Vision models often fail systematically on groups of data that share common semantic characteristics (e.g., rare objects or unusual scenes), but identifying these failure modes is a challenge. We introduce AdaVision, an interactive process for testing vision models which helps users identify and fix coherent failure modes. Given a natural language description of a coherent group, AdaVision retrieves relevant images from LAION-5B with CLIP. The user then labels a small amount of data for model correctness, which is used in successive retrieval rounds to hill-climb towards high-error regions, refining the group definition. Once a group is saturated, AdaVision uses GPT-3 to suggest new group descriptions for the user to explore. We demonstrate the usefulness and generality of AdaVision in user studies, where users find major bugs in state-of-the-art classification, object detection, and image captioning models. These user-discovered groups have failure rates 2-3x higher than those surfaced by automatic error clustering methods. Finally, finetuning on examples found with AdaVision fixes the discovered bugs when evaluated on unseen examples, without degrading in-distribution accuracy, and while also improving performance on out-of-distribution datasets.

Current approaches for fixing systematic problems in NLP models (e.g. regex patches, finetuning on more data) are either brittle, or labor-intensive and liable to shortcuts. In contrast, humans often provide corrections to each other through natural language. Taking inspiration from this, we explore natural language patches -- declarative statements that allow developers to provide corrective feedback at the right level of abstraction, either overriding the model (``if a review gives 2 stars, the sentiment is negative'') or providing additional information the model may lack (``if something is described as the bomb, then it is good''). We model the task of determining if a patch applies separately from the task of integrating patch information, and show that with a small amount of synthetic data, we can teach models to effectively use real patches on real data -- 1 to 7 patches improve accuracy by ~1-4 accuracy points on different slices of a sentiment analysis dataset, and F1 by 7 points on a relation extraction dataset. Finally, we show that finetuning on as many as 100 labeled examples may be needed to match the performance of a small set of language patches.

基于Shapley值的功能归因在解释机器学习模型中很受欢迎。但是，从理论和计算的角度来看，它们的估计是复杂的。我们将这种复杂性分解为两个因素：（1）〜删除特征信息的方法，以及（2）〜可拖动估计策略。这两个因素提供了一种天然镜头，我们可以更好地理解和比较24种不同的算法。基于各种特征删除方法，我们描述了多种类型的Shapley值特征属性和计算每个类型的方法。然后，基于可进行的估计策略，我们表征了两个不同的方法家族：模型 - 不合时宜的和模型特定的近似值。对于模型 - 不合稳定的近似值，我们基准了广泛的估计方法，并将其与Shapley值的替代性但等效的特征联系起来。对于特定于模型的近似值，我们阐明了对每种方法的线性，树和深模型的障碍至关重要的假设。最后，我们确定了文献中的差距以及有希望的未来研究方向。

研究人员提出了多种模型解释方法，但目前尚不清楚大多数方法如何相关或何时一种方法比另一种方法更可取。我们研究了文献，发现许多方法都是基于通过删除来解释的共同原理 - 本质上是测量从模型中删除一组特征的影响。这些方法在几个方面有所不同，因此我们为基于删除的解释开发了一个沿三个维度表征每个方法的框架：1）该方法如何删除特征，2）该方法解释的模型行为以及3）方法如何汇总每个方法功能的影响。我们的框架统一了26种现有方法，其中包括几种最广泛使用的方法（Shap，Lime，有意义的扰动，排列测试）。揭露这些方法之间的基本相似性使用户能够推荐使用哪种工具，并为正在进行的模型解释性研究提出了有希望的方向。

Understanding why a model makes a certain prediction can be as crucial as the prediction's accuracy in many applications. However, the highest accuracy for large modern datasets is often achieved by complex models that even experts struggle to interpret, such as ensemble or deep learning models, creating a tension between accuracy and interpretability. In response, various methods have recently been proposed to help users interpret the predictions of complex models, but it is often unclear how these methods are related and when one method is preferable over another. To address this problem, we present a unified framework for interpreting predictions, SHAP (SHapley Additive exPlanations). SHAP assigns each feature an importance value for a particular prediction. Its novel components include: (1) the identification of a new class of additive feature importance measures, and (2) theoretical results showing there is a unique solution in this class with a set of desirable properties. The new class unifies six existing methods, notable because several recent methods in the class lack the proposed desirable properties. Based on insights from this unification, we present new methods that show improved computational performance and/or better consistency with human intuition than previous approaches.

Data compression is becoming critical for storing scientific data because many scientific applications need to store large amounts of data and post process this data for scientific discovery. Unlike image and video compression algorithms that limit errors to primary data, scientists require compression techniques that accurately preserve derived quantities of interest (QoIs). This paper presents a physics-informed compression technique implemented as an end-to-end, scalable, GPU-based pipeline for data compression that addresses this requirement. Our hybrid compression technique combines machine learning techniques and standard compression methods. Specifically, we combine an autoencoder, an error-bounded lossy compressor to provide guarantees on raw data error, and a constraint satisfaction post-processing step to preserve the QoIs within a minimal error (generally less than floating point error). The effectiveness of the data compression pipeline is demonstrated by compressing nuclear fusion simulation data generated by a large-scale fusion code, XGC, which produces hundreds of terabytes of data in a single day. Our approach works within the ADIOS framework and results in compression by a factor of more than 150 while requiring only a few percent of the computational resources necessary for generating the data, making the overall approach highly effective for practical scenarios.

Algorithms that involve both forecasting and optimization are at the core of solutions to many difficult real-world problems, such as in supply chains (inventory optimization), traffic, and in the transition towards carbon-free energy generation in battery/load/production scheduling in sustainable energy systems. Typically, in these scenarios we want to solve an optimization problem that depends on unknown future values, which therefore need to be forecast. As both forecasting and optimization are difficult problems in their own right, relatively few research has been done in this area. This paper presents the findings of the ``IEEE-CIS Technical Challenge on Predict+Optimize for Renewable Energy Scheduling," held in 2021. We present a comparison and evaluation of the seven highest-ranked solutions in the competition, to provide researchers with a benchmark problem and to establish the state of the art for this benchmark, with the aim to foster and facilitate research in this area. The competition used data from the Monash Microgrid, as well as weather data and energy market data. It then focused on two main challenges: forecasting renewable energy production and demand, and obtaining an optimal schedule for the activities (lectures) and on-site batteries that lead to the lowest cost of energy. The most accurate forecasts were obtained by gradient-boosted tree and random forest models, and optimization was mostly performed using mixed integer linear and quadratic programming. The winning method predicted different scenarios and optimized over all scenarios jointly using a sample average approximation method.

Lack of factual correctness is an issue that still plagues state-of-the-art summarization systems despite their impressive progress on generating seemingly fluent summaries. In this paper, we show that factual inconsistency can be caused by irrelevant parts of the input text, which act as confounders. To that end, we leverage information-theoretic measures of causal effects to quantify the amount of confounding and precisely quantify how they affect the summarization performance. Based on insights derived from our theoretical results, we design a simple multi-task model to control such confounding by leveraging human-annotated relevant sentences when available. Crucially, we give a principled characterization of data distributions where such confounding can be large thereby necessitating the use of human annotated relevant sentences to generate factual summaries. Our approach improves faithfulness scores by 20\% over strong baselines on AnswerSumm \citep{fabbri2021answersumm}, a conversation summarization dataset where lack of faithfulness is a significant issue due to the subjective nature of the task. Our best method achieves the highest faithfulness score while also achieving state-of-the-art results on standard metrics like ROUGE and METEOR. We corroborate these improvements through human evaluation.

As language models (LMs) scale, they develop many novel behaviors, good and bad, exacerbating the need to evaluate how they behave. Prior work creates evaluations with crowdwork (which is time-consuming and expensive) or existing data sources (which are not always available). Here, we automatically generate evaluations with LMs. We explore approaches with varying amounts of human effort, from instructing LMs to write yes/no questions to making complex Winogender schemas with multiple stages of LM-based generation and filtering. Crowdworkers rate the examples as highly relevant and agree with 90-100% of labels, sometimes more so than corresponding human-written datasets. We generate 154 datasets and discover new cases of inverse scaling where LMs get worse with size. Larger LMs repeat back a dialog user's preferred answer ("sycophancy") and express greater desire to pursue concerning goals like resource acquisition and goal preservation. We also find some of the first examples of inverse scaling in RL from Human Feedback (RLHF), where more RLHF makes LMs worse. For example, RLHF makes LMs express stronger political views (on gun rights and immigration) and a greater desire to avoid shut down. Overall, LM-written evaluations are high-quality and let us quickly discover many novel LM behaviors.

This paper describes Waymo's Collision Avoidance Testing (CAT) methodology: a scenario-based testing method that evaluates the safety of the Waymo Driver Automated Driving Systems' (ADS) intended functionality in conflict situations initiated by other road users that require urgent evasive maneuvers. Because SAE Level 4 ADS are responsible for the dynamic driving task (DDT), when engaged, without immediate human intervention, evaluating a Level 4 ADS using scenario-based testing is difficult due to the potentially infinite number of operational scenarios in which hazardous situations may unfold. To that end, in this paper we first describe the safety test objectives for the CAT methodology, including the collision and serious injury metrics and the reference behavior model representing a non-impaired eyes on conflict human driver used to form an acceptance criterion. Afterward, we introduce the process for identifying potentially hazardous situations from a combination of human data, ADS testing data, and expert knowledge about the product design and associated Operational Design Domain (ODD). The test allocation and execution strategy is presented next, which exclusively utilize simulations constructed from sensor data collected on a test track, real-world driving, or from simulated sensor data. The paper concludes with the presentation of results from applying CAT to the fully autonomous ride-hailing service that Waymo operates in San Francisco, California and Phoenix, Arizona. The iterative nature of scenario identification, combined with over ten years of experience of on-road testing, results in a scenario database that converges to a representative set of responder role scenarios for a given ODD. Using Waymo's virtual test platform, which is calibrated to data collected as part of many years of ADS development, the CAT methodology provides a robust and scalable safety evaluation.