Olga Russakovsky

Ziv Epstein, Aaron Hertzmann, Laura Herman et al.

A new class of tools, colloquially called generative AI, can produce high-quality artistic media for visual arts, concept art, music, fiction, literature, video, and animation. The generative capabilities of these tools are likely to fundamentally alter the creative processes by which creators formulate ideas and put them into production. As creativity is reimagined, so too may be many sectors of society. Understanding the impact of generative AI - and making policy decisions around it - requires new interdisciplinary scientific inquiry into culture, economics, law, algorithms, and the interaction of technology and creativity. We argue that generative AI is not the harbinger of art's demise, but rather is a new medium with its own distinct affordances. In this vein, we consider the impacts of this new medium on creators across four themes: aesthetics and culture, legal questions of ownership and credit, the future of creative work, and impacts on the contemporary media ecosystem. Across these themes, we highlight key research questions and directions to inform policy and beneficial uses of the technology.

Zhiwei Deng, Olga Russakovsky

We propose an algorithm that compresses the critical information of a large dataset into compact addressable memories. These memories can then be recalled to quickly re-train a neural network and recover the performance (instead of storing and re-training on the full original dataset). Building upon the dataset distillation framework, we make a key observation that a shared common representation allows for more efficient and effective distillation. Concretely, we learn a set of bases (aka ``memories'') which are shared between classes and combined through learned flexible addressing functions to generate a diverse set of training examples. This leads to several benefits: 1) the size of compressed data does not necessarily grow linearly with the number of classes; 2) an overall higher compression rate with more effective distillation is achieved; and 3) more generalized queries are allowed beyond recalling the original classes. We demonstrate state-of-the-art results on the dataset distillation task across six benchmarks, including up to 16.5% and 9.7% in retained accuracy improvement when distilling CIFAR10 and CIFAR100 respectively. We then leverage our framework to perform continual learning, achieving state-of-the-art results on four benchmarks, with 23.2% accuracy improvement on MANY. The code is released on our project webpage https://github.com/princetonvisualai/RememberThePast-DatasetDistillation.

Vikram V. Ramaswamy, Sunnie S. Y. Kim, Ruth Fong et al.

Concept-based interpretability methods aim to explain deep neural network model predictions using a predefined set of semantic concepts. These methods evaluate a trained model on a new, "probe" dataset and correlate model predictions with the visual concepts labeled in that dataset. Despite their popularity, they suffer from limitations that are not well-understood and articulated by the literature. In this work, we analyze three commonly overlooked factors in concept-based explanations. First, the choice of the probe dataset has a profound impact on the generated explanations. Our analysis reveals that different probe datasets may lead to very different explanations, and suggests that the explanations are not generalizable outside the probe dataset. Second, we find that concepts in the probe dataset are often less salient and harder to learn than the classes they claim to explain, calling into question the correctness of the explanations. We argue that only visually salient concepts should be used in concept-based explanations. Finally, while existing methods use hundreds or even thousands of concepts, our human studies reveal a much stricter upper bound of 32 concepts or less, beyond which the explanations are much less practically useful. We make suggestions for future development and analysis of concept-based interpretability methods. Code for our analysis and user interface can be found at \url{https://github.com/princetonvisualai/OverlookedFactors}

Ye Zhu, Yu Wu, Zhiwei Deng et al.

Applying pre-trained generative denoising diffusion models (DDMs) for downstream tasks such as image semantic editing usually requires either fine-tuning DDMs or learning auxiliary editing networks in the existing literature. In this work, we present our BoundaryDiffusion method for efficient, effective and light-weight semantic control with frozen pre-trained DDMs, without learning any extra networks. As one of the first learning-free diffusion editing works, we start by seeking a comprehensive understanding of the intermediate high-dimensional latent spaces by theoretically and empirically analyzing their probabilistic and geometric behaviors in the Markov chain. We then propose to further explore the critical step for editing in the denoising trajectory that characterizes the convergence of a pre-trained DDM and introduce an automatic search method. Last but not least, in contrast to the conventional understanding that DDMs have relatively poor semantic behaviors, we prove that the critical latent space we found already exhibits semantic subspace boundaries at the generic level in unconditional DDMs, which allows us to do controllable manipulation by guiding the denoising trajectory towards the targeted boundary via a single-step operation. We conduct extensive experiments on multiple DPMs architectures (DDPM, iDDPM) and datasets (CelebA, CelebA-HQ, LSUN-church, LSUN-bedroom, AFHQ-dog) with different resolutions (64, 256), achieving superior or state-of-the-art performance in various task scenarios (image semantic editing, text-based editing, unconditional semantic control) to demonstrate the effectiveness.

Carlos E. Jimenez, Olga Russakovsky, Karthik Narasimhan · princeton

We introduce CARETS, a systematic test suite to measure consistency and robustness of modern VQA models through a series of six fine-grained capability tests. In contrast to existing VQA test sets, CARETS features balanced question generation to create pairs of instances to test models, with each pair focusing on a specific capability such as rephrasing, logical symmetry or image obfuscation. We evaluate six modern VQA systems on CARETS and identify several actionable weaknesses in model comprehension, especially with concepts such as negation, disjunction, or hypernym invariance. Interestingly, even the most sophisticated models are sensitive to aspects such as swapping the order of terms in a conjunction or varying the number of answer choices mentioned in the question. We release CARETS to be used as an extensible tool for evaluating multi-modal model robustness.

Mengxue Qu, Yu Wu, Wu Liu et al.

In this paper, we investigate how to achieve better visual grounding with modern vision-language transformers, and propose a simple yet powerful Selective Retraining (SiRi) mechanism for this challenging task. Particularly, SiRi conveys a significant principle to the research of visual grounding, i.e., a better initialized vision-language encoder would help the model converge to a better local minimum, advancing the performance accordingly. In specific, we continually update the parameters of the encoder as the training goes on, while periodically re-initialize rest of the parameters to compel the model to be better optimized based on an enhanced encoder. SiRi can significantly outperform previous approaches on three popular benchmarks. Specifically, our method achieves 83.04% Top1 accuracy on RefCOCO+ testA, outperforming the state-of-the-art approaches (training from scratch) by more than 10.21%. Additionally, we reveal that SiRi performs surprisingly superior even with limited training data. We also extend it to transformer-based visual grounding models and other vision-language tasks to verify the validity.

Vikram V. Ramaswamy, Sing Yu Lin, Dora Zhao et al.

Current dataset collection methods typically scrape large amounts of data from the web. While this technique is extremely scalable, data collected in this way tends to reinforce stereotypical biases, can contain personally identifiable information, and typically originates from Europe and North America. In this work, we rethink the dataset collection paradigm and introduce GeoDE, a geographically diverse dataset with 61,940 images from 40 classes and 6 world regions, with no personally identifiable information, collected by soliciting images from people around the world. We analyse GeoDE to understand differences in images collected in this manner compared to web-scraping. We demonstrate its use as both an evaluation and training dataset, allowing us to highlight and begin to mitigate the shortcomings in current models, despite GeoDE's relatively small size. We release the full dataset and code at https://geodiverse-data-collection.cs.princeton.edu

Sunnie S. Y. Kim, Elizabeth Anne Watkins, Olga Russakovsky et al.

Despite the proliferation of explainable AI (XAI) methods, little is understood about end-users' explainability needs and behaviors around XAI explanations. To address this gap and contribute to understanding how explainability can support human-AI interaction, we conducted a mixed-methods study with 20 end-users of a real-world AI application, the Merlin bird identification app, and inquired about their XAI needs, uses, and perceptions. We found that participants desire practically useful information that can improve their collaboration with the AI, more so than technical system details. Relatedly, participants intended to use XAI explanations for various purposes beyond understanding the AI's outputs: calibrating trust, improving their task skills, changing their behavior to supply better inputs to the AI, and giving constructive feedback to developers. Finally, among existing XAI approaches, participants preferred part-based explanations that resemble human reasoning and explanations. We discuss the implications of our findings and provide recommendations for future XAI design.

Sunayana Rane, Mira L. Nencheva, Zeyu Wang et al.

For human children as well as machine learning systems, a key challenge in learning a word is linking the word to the visual phenomena it describes. We explore this aspect of word learning by using the performance of computer vision systems as a proxy for the difficulty of learning a word from visual cues. We show that the age at which children acquire different categories of words is correlated with the performance of visual classification and captioning systems, over and above the expected effects of word frequency. The performance of the computer vision systems is correlated with human judgments of the concreteness of words, which are in turn a predictor of children's word learning, suggesting that these models are capturing the relationship between words and visual phenomena.

Angelina Wang, Vikram V. Ramaswamy, Olga Russakovsky

Research in machine learning fairness has historically considered a single binary demographic attribute; however, the reality is of course far more complicated. In this work, we grapple with questions that arise along three stages of the machine learning pipeline when incorporating intersectionality as multiple demographic attributes: (1) which demographic attributes to include as dataset labels, (2) how to handle the progressively smaller size of subgroups during model training, and (3) how to move beyond existing evaluation metrics when benchmarking model fairness for more subgroups. For each question, we provide thorough empirical evaluation on tabular datasets derived from the US Census, and present constructive recommendations for the machine learning community. First, we advocate for supplementing domain knowledge with empirical validation when choosing which demographic attribute labels to train on, while always evaluating on the full set of demographic attributes. Second, we warn against using data imbalance techniques without considering their normative implications and suggest an alternative using the structure in the data. Third, we introduce new evaluation metrics which are more appropriate for the intersectional setting. Overall, we provide substantive suggestions on three necessary (albeit not sufficient!) considerations when incorporating intersectionality into machine learning.

Angelina Wang, Olga Russakovsky

Transfer learning is beneficial by allowing the expressive features of models pretrained on large-scale datasets to be finetuned for the target task of smaller, more domain-specific datasets. However, there is a concern that these pretrained models may come with their own biases which would propagate into the finetuned model. In this work, we investigate bias when conceptualized as both spurious correlations between the target task and a sensitive attribute as well as underrepresentation of a particular group in the dataset. Under both notions of bias, we find that (1) models finetuned on top of pretrained models can indeed inherit their biases, but (2) this bias can be corrected for through relatively minor interventions to the finetuning dataset, and often with a negligible impact to performance. Our findings imply that careful curation of the finetuning dataset is important for reducing biases on a downstream task, and doing so can even compensate for bias in the pretrained model.

Sruthi Sudhakar, Viraj Prabhu, Olga Russakovsky et al.

As computer vision systems are being increasingly deployed at scale in high-stakes applications like autonomous driving, concerns about social bias in these systems are rising. Analysis of fairness in real-world vision systems, such as object detection in driving scenes, has been limited to observing predictive inequity across attributes such as pedestrian skin tone, and lacks a consistent methodology to disentangle the role of confounding variables e.g. does my model perform worse for a certain skin tone, or are such scenes in my dataset more challenging due to occlusion and crowds? In this work, we introduce ICON$^2$, a framework for robustly answering this question. ICON$^2$ leverages prior knowledge on the deficiencies of object detection systems to identify performance discrepancies across sub-populations, compute correlations between these potential confounders and a given sensitive attribute, and control for the most likely confounders to obtain a more reliable estimate of model bias. Using our approach, we conduct an in-depth study on the performance of object detection with respect to income from the BDD100K driving dataset, revealing useful insights.

Xindi Wu, Dingli Yu, Yangsibo Huang et al.

Compositionality is a critical capability in Text-to-Image (T2I) models, as it reflects their ability to understand and combine multiple concepts from text descriptions. Existing evaluations of compositional capability rely heavily on human-designed text prompts or fixed templates, limiting their diversity and complexity, and yielding low discriminative power. We propose ConceptMix, a scalable, controllable, and customizable benchmark which automatically evaluates compositional generation ability of T2I models. This is done in two stages. First, ConceptMix generates the text prompts: concretely, using categories of visual concepts (e.g., objects, colors, shapes, spatial relationships), it randomly samples an object and k-tuples of visual concepts, then uses GPT4-o to generate text prompts for image generation based on these sampled concepts. Second, ConceptMix evaluates the images generated in response to these prompts: concretely, it checks how many of the k concepts actually appeared in the image by generating one question per visual concept and using a strong VLM to answer them. Through administering ConceptMix to a diverse set of T2I models (proprietary as well as open ones) using increasing values of k, we show that our ConceptMix has higher discrimination power than earlier benchmarks. Specifically, ConceptMix reveals that the performance of several models, especially open models, drops dramatically with increased k. Importantly, it also provides insight into the lack of prompt diversity in widely-used training datasets. Additionally, we conduct extensive human studies to validate the design of ConceptMix and compare our automatic grading with human judgement. We hope it will guide future T2I model development.

Nicole Meister, Dora Zhao, Angelina Wang et al.

Gender biases are known to exist within large-scale visual datasets and can be reflected or even amplified in downstream models. Many prior works have proposed methods for mitigating gender biases, often by attempting to remove gender expression information from images. To understand the feasibility and practicality of these approaches, we investigate what $\textit{gender artifacts}$ exist within large-scale visual datasets. We define a $\textit{gender artifact}$ as a visual cue that is correlated with gender, focusing specifically on those cues that are learnable by a modern image classifier and have an interpretable human corollary. Through our analyses, we find that gender artifacts are ubiquitous in the COCO and OpenImages datasets, occurring everywhere from low-level information (e.g., the mean value of the color channels) to the higher-level composition of the image (e.g., pose and location of people). Given the prevalence of gender artifacts, we claim that attempts to remove gender artifacts from such datasets are largely infeasible. Instead, the responsibility lies with researchers and practitioners to be aware that the distribution of images within datasets is highly gendered and hence develop methods which are robust to these distributional shifts across groups.

Xindi Wu, Byron Zhang, Zhiwei Deng et al.

Dataset distillation methods reduce large-scale datasets to smaller sets of synthetic data, preserving sufficient information to quickly train a new model from scratch. However, prior work on dataset distillation has focused exclusively on image classification datasets, whereas modern large-scale datasets are primarily vision-language datasets. In this work, we design the first vision-language dataset distillation method, building on the idea of trajectory matching. A key challenge is that vision-language datasets do not have a set of discrete classes. To overcome this, our proposed method jointly distills image-text pairs in a contrastive formulation. Further, we leverage Low-Rank Adaptation (LoRA) matching to enable more efficient and effective trajectory matching in complex modern vision-language models. Since there are no existing baselines, we compare our distillation approach with three adapted vision-language coreset selection methods. We demonstrate significant improvements on the challenging Flickr30K and COCO retrieval benchmarks: for example, on Flickr30K, the best coreset selection method selecting 1000 image-text pairs for training achieves only 5.6% image-to-text retrieval accuracy (i.e., recall@1); in contrast, our dataset distillation almost doubles that to 9.9% with just 100 training pairs, an order of magnitude fewer.

Vikram V. Ramaswamy, Sunnie S. Y. Kim, Nicole Meister et al.

Deep learning models have achieved remarkable success in different areas of machine learning over the past decade; however, the size and complexity of these models make them difficult to understand. In an effort to make them more interpretable, several recent works focus on explaining parts of a deep neural network through human-interpretable, semantic attributes. However, it may be impossible to completely explain complex models using only semantic attributes. In this work, we propose to augment these attributes with a small set of uninterpretable features. Specifically, we develop a novel explanation framework ELUDE (Explanation via Labelled and Unlabelled DEcomposition) that decomposes a model's prediction into two parts: one that is explainable through a linear combination of the semantic attributes, and another that is dependent on the set of uninterpretable features. By identifying the latter, we are able to analyze the "unexplained" portion of the model, obtaining insights into the information used by the model. We show that the set of unlabelled features can generalize to multiple models trained with the same feature space and compare our work to two popular attribute-oriented methods, Interpretable Basis Decomposition and Concept Bottleneck, and discuss the additional insights ELUDE provides.

Vikram V. Ramaswamy, Sunnie S. Y. Kim, Ruth Fong et al.

Concept-based explanations for convolutional neural networks (CNNs) aim to explain model behavior and outputs using a pre-defined set of semantic concepts (e.g., the model recognizes scene class ``bedroom'' based on the presence of concepts ``bed'' and ``pillow''). However, they often do not faithfully (i.e., accurately) characterize the model's behavior and can be too complex for people to understand. Further, little is known about how faithful and understandable different explanation methods are, and how to control these two properties. In this work, we propose UFO, a unified method for controlling Understandability and Faithfulness Objectives in concept-based explanations. UFO formalizes understandability and faithfulness as mathematical objectives and unifies most existing concept-based explanations methods for CNNs. Using UFO, we systematically investigate how explanations change as we turn the knobs of faithfulness and understandability. Our experiments demonstrate a faithfulness-vs-understandability tradeoff: increasing understandability reduces faithfulness. We also provide insights into the ``disagreement problem'' in explainable machine learning, by analyzing when and how concept-based explanations disagree with each other.

Ye Zhu, Yu Wu, Duo Xu et al.

In this work, we investigate the domain generalization capabilities of diffusion models in the context of synthesizing images that are distinct from the training data. Instead of fine-tuning, we tackle this challenge from a sampling-based perspective using frozen, pre-trained diffusion models. Specifically, we demonstrate that arbitrary out-of-domain (OOD) images establish Gaussian priors in the latent spaces of a given model after inversion, and that these priors are separable from those of the original training domain. This OOD latent property allows us to synthesize new images of the target unseen domain by discovering qualified OOD latent encodings in the inverted noisy spaces, without altering the pre-trained models. Our cross-model and cross-domain experiments show that the proposed sampling-based method can expand the latent space and generate unseen images without impairing the generation quality of the original domain. We also showcase a practical application of our approach using astrophysical data, highlighting the potential of this generalization paradigm in data-sparse fields such as scientific exploration.

William Yang, Byron Zhang, Olga Russakovsky

The task of out-of-distribution (OOD) detection is notoriously ill-defined. Earlier works focused on new-class detection, aiming to identify label-altering data distribution shifts, also known as "semantic shift." However, recent works argue for a focus on failure detection, expanding the OOD evaluation framework to account for label-preserving data distribution shifts, also known as "covariate shift." Intriguingly, under this new framework, complex OOD detectors that were previously considered state-of-the-art now perform similarly to, or even worse than the simple maximum softmax probability baseline. This raises the question: what are the latest OOD detectors actually detecting? Deciphering the behavior of OOD detection algorithms requires evaluation datasets that decouples semantic shift and covariate shift. To aid our investigations, we present ImageNet-OOD, a clean semantic shift dataset that minimizes the interference of covariate shift. Through comprehensive experiments, we show that OOD detectors are more sensitive to covariate shift than to semantic shift, and the benefits of recent OOD detection algorithms on semantic shift detection is minimal. Our dataset and analyses provide important insights for guiding the design of future OOD detectors.

Nobline Yoo, Olga Russakovsky

The goal of 2D human pose estimation (HPE) is to localize anatomical landmarks, given an image of a person in a pose. SOTA techniques make use of thousands of labeled figures (finetuning transformers or training deep CNNs), acquired using labor-intensive crowdsourcing. On the other hand, self-supervised methods re-frame the HPE task as a reconstruction problem, enabling them to leverage the vast amount of unlabeled visual data, though at the present cost of accuracy. In this work, we explore ways to improve self-supervised HPE. We (1) analyze the relationship between reconstruction quality and pose estimation accuracy, (2) develop a model pipeline that outperforms the baseline which inspired our work, using less than one-third the amount of training data, and (3) offer a new metric suitable for self-supervised settings that measures the consistency of predicted body part length proportions. We show that a combination of well-engineered reconstruction losses and inductive priors can help coordinate pose learning alongside reconstruction in a self-supervised paradigm.

Sunnie S. Y. Kim, Nicole Meister, Vikram V. Ramaswamy et al.

As AI technology is increasingly applied to high-impact, high-risk domains, there have been a number of new methods aimed at making AI models more human interpretable. Despite the recent growth of interpretability work, there is a lack of systematic evaluation of proposed techniques. In this work, we introduce HIVE (Human Interpretability of Visual Explanations), a novel human evaluation framework that assesses the utility of explanations to human users in AI-assisted decision making scenarios, and enables falsifiable hypothesis testing, cross-method comparison, and human-centered evaluation of visual interpretability methods. To the best of our knowledge, this is the first work of its kind. Using HIVE, we conduct IRB-approved human studies with nearly 1000 participants and evaluate four methods that represent the diversity of computer vision interpretability works: GradCAM, BagNet, ProtoPNet, and ProtoTree. Our results suggest that explanations engender human trust, even for incorrect predictions, yet are not distinct enough for users to distinguish between correct and incorrect predictions. We open-source HIVE to enable future studies and encourage more human-centered approaches to interpretability research.

Sunnie S. Y. Kim, Sharon Zhang, Nicole Meister et al.

Singh et al. (2020) point out the dangers of contextual bias in visual recognition datasets. They propose two methods, CAM-based and feature-split, that better recognize an object or attribute in the absence of its typical context while maintaining competitive within-context accuracy. To verify their performance, we attempted to reproduce all 12 tables in the original paper, including those in the appendix. We also conducted additional experiments to better understand the proposed methods, including increasing the regularization in CAM-based and removing the weighted loss in feature-split. As the original code was not made available, we implemented the entire pipeline from scratch in PyTorch 1.7.0. Our implementation is based on the paper and email exchanges with the authors. We found that both proposed methods in the original paper help mitigate contextual bias, although for some methods, we could not completely replicate the quantitative results in the paper even after completing an extensive hyperparameter search. For example, on COCO-Stuff, DeepFashion, and UnRel, our feature-split model achieved an increase in accuracy on out-of-context images over the standard baseline, whereas on AwA, we saw a drop in performance. For the proposed CAM-based method, we were able to reproduce the original paper's results to within 0.5$\%$ mAP. Our implementation can be found at https://github.com/princetonvisualai/ContextualBias.

Kaiyu Yang, Jacqueline Yau, Li Fei-Fei et al.

Face obfuscation (blurring, mosaicing, etc.) has been shown to be effective for privacy protection; nevertheless, object recognition research typically assumes access to complete, unobfuscated images. In this paper, we explore the effects of face obfuscation on the popular ImageNet challenge visual recognition benchmark. Most categories in the ImageNet challenge are not people categories; however, many incidental people appear in the images, and their privacy is a concern. We first annotate faces in the dataset. Then we demonstrate that face obfuscation has minimal impact on the accuracy of recognition models. Concretely, we benchmark multiple deep neural networks on obfuscated images and observe that the overall recognition accuracy drops only slightly (<= 1.0%). Further, we experiment with transfer learning to 4 downstream tasks (object recognition, scene recognition, face attribute classification, and object detection) and show that features learned on obfuscated images are equally transferable. Our work demonstrates the feasibility of privacy-aware visual recognition, improves the highly-used ImageNet challenge benchmark, and suggests an important path for future visual datasets. Data and code are available at https://github.com/princetonvisualai/imagenet-face-obfuscation.

Angelina Wang, Olga Russakovsky

Mitigating bias in machine learning systems requires refining our understanding of bias propagation pathways: from societal structures to large-scale data to trained models to impact on society. In this work, we focus on one aspect of the problem, namely bias amplification: the tendency of models to amplify the biases present in the data they are trained on. A metric for measuring bias amplification was introduced in the seminal work by Zhao et al. (2017); however, as we demonstrate, this metric suffers from a number of shortcomings including conflating different types of bias amplification and failing to account for varying base rates of protected attributes. We introduce and analyze a new, decoupled metric for measuring bias amplification, $\text{BiasAmp}_{\rightarrow}$ (Directional Bias Amplification). We thoroughly analyze and discuss both the technical assumptions and normative implications of this metric. We provide suggestions about its measurement by cautioning against predicting sensitive attributes, encouraging the use of confidence intervals due to fluctuations in the fairness of models across runs, and discussing the limitations of what this metric captures. Throughout this paper, we work to provide an interrogative look at the technical measurement of bias amplification, guided by our normative ideas of what we want it to encompass. Code is located at https://github.com/princetonvisualai/directional-bias-amp

Arjun Mani, Nobline Yoo, Will Hinthorn et al.

Visual Question Answering (VQA) has become one of the key benchmarks of visual recognition progress. Multiple VQA extensions have been explored to better simulate real-world settings: different question formulations, changing training and test distributions, conversational consistency in dialogues, and explanation-based answering. In this work, we further expand this space by considering visual questions that include a spatial point of reference. Pointing is a nearly universal gesture among humans, and real-world VQA is likely to involve a gesture towards the target region. Concretely, we (1) introduce and motivate point-input questions as an extension of VQA, (2) define three novel classes of questions within this space, and (3) for each class, introduce both a benchmark dataset and a series of baseline models to handle its unique challenges. There are two key distinctions from prior work. First, we explicitly design the benchmarks to require the point input, i.e., we ensure that the visual question cannot be answered accurately without the spatial reference. Second, we explicitly explore the more realistic point spatial input rather than the standard but unnatural bounding box input. Through our exploration we uncover and address several visual recognition challenges, including the ability to infer human intent, reason both locally and globally about the image, and effectively combine visual, language and spatial inputs. Code is available at: https://github.com/princetonvisualai/pointingqa .

Angelina Wang, Alexander Liu, Ryan Zhang et al.

Machine learning models are known to perpetuate and even amplify the biases present in the data. However, these data biases frequently do not become apparent until after the models are deployed. Our work tackles this issue and enables the preemptive analysis of large-scale datasets. REVISE (REvealing VIsual biaSEs) is a tool that assists in the investigation of a visual dataset, surfacing potential biases along three dimensions: (1) object-based, (2) person-based, and (3) geography-based. Object-based biases relate to the size, context, or diversity of the depicted objects. Person-based metrics focus on analyzing the portrayal of people within the dataset. Geography-based analyses consider the representation of different geographic locations. These three dimensions are deeply intertwined in how they interact to bias a dataset, and REVISE sheds light on this; the responsibility then lies with the user to consider the cultural and historical context, and to determine which of the revealed biases may be problematic. The tool further assists the user by suggesting actionable steps that may be taken to mitigate the revealed biases. Overall, the key aim of our work is to tackle the machine learning bias problem early in the pipeline. REVISE is available at https://github.com/princetonvisualai/revise-tool

Kaiyu Yang, Olga Russakovsky, Jia Deng

Understanding the spatial relations between objects in images is a surprisingly challenging task. A chair may be "behind" a person even if it appears to the left of the person in the image (depending on which way the person is facing). Two students that appear close to each other in the image may not in fact be "next to" each other if there is a third student between them. We introduce SpatialSense, a dataset specializing in spatial relation recognition which captures a broad spectrum of such challenges, allowing for proper benchmarking of computer vision techniques. SpatialSense is constructed through adversarial crowdsourcing, in which human annotators are tasked with finding spatial relations that are difficult to predict using simple cues such as 2D spatial configuration or language priors. Adversarial crowdsourcing significantly reduces dataset bias and samples more interesting relations in the long tail compared to existing datasets. On SpatialSense, state-of-the-art recognition models perform comparably to simple baselines, suggesting that they rely on straightforward cues instead of fully reasoning about this complex task. The SpatialSense benchmark provides a path forward to advancing the spatial reasoning capabilities of computer vision systems. The dataset and code are available at https://github.com/princeton-vl/SpatialSense.

Sunnie S. Y. Kim, Jennifer Wortman Vaughan, Q. Vera Liao et al. · microsoft-research

Large language models (LLMs) can produce erroneous responses that sound fluent and convincing, raising the risk that users will rely on these responses as if they were correct. Mitigating such overreliance is a key challenge. Through a think-aloud study in which participants use an LLM-infused application to answer objective questions, we identify several features of LLM responses that shape users' reliance: explanations (supporting details for answers), inconsistencies in explanations, and sources. Through a large-scale, pre-registered, controlled experiment (N=308), we isolate and study the effects of these features on users' reliance, accuracy, and other measures. We find that the presence of explanations increases reliance on both correct and incorrect responses. However, we observe less reliance on incorrect responses when sources are provided or when explanations exhibit inconsistencies. We discuss the implications of these findings for fostering appropriate reliance on LLMs.

Xindi Wu, Mengzhou Xia, Rulin Shao et al.

Training vision-language models via instruction tuning often relies on large mixtures of data spanning diverse tasks and domains. However, these mixtures frequently include redundant information, increasing computational costs without proportional performance gains, necessitating more effective data selection strategies. Existing methods typically rely on task-agnostic heuristics to estimate data importance or focus on optimizing single tasks in isolation, limiting their effectiveness in multitask settings. In this work, we introduce ICONS, a gradient-based Influence CONsensus approach for vision-language data Selection. Our method leverages first-order training dynamics to estimate the influence of individual training examples on validation performance and aggregates these estimates across tasks via majority voting over task-specific influences. This cross-task consensus identifies data points that are consistently valuable across tasks, enabling us to prioritize examples that drive overall performance. The voting-based design further mitigates issues such as score calibration and outlier sensitivity, resulting in robust and scalable data selection for diverse multitask mixtures. With only 20% of the data from LLaVA-665K and Cambrian-7M, our selected subsets retain 98.6% and 98.8% of the performance achieved with full datasets, and can even surpass full data training at a 60% selection ratio on LLaVA-665K. Our approach also generalizes to unseen tasks and architectures, demonstrating strong transfer. We release two compact, high-utility subsets, LLaVA-ICONS-133K and Cambrian-ICONS-1.4M, preserving impactful training examples for efficient and scalable vision-language model development.

Jihoon Chung, Tyler Zhu, Max Gonzalez Saez-Diez et al. · salesforce, stanford

The recent advent of Large Language Models (LLMs) has ushered sophisticated reasoning capabilities into the realm of video through Video Large Language Models (VideoLLMs). However, VideoLLMs currently rely on a single vision encoder for all of their visual processing, which limits the amount and type of visual information that can be conveyed to the LLM. Our method, MERV, Multi-Encoder Representation of Videos, instead leverages multiple frozen visual encoders to create a unified representation of a video, providing the VideoLLM with a comprehensive set of specialized visual knowledge. Spatio-temporally aligning the features from each encoder allows us to tackle a wider range of open-ended and multiple-choice video understanding questions and outperform prior state-of-the-art works. MERV is up to 3.7% better in accuracy than Video-LLaVA across the standard suite video understanding benchmarks, while also having a better Video-ChatGPT score. We also improve upon SeViLA, the previous best on zero-shot Perception Test accuracy, by 2.2%. MERV introduces minimal extra parameters and trains faster than equivalent single-encoder methods while parallelizing the visual processing. Finally, we provide qualitative evidence that MERV successfully captures domain knowledge from each of its encoders. Our results offer promising directions in utilizing multiple vision encoders for comprehensive video understanding.

Ruoyu Wang, Huayang Huang, Ye Zhu et al.

In this work, we introduce NoiseQuery as a novel method for enhanced noise initialization in versatile goal-driven text-to-image (T2I) generation. Specifically, we propose to leverage an aligned Gaussian noise as implicit guidance to complement explicit user-defined inputs, such as text prompts, for better generation quality and controllability. Unlike existing noise optimization methods designed for specific models, our approach is grounded in a fundamental examination of the generic finite-step noise scheduler design in diffusion formulation, allowing better generalization across different diffusion-based architectures in a tuning-free manner. This model-agnostic nature allows us to construct a reusable noise library compatible with multiple T2I models and enhancement techniques, serving as a foundational layer for more effective generation. Extensive experiments demonstrate that NoiseQuery enables fine-grained control and yields significant performance boosts not only over high-level semantics but also over low-level visual attributes, which are typically difficult to specify through text alone, with seamless integration into current workflows with minimal computational overhead.

Allison Chen, Ilia Sucholutsky, Olga Russakovsky et al.

Does language help make sense of the visual world? How important is it to actually see the world rather than having it described with words? These basic questions about the nature of intelligence have been difficult to answer because we only had one example of an intelligent system -- humans -- and limited access to cases that isolated language or vision. However, the development of sophisticated Vision-Language Models (VLMs) by artificial intelligence researchers offers us new opportunities to explore the contributions that language and vision make to learning about the world. We ablate components from the cognitive architecture of these models to identify their contributions to learning new tasks from limited data. We find that a language model leveraging all components recovers a majority of a VLM's performance, despite its lack of visual input, and that language seems to allow this by providing access to prior knowledge and reasoning.

Sai Wang, Ye Zhu, Ruoyu Wang et al.

Generative AI capabilities have grown substantially in recent years, raising renewed concerns about potential malicious use of generated data, or "deep fakes". However, deep fake datasets have not kept up with generative AI advancements sufficiently to enable the development of deep fake detection technology which can meaningfully alert human users in real-world settings. Existing datasets typically use GAN-based models and introduce spurious correlations by always editing similar face regions. To counteract the shortcomings, we introduce DETER, a large-scale dataset for DETEcting edited image Regions and deterring modern advanced generative manipulations. DETER includes 300,000 images manipulated by four state-of-the-art generators with three editing operations: face swapping (a standard coarse image manipulation), inpainting (a novel manipulation for deep fake datasets), and attribute editing (a subtle fine-grained manipulation). While face swapping and attribute editing are performed on similar face regions such as eyes and nose, the inpainting operation can be performed on random image regions, removing the spurious correlations of previous datasets. Careful image post-processing is performed to ensure deep fakes in DETER look realistic, and human studies confirm that human deep fake detection rate on DETER is 20.4% lower than on other fake datasets. Equipped with the dataset, we conduct extensive experiments and break-down analysis using our rich annotations and improved benchmark protocols, revealing future directions and the next set of challenges in developing reliable regional fake detection models.

Indu Panigrahi, Sunnie S. Y. Kim, Amna Liaqat et al.

Explanations for computer vision models are important tools for interpreting how the underlying models work. However, they are often presented in static formats, which pose challenges for users, including information overload, a gap between semantic and pixel-level information, and limited opportunities for exploration. We investigate interactivity as a mechanism for tackling these issues in three common explanation types: heatmap-based, concept-based, and prototype-based explanations. We conducted a study (N=24), using a bird identification task, involving participants with diverse technical and domain expertise. We found that while interactivity enhances user control, facilitates rapid convergence to relevant information, and allows users to expand their understanding of the model and explanation, it also introduces new challenges. To address these, we provide design recommendations for interactive computer vision explanations, including carefully selected default views, independent input controls, and constrained output spaces.

Amaya Dharmasiri, William Yang, Polina Kirichenko et al.

Coreset selection methods have shown promise in reducing the training data size while maintaining model performance for data-efficient machine learning. However, as many datasets suffer from biases that cause models to learn spurious correlations instead of causal features, it is important to understand whether and how dataset reduction methods may perpetuate, amplify, or mitigate these biases. In this work, we conduct the first comprehensive analysis of the implications of data selection on the spurious bias levels of the selected coresets and the robustness of downstream models trained on them. We use an extensive experimental setting spanning ten different spurious correlations benchmarks, five score metrics to characterize sample importance/ difficulty, and five data selection policies across a broad range of coreset sizes. Thereby, we unravel a series of nontrivial nuances in interactions between sample difficulty and bias alignment, as well as dataset bias and resultant model robustness. For example, we find that selecting coresets using embedding-based sample characterization scores runs a comparatively lower risk of inadvertently exacerbating bias than selecting using characterizations based on learning dynamics. Most importantly, our analysis reveals that although some coreset selection methods could achieve lower bias levels by prioritizing difficult samples, they do not reliably guarantee downstream robustness.

Ellie Zhou, Jihoon Chung, Olga Russakovsky

Human action recognition models often rely on background cues rather than human movement and pose to make predictions, a behavior known as background bias. We present a systematic analysis of background bias across classification models, contrastive text-image pretrained models, and Video Large Language Models (VLLM) and find that all exhibit a strong tendency to default to background reasoning. Next, we propose mitigation strategies for classification models and show that incorporating segmented human input effectively decreases background bias by 3.78%. Finally, we explore manual and automated prompt tuning for VLLMs, demonstrating that prompt design can steer predictions towards human-focused reasoning by 9.85%.

Nobline Yoo, Olga Russakovsky, Ye Zhu

Text-to-image (T2I) diffusion models have achieved strong performance in semantic alignment, yet they still struggle with generating the correct number of objects specified in prompts. Existing approaches typically incorporate auxiliary counting networks as external critics to enhance numeracy. However, since these critics must provide gradient guidance during generation, they are restricted to regression-based models that are inherently differentiable, thus excluding detector-based models with superior counting ability, whose count-via-enumeration nature is non-differentiable. To overcome this limitation, we propose Detector-to-Differentiable (D2D), a novel framework that transforms non-differentiable detection models into differentiable critics, thereby leveraging their superior counting ability to guide numeracy generation. Specifically, we design custom activation functions to convert detector logits into soft binary indicators, which are then used to optimize the noise prior at inference time with pre-trained T2I models. Our extensive experiments on SDXL-Turbo, SD-Turbo, and Pixart-DMD across four benchmarks of varying complexity (low-density, high-density, and multi-object scenarios) demonstrate consistent and substantial improvements in object counting accuracy (e.g., boosting up to 13.7% on D2D-Small, a 400-prompt, low-density benchmark), with minimal degradation in overall image quality and computational overhead.

Ye Zhu, Duo Xu, Zhiwei Deng et al.

We study Diffusion Schrödinger Bridge (DSB) models in the context of dynamical astrophysical systems, specifically tackling observational inverse prediction tasks within Giant Molecular Clouds (GMCs) for star formation. We introduce the Astro-DSB model, a variant of DSB with the pairwise domain assumption tailored for astrophysical dynamics. By investigating its learning process and prediction performance in both physically simulated data and in real observations (the Taurus B213 data), we present two main takeaways. First, from the astrophysical perspective, our proposed paired DSB method improves interpretability, learning efficiency, and prediction performance over conventional astrostatistical and other machine learning methods. Second, from the generative modeling perspective, probabilistic generative modeling reveals improvements over discriminative pixel-to-pixel modeling in Out-Of-Distribution (OOD) testing cases of physical simulations with unseen initial conditions and different dominant physical processes. Our study expands research into diffusion models beyond the traditional visual synthesis application and provides evidence of the models' learning abilities beyond pure data statistics, paving a path for future physics-aware generative models which can align dynamics between machine learning and real (astro)physical systems.

William Yang, Ye Zhu, Zhiwei Deng et al.

Dataset distillation has emerged as a strategy to overcome the hurdles associated with large datasets by learning a compact set of synthetic data that retains essential information from the original dataset. While distilled data can be used to train high performing models, little is understood about how the information is stored. In this study, we posit and answer three questions about the behavior, representativeness, and point-wise information content of distilled data. We reveal distilled data cannot serve as a substitute for real data during training outside the standard evaluation setting for dataset distillation. Additionally, the distillation process retains high task performance by compressing information related to the early training dynamics of real models. Finally, we provide an framework for interpreting distilled data and reveal that individual distilled data points contain meaningful semantic information. This investigation sheds light on the intricate nature of distilled data, providing a better understanding on how they can be effectively utilized.

Sunnie S. Y. Kim, Elizabeth Anne Watkins, Olga Russakovsky et al.

Trust is an important factor in people's interactions with AI systems. However, there is a lack of empirical studies examining how real end-users trust or distrust the AI system they interact with. Most research investigates one aspect of trust in lab settings with hypothetical end-users. In this paper, we provide a holistic and nuanced understanding of trust in AI through a qualitative case study of a real-world computer vision application. We report findings from interviews with 20 end-users of a popular, AI-based bird identification app where we inquired about their trust in the app from many angles. We find participants perceived the app as trustworthy and trusted it, but selectively accepted app outputs after engaging in verification behaviors, and decided against app adoption in certain high-stakes scenarios. We also find domain knowledge and context are important factors for trust-related assessment and decision-making. We discuss the implications of our findings and provide recommendations for future research on trust in AI.

Zeyu Wang, Yu Wu, Karthik Narasimhan et al.

Retrieving target videos based on text descriptions is a task of great practical value and has received increasing attention over the past few years. Despite recent progress, imperfect annotations in existing video retrieval datasets have posed significant challenges on model evaluation and development. In this paper, we tackle this issue by focusing on the less-studied setting of multi-query video retrieval, where multiple descriptions are provided to the model for searching over the video archive. We first show that multi-query retrieval task effectively mitigates the dataset noise introduced by imperfect annotations and better correlates with human judgement on evaluating retrieval abilities of current models. We then investigate several methods which leverage multiple queries at training time, and demonstrate that the multi-query inspired training can lead to superior performance and better generalization. We hope further investigation in this direction can bring new insights on building systems that perform better in real-world video retrieval applications.

Dora Zhao, Angelina Wang, Olga Russakovsky

Image captioning is an important task for benchmarking visual reasoning and for enabling accessibility for people with vision impairments. However, as in many machine learning settings, social biases can influence image captioning in undesirable ways. In this work, we study bias propagation pathways within image captioning, focusing specifically on the COCO dataset. Prior work has analyzed gender bias in captions using automatically-derived gender labels; here we examine racial and intersectional biases using manual annotations. Our first contribution is in annotating the perceived gender and skin color of 28,315 of the depicted people after obtaining IRB approval. Using these annotations, we compare racial biases present in both manual and automatically-generated image captions. We demonstrate differences in caption performance, sentiment, and word choice between images of lighter versus darker-skinned people. Further, we find the magnitude of these differences to be greater in modern captioning systems compared to older ones, thus leading to concerns that without proper consideration and mitigation these differences will only become increasingly prevalent. Code and data is available at https://princetonvisualai.github.io/imagecaptioning-bias .

Vikram V. Ramaswamy, Sunnie S. Y. Kim, Olga Russakovsky

Fairness in visual recognition is becoming a prominent and critical topic of discussion as recognition systems are deployed at scale in the real world. Models trained from data in which target labels are correlated with protected attributes (e.g., gender, race) are known to learn and exploit those correlations. In this work, we introduce a method for training accurate target classifiers while mitigating biases that stem from these correlations. We use GANs to generate realistic-looking images, and perturb these images in the underlying latent space to generate training data that is balanced for each protected attribute. We augment the original dataset with this perturbed generated data, and empirically demonstrate that target classifiers trained on the augmented dataset exhibit a number of both quantitative and qualitative benefits. We conduct a thorough evaluation across multiple target labels and protected attributes in the CelebA dataset, and provide an in-depth analysis and comparison to existing literature in the space.

Zeyu Wang, Berthy Feng, Karthik Narasimhan et al.

Despite considerable progress, state of the art image captioning models produce generic captions, leaving out important image details. Furthermore, these systems may even misrepresent the image in order to produce a simpler caption consisting of common concepts. In this paper, we first analyze both modern captioning systems and evaluation metrics through empirical experiments to quantify these phenomena. We find that modern captioning systems return higher likelihoods for incorrect distractor sentences compared to ground truth captions, and that evaluation metrics like SPICE can be 'topped' using simple captioning systems relying on object detectors. Inspired by these observations, we design a new metric (SPICE-U) by introducing a notion of uniqueness over the concepts generated in a caption. We show that SPICE-U is better correlated with human judgements compared to SPICE, and effectively captures notions of diversity and descriptiveness. Finally, we also demonstrate a general technique to improve any existing captioning model -- by using mutual information as a re-ranking objective during decoding. Empirically, this results in more unique and informative captions, and improves three different state-of-the-art models on SPICE-U as well as average score over existing metrics.

Zhiwei Deng, Karthik Narasimhan, Olga Russakovsky

The ability to perform effective planning is crucial for building an instruction-following agent. When navigating through a new environment, an agent is challenged with (1) connecting the natural language instructions with its progressively growing knowledge of the world; and (2) performing long-range planning and decision making in the form of effective exploration and error correction. Current methods are still limited on both fronts despite extensive efforts. In this paper, we introduce the Evolving Graphical Planner (EGP), a model that performs global planning for navigation based on raw sensory input. The model dynamically constructs a graphical representation, generalizes the action space to allow for more flexible decision making, and performs efficient planning on a proxy graph representation. We evaluate our model on a challenging Vision-and-Language Navigation (VLN) task with photorealistic images and achieve superior performance compared to previous navigation architectures. For instance, we achieve a 53% success rate on the test split of the Room-to-Room navigation task through pure imitation learning, outperforming previous navigation architectures by up to 5%.

Felix Yu, Zhiwei Deng, Karthik Narasimhan et al.

In the Vision-and-Language Navigation (VLN) task, an agent with egocentric vision navigates to a destination given natural language instructions. The act of manually annotating these instructions is timely and expensive, such that many existing approaches automatically generate additional samples to improve agent performance. However, these approaches still have difficulty generalizing their performance to new environments. In this work, we investigate the popular Room-to-Room (R2R) VLN benchmark and discover that what is important is not only the amount of data you synthesize, but also how you do it. We find that shortest path sampling, which is used by both the R2R benchmark and existing augmentation methods, encode biases in the action space of the agent which we dub as action priors. We then show that these action priors offer one explanation toward the poor generalization of existing works. To mitigate such priors, we propose a path sampling method based on random walks to augment the data. By training with this augmentation strategy, our agent is able to generalize better to unknown environments compared to the baseline, significantly improving model performance in the process.

Kaiyu Yang, Klint Qinami, Li Fei-Fei et al.

Computer vision technology is being used by many but remains representative of only a few. People have reported misbehavior of computer vision models, including offensive prediction results and lower performance for underrepresented groups. Current computer vision models are typically developed using datasets consisting of manually annotated images or videos; the data and label distributions in these datasets are critical to the models' behavior. In this paper, we examine ImageNet, a large-scale ontology of images that has spurred the development of many modern computer vision methods. We consider three key factors within the "person" subtree of ImageNet that may lead to problematic behavior in downstream computer vision technology: (1) the stagnant concept vocabulary of WordNet, (2) the attempt at exhaustive illustration of all categories with images, and (3) the inequality of representation in the images within concepts. We seek to illuminate the root causes of these concerns and take the first steps to mitigate them constructively.

Jonathan C. Stroud, Ryan McCaffrey, Rada Mihalcea et al.

Temporal grounding entails establishing a correspondence between natural language event descriptions and their visual depictions. Compositional modeling becomes central: we first ground atomic descriptions "girl eating an apple," "batter hitting the ball" to short video segments, and then establish the temporal relationships between the segments. This compositional structure enables models to recognize a wider variety of events not seen during training through recognizing their atomic sub-events. Explicit temporal modeling accounts for a wide variety of temporal relationships that can be expressed in language: e.g., in the description "girl stands up from the table after eating an apple" the visual ordering of the events is reversed, with first "eating an apple" followed by "standing up from the table." We leverage these observations to develop a unified deep architecture, CTG-Net, to perform temporal grounding of natural language event descriptions to videos. We demonstrate that our system outperforms prior state-of-the-art methods on the DiDeMo, Tempo-TL, and Tempo-HL temporal grounding datasets.

Zeyu Wang, Klint Qinami, Ioannis Christos Karakozis et al.

Computer vision models learn to perform a task by capturing relevant statistics from training data. It has been shown that models learn spurious age, gender, and race correlations when trained for seemingly unrelated tasks like activity recognition or image captioning. Various mitigation techniques have been presented to prevent models from utilizing or learning such biases. However, there has been little systematic comparison between these techniques. We design a simple but surprisingly effective visual recognition benchmark for studying bias mitigation. Using this benchmark, we provide a thorough analysis of a wide range of techniques. We highlight the shortcomings of popular adversarial training approaches for bias mitigation, propose a simple but similarly effective alternative to the inference-time Reducing Bias Amplification method of Zhao et al., and design a domain-independent training technique that outperforms all other methods. Finally, we validate our findings on the attribute classification task in the CelebA dataset, where attribute presence is known to be correlated with the gender of people in the image, and demonstrate that the proposed technique is effective at mitigating real-world gender bias.

Joshua C. Peterson, Ruairidh M. Battleday, Thomas L. Griffiths et al.

The classification performance of deep neural networks has begun to asymptote at near-perfect levels. However, their ability to generalize outside the training set and their robustness to adversarial attacks have not. In this paper, we make progress on this problem by training with full label distributions that reflect human perceptual uncertainty. We first present a new benchmark dataset which we call CIFAR10H, containing a full distribution of human labels for each image of the CIFAR10 test set. We then show that, while contemporary classifiers fail to exhibit human-like uncertainty on their own, explicit training on our dataset closes this gap, supports improved generalization to increasingly out-of-training-distribution test datasets, and confers robustness to adversarial attacks.