributions of incompatible points align with noise bias, motivating the need for compatibility aware de-noising and robustness methods.

Predictive models are susceptible to errors called unknown unknowns, in which the model assigns incorrect labels to instances with high confidence. These commonly arise when training data does not represent variations of a class encountered at model deployment. Prior work showed that crowd workers can identify instances of unknown unknowns, but asking the crowd to identify a sufficient number of individual instances can be costly to acquire [2]. Instead, this paper presents an approach that leverages people’s ability to find patterns to retrain classifiers more effectively with fewer examples. We ask crowd workers to suggest and verify patterns in unknown unknowns. We then use these patterns to train an expansion classifier to identify additional examples from existing data that the primary classifier has encountered (and potentially misclassified) in the past. Our experiments show that our approach outperforms existing unknown unknown detection methods at improving classifier performance. This work is the first to leverage crowds to identify error patterns in large datasets to improve ML training.

Agents trained in simulation may make errors when performing actions in the real world due to mismatches between training and execution environments. These mistakes can be dangerous and difficult for the agent to discover because the agent is unable to predict them a priori. In this work, we propose the use of oracle feedback to learn a predictive model of these blind spots in order to reduce costly errors in real-world applications. We focus on blind spots in reinforcement learning (RL) that occur due to incomplete state representation: when the agent lacks necessary features to represent the true state of the world, and thus cannot distinguish between numerous states. We formalize the problem of discovering blind spots in RL as a noisy supervised learning problem with class imbalance. Our system learns models for predicting blind spots within unseen regions of the state space by combining techniques for label aggregation, calibration, and supervised learning. These models take into consideration noise emerging from different forms of oracle feedback, including demonstrations and corrections. We evaluate our approach across two domains and demonstrate that it achieves higher predictive performance than baseline methods, and also that the learned model can be used to selectively query an oracle at execution time to prevent errors. We also empirically analyze the biases of various feedback types and how these biases influence the discovery of blind spots. Further, we include analyses of our approach that incorporate relaxed initial optimality assumptions. (Interestingly, relaxing the assumptions of an optimal oracle and an optimal simulator policy helped our models to perform better.) We also propose extensions to our method that are intended to improve performance when using corrections and demonstrations data.

AI systems are being deployed to support human decision making in high-stakes domains such as healthcare and criminal justice. In many cases, the human and AI form a team, in which the human makes decisions after reviewing the AI’s inferences. A successful partnership requires that the human develops insights into the performance of the AI system, including its failures. We study the influence of updates to an AI system in this setting. While updates can increase the AI’s predictive performance, they may also lead to behavioral changes that are at odds with the user’s prior experiences and confidence in the AI’s inferences. We show that updates that increase AI performance may actually hurt team performance. We introduce the notion of the compatibility of an AI update with prior user experience and present methods for studying the role of compatibility in human-AI teams. Empirical results on three high-stakes classification tasks show that current machine learning algorithms do not produce compatible updates. We propose a re-training objective to improve the compatibility of an update by penalizing new errors. The objective offers full leverage of the performance/compatibility tradeoff across different datasets, enabling more compatible yet accurate updates.

Simulators are being increasingly used to train agents before deploying them in real-world environments. While training in simulation provides a cost-effective way to learn, poorly modeled aspects of the simulator can lead to costly mistakes, or blind spots. While humans can help guide an agent towards identifying these error regions, humans themselves have blind spots and noise in execution. We study how learning about blind spots of both can be used to manage hand-off decisions when humans and agents jointly act in the real-world in which neither of them are trained or evaluated fully. The formulation assumes that agent blind spots result from representational limitations in the simulation world, which leads the agent to ignore important features that are relevant for acting in the open world. Our approach for blind spot discovery combines experiences collected in simulation with limited human demonstrations. The first step applies imitation learning to demonstration data to identify important features that the human is using but that the agent is missing. The second step uses noisy labels extracted from action mismatches between the agent and the human across simulation and demonstration data to train blind spot models. We show through experiments on two domains that our approach is able to learn a succinct representation that accurately captures blind spot regions and avoids dangerous errors in the real world through transfer of control between the agent and the human.

As predictive models increasingly assist human experts (e.g., doctors) in day-to-day decision making, it is crucial for experts to be able to explore and understand how such models behave in different feature subspaces in order to know if and when to trust them. To this end, we propose Model Understanding through Subspace Explanations (MUSE), a novel model agnostic framework which facilitates understanding of a given black box model by explaining how it behaves in subspaces characterized by certain features of interest. Our framework provides end users (e.g., doctors) with the flexibility of customizing the model explanations by allowing them to input the features of interest. The construction of explanations is guided by a novel objective function that we propose to simultaneously optimize for fidelity to the original model, unambiguity and interpretability of the explanation. More specifically, our objective allows us to learn, with optimality guarantees, a small number of compact decision sets each of which captures the behavior of a given black box model in unambiguous, well-defined regions of the feature space. Experimental evaluation with real-world datasets and user studies demonstrate that our approach can generate customizable, highly compact, easy-to-understand, yet accurate explanations of various kinds of predictive models compared to state-of-the-art baselines.

Decisions made by human-AI teams (e.g., AI-advised humans) are increasingly common in high-stakes domains such as healthcare, criminal justice, and finance. Achieving high team performance depends on more than just the accuracy of the AI system: Since the human and the AI may have different expertise, the highest team performance is often reached when they both know how and when to complement one another. We focus on a factor that is crucial to supporting such complementary: the human’s mental model of the AI capabilities, specifically the AI system’s error boundary (i.e. knowing “When does the AI err?”). Awareness of this lets the human decide when to accept or override the AI’s recommendation. We highlight two key properties of an AI’s error boundary, parsimony and stochasticity, and a property of the task, dimensionality. We show experimentally how these properties affect humans’ mental models of AI capabilities and the resulting team performance. We connect our evaluations to related work and propose goals, beyond accuracy, that merit consideration during model selection and optimization to improve overall human-AI team performance.

In many online environments, such as massive open online courses and crowdsourcing platforms, many people solve similar complex tasks. As a byproduct of solving these tasks, a pool of artifacts are created that may be able to help others perform better on similar tasks. In this paper, we explore whether work that is naturally done by crowdworkers can be used as examples to help future crowdworkers perform better on similar tasks. We explore this in the context of a product comparison review task, where workers must compare and contrast pairs of similar products. We first show that randomly presenting one or two peer-generated examples does not significantly improve performance on future tasks. In a second experiment, we show that presenting examples that are of sufficiently high quality leads to a statistically significant improvement in performance of future workers on a near transfer task. Moreover, our results suggest that even among high quality examples, there are differences in how effective the examples are, indicating that quality is not a perfect proxy for pedagogical value.

Although systematic biases in decision-making are widely documented, the ways in which they emerge from different sources is less understood. We present a controlled experimental platform to study gender bias in hiring by decoupling the effect of world distribution (the gender breakdown of candidates in a specific profession) from bias in human decision-making. We explore the effectiveness of representation criteria, fixed proportional display of candidates, as an intervention strategy for mitigation of gender bias by conducting experiments measuring human decision-makers’ rankings for who they would recommend as potential hires. Experiments across professions with varying gender proportions show that balancing gender representation in candidate slates can correct biases for some professions where the world distribution is skewed, although doing so has no impact on other professions where human persistent preferences are at play. We show that the gender of the decision-maker, complexity of the decision-making task and over- and under-representation of genders in the candidate slate can all impact the final decision. By decoupling sources of bias, we can better isolate strategies for bias mitigation in human-in-the-loop systems.

People receive dozens, or hundreds, of notifications per day and each notification poses some risk of accidental information disclosure in the presence of others; onlookers may see notifications on a mobile phone lock screen, on the periphery of a desktop or laptop display. We quantify the prevalence of these accidental disclosures in the context of email notifications, and we study people's relevant preferences and concerns. Our results are compiled from a retrospective survey of 131 respondents, and a contextual-labeling study where 169 participants labeled 1,040 meeting-email pairs. We find that, for 53% of people, at least 1 in 10 email notifications poses an information disclosure risk, and the real or perceived severity of these risks depend both on user characteristics and the meeting or email attributes. We conclude by exploring machine learning for predicting people's comfort levels, and we present implications for the design of future social-context aware notification systems.

Volunteers who help with online crowdsourcing such as citizen science tasks typically make only a few contributions before exiting. We propose a computational approach for increasing users' engagement in such settings that is based on optimizing policies for displaying motivational messages to users. The approach, which we refer to as Trajectory Corrected Intervention (TCI), reasons about the tradeoff between the long-term influence of engagement messages on participants' contributions and the potential risk of disrupting their current work. We combine model-based reinforcement learning with off-line policy evaluation to generate intervention policies, without relying on a fixed representation of the domain. TCI works iteratively to learn the best representation from a set of random intervention trials and to generate candidate intervention policies. It is able to refine selected policies off-line by exploiting the fact that users can only be interrupted once per session.We implemented TCI in the wild with Galaxy Zoo, one of the largest citizen science platforms on the web. We found that TCI was able to outperform the state-of-the-art intervention policy for this domain, and significantly increased the contributions of thousands of users. This work demonstrates the benefit of combining traditional AI planning with off-line policy methods to generate intelligent intervention strategies.

As machine learning systems move from computer-science laboratories into the open world, their accountability becomes a high priority problem. Accountability requires deep understanding of system behavior and its failures. Current evaluation methods such as single-score error metrics and confusion matrices provide aggregate views of system performance that hide important shortcomings. Understanding details about failures is important for identifying pathways for refinement, communicating the reliability of systems in different settings, and for specifying appropriate human oversight and engagement. Characterization of failures and shortcomings is particularly complex for systems composed of multiple machine learned components. For such systems, existing evaluation methods have limited expressiveness in describing and explaining the relationship among input content, the internal states of system components, and final output quality. We present Pandora, a set of hybrid human-machine methods and tools for describing and explaining system failures. Pandora leverages both human and system-generated observations to summarize conditions of system malfunction with respect to the input content and system architecture. We share results of a case study with a machine learning pipeline for image captioning that show how detailed performance views can be beneficial for analysis and debugging.

We study how real-time crowdsourcing can be used both for evaluating the value provided by existing automated approaches and for enabling workflows that provide scalable and useful alt text to blind users. We show that the shortcomings of existing AI image captioning systems frequently hinder a user's understanding of an image they cannot see to a degree that even clarifying conversations with sighted assistants cannot correct. Based on analysis of clarifying conversations collected from our studies, we design experiences that can effectively assist users in a scalable way without the need for real-time interaction. Our results provide lessons and guidelines that the designers of future AI captioning systems can use to improve labeling of social media imagery for blind users.

Predictive models deployed in the real world may assign incorrect labels to instances with high confidence. Such errors or unknown unknowns are rooted in model incompleteness, and typically arise because of the mismatch between training data and the cases encountered at test time. As the models are blind to such errors, input from an oracle is needed to identify these failures. In this paper, we formulate and address the problem of informed discovery of unknown unknowns of any given predictive model where unknown unknowns occur due to systematic biases in the training data.We propose a model-agnostic methodology which uses feedback from an oracle to both identify unknown unknowns and to intelligently guide the discovery. We employ a two-phase approach which first organizes the data into multiple partitions based on the feature similarity of instances and the confidence scores assigned by the predictive model, and then utilizes an explore-exploit strategy for discovering unknown unknowns across these partitions. We demonstrate the efficacy of our framework by varying the underlying causes of unknown unknowns across various applications. To the best of our knowledge, this paper presents the first algorithmic approach to the problem of discovering unknown unknowns of predictive models.

We study the problem of troubleshooting machine learning systems that rely on analytical pipelines of distinct components. Understanding and fixing errors that arise in such integrative systems is difficult as failures can occur at multiple points in the execution workflow. Moreover, errors can propagate, become amplified or be suppressed, making blame assignment difficult. We propose a human-in-the-loop methodology which leverages human intellect for troubleshooting system failures. The approach simulates potential component fixes through human computation tasks and measures the expected improvements in the holistic behavior of the system. The method provides guidance to designers about how they can best improve the system. We demonstrate the effectiveness of the approach on an automated image captioning system that has been pressed into real-world use.

Crowdsourcing provides a scalable and efficient way to construct labeled datasets for training machine learning systems. However, creating comprehensive label guidelines for crowdworkers is often prohibitive even for seemingly simple concepts. Incomplete or ambiguous label guidelines can then result in differing interpretations of concepts and inconsistent labels. Existing approaches for improving label quality, such as worker screening or detection of poor work, are ineffective for this problem and can lead to rejection of honest work and a missed opportunity to capture rich interpretations about data. We introduce Revolt, a collaborative approach that brings ideas from expert annotation workflows to crowd-based labeling. Revolt eliminates the burden of creating detailed label guidelines by harnessing crowd disagreements to identify ambiguous concepts and create rich structures (groups of semantically related items) for post-hoc label decisions. Experiments comparing Revolt to traditional crowdsourced labeling show that Revolt produces high quality labels without requiring label guidelines in turn for an increase in monetary cost. This up front cost, however, is mitigated by Revolt's ability to produce reusable structures that can accommodate a variety of label boundaries without requiring new data to be collected. Further comparisons of Revolt's collaborative and non-collaborative variants show that collaboration reaches higher label accuracy with lower monetary cost.

Crowd work is typically limited to simple, context-free tasks because they are easy to describe and understand. In contrast, complex tasks require communication between the requester and workers to achieve mutual understanding, which can be more work than it is worth. This paper explores the notion of structured communication: using structured microtasks to support communication in the domain of complex writing. Our studies compare a variety of communication mechanisms with respect to the costs to the requester in providing information and the value of that information to workers while performing the task. We find that different mechanisms are effective at different stages of writing. For early drafts, asking the requester to state the biggest problem in the current write-up is valuable and low cost, while later it is more useful for the worker if the requester highlights the text that needs to be improved. These findings can be used to enable richer, more interactive crowd work than what currently seems possible. We incorporate the findings in a workflow for crowdsourcing written content using appropriately timed mechanisms for communicating with the crowd.

Crowd-powered systems leverage human intelligence to go beyond the capabilities of automated systems, but also introduce privacy and security concerns because unknown people must view the data that the system processes. While automated approaches cannot robustly filter private information from these datasets, people have the ability to do so if the risk from them viewing the data can be mitigated. We present a crowd-powered approach to masking private content in data by segmenting and distributing smaller segments to crowd workers so that individual workers can identify potentially private content without being able to fully view it themselves. We introduce a novel pyramid workflow for segmentation that uses segments at multiple levels of granularity to overcome problems with fixed-sized approaches. We implement our approach in CrowdMask, a system that allows images with potentially sensitive content to be masked by appearing in progressively larger, more identifiable segments, and masking portions of the image as soon as a risk is identified. Our experiments with 4134 Mechanical Turk workers show that CrowdMask can effectively mask private content from images without revealing sensitive content to constituent workers, while still enabling future systems to use the filtered result.

The access of visually impaired users to imagery in social media is constrained by the availability of suitable alt text. It is unknown how imperfections in emerging tools for automatic caption generation may help or hinder blind users' understanding of social media posts with embedded imagery. In this paper, we study how crowdsourcing can be used both for evaluating the value provided by existing automated approaches and for enabling workflows that provide scalable and useful alt text to blind users. Using real-time crowdsourcing, we designed experiences that varied the depth of interaction of the crowd in assisting visually impaired users at caption interpretation, and measured trade-offs in effectiveness, scalability, and reusability. We show that the shortcomings of existing AI image captioning systems frequently hinder a user's understanding of an image they cannot see to a degree that even clarifying conversations with sighted assistants cannot correct. Our detailed analysis of the set of clarifying conversations collected from our studies led to the design of experiences that can effectively assist users in a scalable way without the need for real-time interaction. They also provide lessons and guidelines that human captioners and the designers of future iterations of AI captioning systems can use to improve labeling of social media imagery for blind users.

We explore how crowdworkers can be trained to tackle complex crowdsourcing tasks. We are particularly interested in training novice workers to perform well on solving tasks in situations where the space of strategies is large and workers need to discover and try different strategies to be successful. In a first experiment, we perform a comparison of five different training strategies. For complex web search challenges, we show that providing expert examples is an effective form of training, surpassing other forms of training in nearly all measures of interest. However, such training relies on access to domain expertise, which may be expensive or lacking. Therefore, in a second experiment we study the feasibility of training workers in the absence of domain expertise. We show that having workers validate the work of their peer workers can be even more effective than having them review expert examples if we only present solutions filtered by a threshold length. The results suggest that crowdsourced solutions of peer workers may be harnessed in an automated training pipeline.

Models for aggregating contributions by crowd workers have been shown to be challenged by the rise of task-specific biases or errors. Task-dependent errors in assessment may shift the majority opinion of even large numbers of workers to an incorrect answer. We introduce and evaluate probabilistic models that can detect and correct task-dependent bias automatically. First, we show how to build and use probabilistic graphical models for jointly modeling task features, workers' biases, worker contributions and ground truth answers of tasks so that task-dependent bias can be corrected. Second, we show how the approach can perform a type of transfer learning among workers to address the issue of annotation sparsity. We evaluate the models with varying complexity on a large data set collected from a citizen science project and show that the models are effective at correcting the task-dependent worker bias. Finally, we investigate the use of active learning to guide the acquisition of expert assessments to enable automatic detection and correction of worker bias.

We exploit the absence of signals as informative observations in the context of providing task recommendations in crowdsourcing. Workers on crowdsourcing platforms do not provide explicit ratings about tasks. We present methods that enable a system to leverage implicit signals about task preferences. These signals include types of tasks that have been available and have been displayed, and the number of tasks workers select and complete. In contrast to previous work, we present a general model that can represent both positive and negative implicit signals. We introduce algorithms that can learn these models without exceeding the computational complexity of existing approaches. Finally, using data from a high-throughput crowdsourcing platform, we show that reasoning about both positive and negative implicit feedback can improve the quality of task recommendations.

Online services such as web search and e-commerce applications typically rely on the collection of data about users, including details of their activities on the web. Such personal data is used to maximize revenues via targeting of advertisements and longer engagements of users, and to enhance the quality of service via personalization of content. To date, service providers have largely followed the approach of either requiring or requesting consent for collecting user data. Users may be willing to share private information in return for incentives, enhanced services, or assurances about the nature and extent of the logged data. We introduce stochastic privacy, an approach to privacy centering on the simple concept of providing people with a guarantee that the probability that their personal data will be shared does not exceed a given bound. Such a probability, which we refer to as the privacy risk, can be given by users as a preference or communicated as a policy by a service provider. Service providers can work to personalize and to optimize revenues in accordance with preferences about privacy risk. We present procedures, proofs, and an overall system for maximizing the quality of services, while respecting bounds on privacy risk. We demonstrate the methodology with a case study and evaluation of the procedures applied to web search personalization. We show how we can achieve near-optimal utility of accessing information with provable guarantees on the probability of sharing data.

Crowd-powered systems have become a popular way to augment the capabilities of automated systems in real-world settings. Many of these systems rely on human workers to process potentially sensitive data or make important decisions. This puts these systems at risk of unintentionally releasing sensitive data or having their outcomes maliciously manipulated. While almost all crowd-powered approaches account for errors made by individual workers, few factor in active attacks on the system. In this paper, we analyze different forms of threats from individuals and groups of workers extracting information from crowd-powered systems or manipulating these systems' outcomes. Via a set of studies performed on Amazon's Mechanical Turk platform and involving 1,140 unique workers, we demonstrate the viability of these threats. We show that the current system is vulnerable to coordinated attacks on a task based on the requests of another task and that a significant portion of Mechanical Turk workers are willing to contribute to an attack. We propose several possible approaches to mitigating these threats, including leveraging workers who are willing to go above and beyond to help, automatically flagging sensitive content, and using workflows that conceal information from each individual, while still allowing the group to complete a task. Our findings enable the crowd to continue to play an important part in automated systems, even as the data they use and the decisions they support become increasingly important.

We explore the use of crowdsourcing to generate natural language in spoken dia-logue systems. We introduce a method-ology to elicit novel templates from the crowd based on a dialogue seed corpus, and investigate the effect that the amount of surrounding dialogue context has on the generation task. Evaluation is performed both with a crowd and with a system de-veloper to assess the naturalness and suit-ability of the elicited phrases. Results indi-cate that the crowd is able to provide rea-sonable and diverse templates within this methodology. More work is necessary be-fore elicited templates can be automati-cally plugged into the system. 1

Paid and volunteer crowd work have emerged as a means for harnessing human intelligence for performing diverse tasks. However, little is known about the relative performance of volunteer versus paid crowd work, and how financial incentives influence the quality and efficiency of output. We study the performance of volunteers as well as workers paid with different monetary schemes on a difficult real-world crowdsourcing task. We observe that performance by unpaid and paid workers can be compared in carefully designed tasks, that financial incentives can be used to trade quality for speed, and that the compensation system on Amazon Mechanical Turk creates particular indirect incentives for workers. Our methodology and results have implications for the ideal choice of financial incentives and motivates further study on how monetary incentives influence worker behavior in crowdsourcing.

We present studies of the attention and time, or engagement, invested by crowd workers on tasks. Consideration of worker engagement is especially important in volunteer settings such as online citizen science. Using data from Galaxy Zoo, a prominent citizen science project, we design and construct statistical models that provide predictions about the forthcoming engagement of volunteers. We characterize the accuracy of predictions with respect to different sets of features that describe user behavior and study the sensitivity of predictions to variations in the amount of data and retraining. We design our model for guiding system actions in real-time settings, and discuss the prospect for harnessing predictive models of engagement to enhance user attention and effort on volunteer tasks.

We study the opportunity for using crowdsourcing methods to acquire language corpora for use in natural language processing systems. Specifically, we empirically investigate three methods for eliciting natural language sentences that correspond to a given semantic form. The methods convey frame semantics to crowd workers by means of sentences, scenarios, and list-based descriptions. We discuss various performance measures of the crowdsourcing process, and analyze the semantic correctness, naturalness, and biases of the collected language. We highlight research challenges and directions in applying these methods to acquire corpora for natural language processing applications.