The performance of deep neural networks is strongly influenced by the quality of their training data. However, mitigating dataset bias by manually curating challenging edge cases remains a major bottleneck. To address this, we propose an automated pipeline for text-guided edge-case synthesis. Our approach employs a Large Language Model, fine-tuned via preference learning, to rephrase image captions into diverse textual prompts that steer a Text-to-Image model toward generating difficult visual scenarios. Evaluated on the FishEye8K object detection benchmark, our method achieves superior robustness, surpassing both naive augmentation and manually engineered prompts. This work establishes a scalable framework that shifts data curation from manual effort to automated, targeted synthesis, offering a promising direction for developing more reliable and continuously improving AI systems. Code is available at https://github.com/gokyeongryeol/ATES.

Pairwise preferences over model responses are widely collected to evaluate and provide feedback to large language models (LLMs). Given two alternative model responses to the same input, a human or AI annotator selects the "better" response. This approach can provide feedback for domains where other hard-coded metrics are difficult to obtain (e.g., chat response quality), thereby helping model evaluation or training. However, for some domains high-quality pairwise comparisons can be tricky to obtain - from AI and humans. For example, for responses with many factual statements, annotators may disproportionately weigh writing quality rather than underlying facts. In this work, we explore augmenting standard AI annotator systems with additional tools to improve performance on three challenging response domains: long-form factual, math and code tasks. We propose a tool-using agentic system to provide higher quality feedback on these domains. Our system uses web-search and code execution to ground itself based on external validation, independent of the LLM's internal knowledge and biases. We provide extensive experimental results evaluating our method across the three targeted response domains as well as general annotation tasks, using RewardBench (incl. AlpacaEval and LLMBar), RewardMath, as well as three new datasets for domains with saturated pre-existing datasets. Our results indicate that external tools can indeed improve performance in many, but not all, cases. More generally, our experiments highlight the sensitivity of performance to simple parameters (e.g., prompt) and the need for improved (non-saturated) annotator benchmarks. We share our code at https://github.com/apple/ml-agent-evaluator.

Street scene datasets, collected from Street View or dashboard cameras, offer a promising means of detecting urban objects and incidents like street flooding. However, a major challenge in using these datasets is their lack of reliable labels: there are myriad types of incidents, many types occur rarely, and ground-truth measures of where incidents occur are lacking. Here, we propose BayFlood, a two-stage approach which circumvents this difficulty. First, we perform zero-shot classification of where incidents occur using a pretrained vision-language model (VLM). Second, we fit a spatial Bayesian model on the VLM classifications. The zero-shot approach avoids the need to annotate large training sets, and the Bayesian model provides frequent desiderata in urban settings - principled measures of uncertainty, smoothing across locations, and incorporation of external data like stormwater accumulation zones. We comprehensively validate this two-stage approach, showing that VLMs provide strong zero-shot signal for floods across multiple cities and time periods, the Bayesian model improves out-of-sample prediction relative to baseline methods, and our inferred flood risk correlates with known external predictors of risk. Having validated our approach, we show it can be used to improve urban flood detection: our analysis reveals 113,738 people who are at high risk of flooding overlooked by current methods, identifies demographic biases in existing methods, and suggests locations for new flood sensors. More broadly, our results showcase how Bayesian modeling of zero-shot LM annotations represents a promising paradigm because it avoids the need to collect large labeled datasets and leverages the power of foundation models while providing the expressiveness and uncertainty quantification of Bayesian models.

Curating annotations for medical image segmentation is a labor-intensive and time-consuming task that requires domain expertise, resulting in "narrowly" focused deep learning (DL) models with limited translational utility. Recently, foundation models like the Segment Anything Model (SAM) have revolutionized semantic segmentation with exceptional zero-shot generalizability across various domains, including medical imaging, and hold a lot of promise for streamlining the annotation process. However, SAM has yet to be evaluated in a crowd-sourced setting to curate annotations for training 3D DL segmentation models. In this work, we explore the potential of SAM for crowd-sourcing "sparse" annotations from non-experts to generate "dense" segmentation masks for training 3D nnU-Net models, a state-of-the-art DL segmentation model. Our results indicate that while SAM-generated annotations exhibit high mean Dice scores compared to ground-truth annotations, nnU-Net models trained on SAM-generated annotations perform significantly worse than nnU-Net models trained on ground-truth annotations ($p<0.001$, all).

A clinical artificial intelligence (AI) system is often validated on a held-out set of data which it has not been exposed to before (e.g., data from a different hospital with a distinct electronic health record system). This evaluation process is meant to mimic the deployment of an AI system on data in the wild; those which are currently unseen by the system yet are expected to be encountered in a clinical setting. However, when data in the wild differ from the held-out set of data, a phenomenon referred to as distribution shift, and lack ground-truth annotations, it becomes unclear the extent to which AI-based findings can be trusted on data in the wild. Here, we introduce SUDO, a framework for evaluating AI systems without ground-truth annotations. SUDO assigns temporary labels to data points in the wild and directly uses them to train distinct models, with the highest performing model indicative of the most likely label. Through experiments with AI systems developed for dermatology images, histopathology patches, and clinical reports, we show that SUDO can be a reliable proxy for model performance and thus identify unreliable predictions. We also demonstrate that SUDO informs the selection of models and allows for the previously out-of-reach assessment of algorithmic bias for data in the wild without ground-truth annotations. The ability to triage unreliable predictions for further inspection and assess the algorithmic bias of AI systems can improve the integrity of research findings and contribute to the deployment of ethical AI systems in medicine.

Neural networks (NNs) can learn to rely on spurious signals in the training data, leading to poor generalisation. Recent methods tackle this problem by training NNs with additional ground-truth annotations of such signals. These methods may, however, let spurious signals re-emerge in deep convolutional NNs (CNNs). We propose Targeted Activation Penalty (TAP), a new method tackling the same problem by penalising activations to control the re-emergence of spurious signals in deep CNNs, while also lowering training times and memory usage. In addition, ground-truth annotations can be expensive to obtain. We show that TAP still works well with annotations generated by pre-trained models as effective substitutes of ground-truth annotations. We demonstrate the power of TAP against two state-of-the-art baselines on the MNIST benchmark and on two clinical image datasets, using four different CNN architectures.

The hands are the primary means by which humans interact with the outside world. As such, accurate hand pose estimation is a necessary requirement for many vision-based systems and enables many applications in areas such as augmented reality (AR), virtual reality (VR) and gesture recognition. Recently, the availability of more accurate and affordable commodity depth cameras coupled with the success of Deep Neural Networks (DNN) has led to significant progress in depth-based 3D hand pose estimation and segmentation [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]. Despite these advancements, one major challenge that remains is that DNN-based methods require large amounts of annotated training data to realize their full potential. A straightforward approach to mitigate this requirement is to use synthesized training data with accurate annotations [11], which can be generated with minimal human effort. However, models trained on synthesized data generalize poorly to the real-world data due to the significant domain gap between synthetic and real-world data. A popular alternative is semi-supervised learning (SSL) [12], where the goal is to leverage unlabeled data along with the labeled data, hence reducing the amount of labeled data required for training. Most of the recent advancements of SSL methods have been focused on image classification [13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23].

Accurate, real-time segmentation of vessel structures in ultrasound image sequences can aid in the measurement of lumen diameters and assessment of vascular diseases. This, however, remains a challenging task, particularly for extremely small vessels that are difficult to visualize. We propose to leverage the rich spatiotemporal context available in ultrasound to improve segmentation of small-scale lower-extremity arterial vasculature. We describe efficient deep learning methods that incorporate temporal, spatial, and feature-aware contextual embeddings at multiple resolution scales while jointly utilizing information from B-mode and Color Doppler signals. Evaluating on femoral and tibial artery scans performed on healthy subjects by an expert ultrasonographer, and comparing to consensus expert ground-truth annotations of inner lumen boundaries, we demonstrate real-time segmentation using the context-aware models and show that they significantly outperform comparable baseline approaches.

We consider the problem of predicting to which classes an image belongs, where the number of classes is large (many thousands or tens of thousands) and where each image typically belongs to multiple classes that should all be properly identified: multi-label, massively multi-class classification. In such classification problems, the best practice until now (for instance in use at Google, Inc.) has been to use a deep convolutional neural network such as the ones described in [19] or [18], culminating in a logistic regression layer with a sigmoid cross-entropy loss, with target labels encoded as high-dimensional sparse binary vectors. The use of logistic regression implies an important yet oft overlooked assumption made about the label space: the classes are considered to be statistically independent, each class being treated as an independent dimension in the label space. This is generally not the case in practice: mirroring statistical dependencies found in the real world, label spaces often have a well-defined internal structure, with some labels being more likely to cooccur than other labels. For instance, "sky" and "beach" are frequently cooccurring labels, while "crane" and "manta ray" are rarely cooccurring. The sigmoid cross-entropy loss with sparse binary targets does not allow to leverage such observations about the structure of the label space. 1 There is therefore an opportunity to exploit the internal structure of the label space for gains in training speed, precision, and recall. One simple way to achieve this is to project the labels onto a lower-dimensional manifold -an embedding space-where a distance function between embedded labels would capture useful statistical dependencies. An appropriate loss function may then allow a parametric model trained via stochastic gradient descent to benefit from the structure of the manifold during training and inference.