Intel Labs Presents 24 Papers on Innovative AI and Computer Vision Research at CVPR 2024

Scott Bair is a key voice at Intel Labs, sharing insights into innovative research for inventing tomorrow’s technology.

Highlights

• Intel Labs researchers will present 24 papers at CVPR 2024 on June 17-21.
• Six Intel Labs papers have been accepted as main conference papers, including highlight paper LiSA: LiDAR Localization with Semantic Awareness, which is the first method that incorporates semantic awareness into scene coordinate regression to boost localization accuracy.
• The other main conference papers cover topics including using egocentric action scene graphs for long-form understanding of videos, a new framework for editing 2D images by incorporating 3D tools, a generalizable AI framework for creating panorama and 3D images from multiple input modalities, improving training efficiency with small-scale inverted data for knowledge distillation, and a novel approach to model fingerprinting.
• Intel Labs researchers co-organized three workshops, including Urban Scene Modeling, Deep Learning for Geometric Computing, and Safe Artificial Intelligence for All Domains.
  
  

Intel Labs will present 24 papers accepted at the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2024) on June 17-21 in Seattle. The event features the latest advances in computer vision, pattern recognition, machine learning, robotics, and artificial intelligence. Six Intel Labs papers have been accepted as main conference papers, including highlight paper LiSA: LiDAR Localization with Semantic Awareness, which is the first method that incorporates semantic awareness into scene coordinate regression to boost localization accuracy. The other main conference papers cover topics including using egocentric action scene graphs for long-form understanding of egocentric videos, a new framework for editing 2D images by incorporating 3D tools, a generalizable AI framework that can create panorama and 3D images from multiple input modalities, improving training efficiency with small-scale inverted data for knowledge distillation, and a novel approach to model fingerprinting.

Intel Labs researchers co-organized three workshops during CVPR, including Urban Scene Modeling: Where Vision Meets Photogrammetry and Graphics on June 17, and Deep Learning for Geometric Computing and Safe Artificial Intelligence for All Domains on June 18.

Intel is also sponsoring an AI Summit on Monday, June 17 from 1-5 pm, and a networking meetup on Thursday, June 20 from 7-10 pm. Stop by the Intel booth #1431 for more details.

Main Conference Papers

Main Conference Paper Highlight:

LiSA: LiDAR Localization with Semantic Awareness

LiDAR localization is a fundamental task in robotics and computer vision, which estimates the pose of a LiDAR point cloud within a global map. Scene coordinate regression (SCR) has demonstrated state-of-the-art performance in this task. In SCR, a scene is represented as a neural network, which outputs the world coordinates for each point in the input point cloud. However, SCR treats all points equally during localization, ignoring the fact that not all objects are beneficial for localization. For example, dynamic objects and repeating structures often negatively impact SCR. To address this problem, we introduce LiSA, the first method that incorporates semantic awareness into SCR to boost the localization robustness and accuracy. To avoid extra computation or network parameters during inference, we distill the knowledge from a segmentation model to the original SCR network. Experiments show the superior performance of LiSA on standard LiDAR localization benchmarks compared to state-of-the-art methods. Applying knowledge distillation not only preserves high efficiency but also achieves higher localization accuracy than introducing extra semantic segmentation modules. We also analyze the benefit of semantic information for LiDAR localization.

Action Scene Graphs for Long-Form Understanding of Egocentric Videos

We present Egocentric Action Scene Graphs (EASGs), a new representation for long-form understanding of egocentric videos. EASGs extend standard manually-annotated representations of egocentric videos, such as verb-noun action labels, by providing a temporally evolving graph-based description of the actions performed by the camera wearer, including interacted objects, their relationships, and how actions unfold in time. Through a novel annotation procedure, we extend the Ego4D dataset by adding manually labeled Egocentric Action Scene Graphs offering a rich set of annotations designed for long-from egocentric video understanding. We hence define the EASG generation task and provide a baseline approach, establishing preliminary benchmarks. Experiments on two downstream tasks, egocentric action anticipation and egocentric activity summarization, highlight the effectiveness of EASGs for long-form egocentric video understanding.

Image Sculpting: Precise Object Editing with 3D Geometry Control

We present Image Sculpting, a new framework for editing 2D images by incorporating tools from 3D geometry and graphics. This approach differs markedly from existing methods, which are confined to 2D spaces and typically rely on textual instructions, leading to ambiguity and limited control. Image Sculpting converts 2D objects into 3D, enabling direct interaction with their 3D geometry. Post-editing, these objects are re-rendered into 2D, merging into the original image to produce high-fidelity results through a coarse-to-fine enhancement process. The framework supports precise, quantifiable, and physically plausible editing options such as pose editing, rotation, translation, 3D composition, carving, and serial addition. It marks an initial step towards combining the creative freedom of generative models with the precision of graphics pipelines.

Language Model Assisted Generation of Images with Consistency

In the current era of generative AI breakthroughs, generating panoramic scenes from a single input image remains a key challenge. Most existing methods use diffusion-based iterative or simultaneous multi-view inpainting. However, the lack of global scene layout priors leads to subpar outputs with duplicated objects (e.g., multiple beds in a bedroom) or requires time-consuming human text inputs for each view. We propose Language Model Assisted Generation of Images with Consistency, a novel method leveraging large language models for guidance while diffusing multiple coherent views of 360-degree panoramic scenes. The model harnesses pre-trained diffusion and language models without fine-tuning, ensuring zero-shot performance. The output quality is further enhanced by super-resolution and multi-view fusion techniques. Extensive experiments demonstrate that the resulting panoramic scenes feature better scene layouts and perspective view rendering quality compared to related works, with more than 70% preference in human evaluations. Combined with conditional diffusion models, the model can accept various input modalities, including but not limited to text, depth maps, sketches, and colored scripts. Applying depth estimation further enables 3D point cloud generation and dynamic scene exploration with fluid camera motion.

Small Scale Data-Free Knowledge Distillation

Data-free knowledge distillation is able to utilize the knowledge learned by a large teacher network to augment the training of a smaller student network without accessing the original training data, avoiding privacy, security, and proprietary risks in real applications. In this line of research, existing methods typically follow an inversion and distillation paradigm in which a generative adversarial network on-the-fly trained with the guidance of the pre-trained teacher network is used to synthesize a large-scale sample set for knowledge distillation. In this paper, we reexamine this common data-free knowledge distillation paradigm, showing that there is considerable room to improve the overall training efficiency through a lens of “small-scale inverted data for knowledge distillation.” In light of three empirical observations indicating the importance of how to balance class distributions in terms of synthetic sample diversity and difficulty during both data inversion and distillation processes, we propose Small Scale Data-free Knowledge Distillation (SSD-KD). In formulation, SSD-KD introduces a modulating function to balance synthetic samples and a priority sampling function to select proper samples, facilitated by a dynamic replay buffer and a reinforcement learning strategy. As a result, SSD-KD can perform distillation training conditioned on an extremely small scale of synthetic samples (e.g., 10× less than the original training data scale), making the overall training efficiency one or two orders of magnitude faster than many mainstream methods while retaining superior or competitive model performance, as demonstrated on popular image classification and semantic segmentation benchmarks.

WOUAF: Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models

The rapid advancement of generative models, facilitating the creation of hyper-realistic images from textual descriptions, has concurrently escalated critical societal concerns such as misinformation. Although providing some mitigation, traditional fingerprinting mechanisms fall short in attributing responsibility for the malicious use of synthetic images. This paper introduces a novel approach to model fingerprinting that assigns responsibility for the generated images, thereby serving as a potential countermeasure to model misuse. Our method modifies generative models based on each user's unique digital fingerprint, imprinting a unique identifier onto the resultant content that can be traced back to the user. This approach, incorporating fine-tuning into Text-to-Image (T2I) tasks using the Stable Diffusion Model, demonstrates near-perfect attribution accuracy with a minimal impact on output quality. Through extensive evaluation, we show that our method outperforms baseline methods with an average improvement of 11% in handling image post-processes. Our method presents a promising and novel avenue for accountable model distribution and responsible use.

Workshops

Urban Scene Modeling: Where Vision Meets Photogrammetry and Graphics

Monday, June 17, 8:30 a.m.–5:30 p.m., Summit 443

Rapid urbanization poses social and environmental challenges. Addressing these issues effectively requires access to accurate and up-to-date 3D building models, obtained promptly and cost-effectively. Urban modeling is an interdisciplinary topic among computer vision, graphics, and photogrammetry. The demand for automated interpretation of scene geometry and semantics has surged due to various applications, including autonomous navigation, augmented reality, smart cities, and digital twins. As a result, substantial research effort has been dedicated to urban scene modeling within the computer vision and graphics communities, with a particular focus on photogrammetry, which has coped with urban modeling challenges for decades. This workshop is intended to bring researchers from these communities together. Through invited talks, spotlight presentations, a workshop challenge, and a poster session, it will increase interdisciplinary interaction and collaboration among photogrammetry, computer vision, and graphics. We also solicit original contributions in the areas related to urban scene modeling.

Deep Learning for Geometric Computing

Tuesday, June 18, 8:30 a.m.–5:30 p.m., Summit 448

Computer vision approaches have made tremendous efforts toward understanding shape from various data formats, especially since entering the deep learning era. Although accurate results have been obtained in detection, recognition, and segmentation, there is less attention and research on extracting topological and geometric information from shapes. These geometric representations provide compact and intuitive abstractions for modeling, synthesis, compression, matching, and analysis. Extracting such representations is significantly different from segmentation and recognition tasks, as they contain both local and global information about the shape. To advance the state of the art in topological and geometric shape analysis using deep learning, we aim to gather researchers from computer vision, computational geometry, computer graphics, and machine learning in this sixth edition of “Deep Learning for Geometric Computing” workshop at CVPR 2024. The workshop encapsulates competitions with prizes, proceedings, keynotes, paper presentations, and a fair and diverse environment for brainstorming about future research collaborations.

Safe Artificial Intelligence for All Domains

Tuesday, June 18, 9 a.m.-5 p.m., Arch 304

Being able to ensure safety of machine learning (ML) based computer vision is key to unlocking its potential in a broad range of safety related applications and future products. In domains like automotive, aviation and medical applications, it paves the way towards systems with a greater degree of autonomy and assistance for humans. The workshop focuses on bringing together researchers, engineers, and practitioners from academia, industry, and government to exchange ideas, share their latest research, and discuss the latest trends and challenges in this field. The workshop also aims to foster collaboration between different stakeholders, including computer vision researchers, machine learning experts, robotics engineers and safety experts, to create a comprehensive framework for developing safe AI systems for all domains. Overall, the SAIAD workshop aims to advance the state-of-the-art in safe AI, address the most pressing challenges, and provide a platform for networking and knowledge sharing among the experts in this field.

Oral Workshop Papers

BAA-NGP: Bundle-Adjusting Accelerated Neural Graphics Primitives

Implicit neural representations have become pivotal in robotic perception, enabling robots to comprehend 3D environments from 2D images. Given a set of camera poses and associated images, the models can be trained to synthesize novel, unseen views. To successfully navigate and interact in dynamic settings, robots require the understanding of their spatial surroundings driven by unassisted reconstruction of 3D scenes and camera poses from real-time video footage. Existing approaches like COLMAP and bundle-adjusting neural radiance field methods take hours to days to process due to the high computational demands of feature matching, dense point sampling, and training of a multi-layer perceptron structure with a large number of parameters. To address these challenges, we propose a framework called bundle-adjusting accelerated neural graphics primitives (BAA-NGP) which leverages accelerated sampling and hash encoding to expedite automatic pose refinement/estimation and 3D scene reconstruction. Experimental results demonstrate 10 to 20x speed improvement compared to other bundle-adjusting neural radiance field methods without sacrificing the quality of pose estimation.

Block Selective Reprogramming for On-device Training of Vision Transformers

The ubiquity of vision transformers (ViTs) for various edge applications, including personalized learning, has created the demand for on-device fine-tuning. However, training with the limited memory and computation power of edge devices remains a significant challenge. In particular, the memory required for training is much higher than that needed for inference, primarily due to the need to store activations across all layers in order to compute the gradients needed for weight updates. In this paper, we first investigate the limitations of existing on-device training methods aimed at reducing memory and compute requirements. We then present block selective reprogramming (BSR) in which we fine-tune only a fraction of total blocks of a pre-trained model and selectively drop tokens based on self-attention scores of the frozen layers. To show the efficacy of BSR, we present extensive evaluations on ViT-B and DeiT-S with five different datasets. Compared to the existing alternatives, our approach simultaneously reduces training memory by up to 1.4x and compute cost by up to 2x while maintaining similar accuracy. We also showcase results for mixture-of-expert (MoE) models, demonstrating the effectiveness of our approach in multitask learning scenarios.

My Art My Choice: Adversarial Protection Against Unruly AI

Generative AI is on the rise, enabling everyone to produce realistic content via publicly available interfaces. Especially for guided image generation, diffusion models are changing the creator economy by producing high-quality low-cost content. In parallel, artists are rising against unruly AI, since their artwork are leveraged, distributed, and dissimulated by large generative models. Our approach, My Art My Choice (MAMC), aims to empower content owners by protecting their copyrighted materials from being utilized by diffusion models in an adversarial fashion. MAMC learns to generate adversarially perturbed "protected" versions of images which can in turn "break" diffusion models. The perturbation amount is decided by the artist to balance distortion vs. protection of the content. MAMC is designed with a simple UNet-based generator, attacking black box diffusion models, combining several losses to create adversarial twins of the original artwork. We experiment on three datasets for various image-to-image tasks, with different user control values. Both protected image and diffusion output results are evaluated in visual, noise, structure, pixel, and generative spaces to validate our claims. We believe that MAMC is a crucial step for preserving ownership information for AI generated content in a flawless, based-on-need, and human-centric way.

Quantifying Explainability with Multi-Scale Gaussian Mixture Models

With the increasing complexity and influence of machine learning models, the development of model explanation techniques has recently gained significant attention, giving rise to the field of explainable artificial intelligence (XAI). Although there exists vast literature on XAI methods, they are usually compared with human evaluations, model-dependent metrics, or distribution shifts. In the present work, we introduce a novel explainability comparison metric, eXplainable Multi-Scale Gmm Distance (XMGD). XMGD provides a principled probabilistic framework for analyzing and quantifying any model or dataset similarity through the lens of explainability. Through experimental results, we demonstrate several critical advantages of XMGD over alternative saliency comparison metrics, including improved robustness and the ability of XMGD to illuminate fine-grain saliency comparison distinctions.

RLNet: Robust Linearized Networks for Efficient Private Inference

This paper presents RLNet, a class of robust linearized networks that can yield latency improvement via reduction of high-latency ReLU operations while improving the model performance on both clean and corrupted images. In particular, RLNet models provide a "triple win ticket" of improved classification accuracy on clean, naturally perturbed, and gradient-based perturbed images using a shared-mask shared-weight architecture with over an order of magnitude fewer ReLUs than baseline models. To demonstrate the efficacy of RLNet, we perform extensive experiments with ResNet and WRN model variants on CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets. Our experimental evaluations show that RLNet can yield models with up to 11.14x fewer ReLUs, with accuracy close to the all-ReLU models, on clean, naturally perturbed, and gradient-based perturbed images. Compared with the SoTA non-robust linearized models at similar ReLU budgets, RLNet achieves an improvement in adversarial accuracy of up to approximately 47%, naturally perturbed accuracy up to approximately 16.4%, while improving clean image accuracy up to approximately 1.5%.

VideoSAGE: Video Summarization with Graph Representation Learning

We propose a graph-based representation learning framework for video summarization. First, we convert an input video to a graph where nodes correspond to each of the video frames. Then, we impose sparsity on the graph by connecting only those pairs of nodes that are within a specified temporal distance. We then formulate the video summarization task as a binary node classification problem, precisely classifying video frames whether they should belong to the output summary video. A graph constructed this way aims to capture long-range interactions among video frames, and the sparsity ensures the model trains without hitting the memory and compute bottleneck. Experiments on two datasets (SumMe and TVSum) demonstrate the effectiveness of the proposed nimble model compared to existing state-of-the-art summarization approaches while being one order of magnitude more efficient in compute time and memory.

Workshop Papers

Workshop Paper Highlight:

LVLM-InterpreT: An Interpretability Tool for Large Vision-Language Models

In the rapidly evolving landscape of artificial intelligence, multi-modal large language models are emerging as a significant area of interest. These models, which combine various forms of data input, are becoming increasingly popular. However, understanding their internal mechanisms remains a complex task. Numerous advancements have been made in the field of explainability tools and mechanisms, yet there is still much to explore. In this work, we present a novel interactive application aimed towards understanding the internal mechanisms of large vision-language models. Our interface is designed to enhance the interpretability of the image patches, which are instrumental in generating an answer, and assess the efficacy of the language model in grounding its output in the image. With our application, a user can systematically investigate the model and uncover system limitations, paving the way for enhancements in system capabilities. Finally, we present a case study of how our application can aid in understanding failure mechanisms in a popular large multi-modal model: LLaVA.

Contrastive Language Video Time Pre-Training

We introduce LAVITI, a novel approach to learning language, video, and temporal representations in long-form videos via contrastive learning. Different from pretraining on video-text pairs like EgoVLP, LAVITI aims to align language, video, and temporal features by extracting meaningful moments in untrimmed videos. Our model employs a set of learnable moment queries to decode clip-level visual, language, and temporal features. In addition to vision and language alignment, we introduce relative temporal embeddings (TE) to represent timestamps in videos, which enables contrastive learning of time. Significantly different from traditional approaches, the prediction of a particular timestamp is transformed by computing the similarity score between the predicted TE and all TEs. Furthermore, existing approaches for video understanding are mainly designed for short videos due to high computational complexity and memory footprint. Our method can be trained on the Ego4D dataset with only 8 NVIDIA RTX-3090 GPUs in a day. We validated our method on CharadesEgo action recognition, achieving state-of-the-art results.

ICSVR: Investigating Compositional and Syntactic Understanding in Video Retrieval Models

Video retrieval (VR) involves retrieving the ground truth video from the video database given a text caption or vice versa. The two important components of compositionality: objects & attributes and actions are joined using correct syntax to form a proper text query. These components (objects & attributes, actions, and syntax) each play an important role to help distinguish among videos and retrieve the correct ground truth video. However, it is unclear what is the effect of these components on the video retrieval performance. We therefore, conduct a systematic study to evaluate the compositional and syntactic understanding of video retrieval models on standard benchmarks such as MSRVTT, MSVD, and DIDEMO. The study is performed on two categories of video retrieval models: (i) which are pre-trained on video-text pairs and fine-tuned on downstream video retrieval datasets (e.g., Frozen-in-Time, Violet, MCQ, etc.) (ii) which adapt pre-trained image-text representations like CLIP for video retrieval (e.g., CLIP4Clip, XCLIP, CLIP2Video, etc.). Our experiments reveal that actions and syntax play a minor role compared to objects & attributes in video understanding. Moreover, video retrieval models that use pre-trained image-text representations (CLIP) have better syntactic and compositional understanding as compared to models pre-trained on video-text data.

LLaVA-Gemma: Accelerating Multimodal Foundation Models with a Compact Language Model

We train a suite of multimodal foundation models using the popular LLaVA framework with the recently released Gemma family of large language models (LLMs). Of particular interest is the 2B parameter Gemma model, which provides opportunities to construct capable small-scale MMFMs. In line with findings from other papers in this space, we test the effect of ablating three design features: pretraining the connector, utilizing a more powerful image backbone, and increasing the size of the language backbone. The resulting models, which we call LLaVA-Gemma, exhibit moderate performance on an array of evaluations, but fail to improve past the current comparably sized state-of-the-art (SOTA) models. Closer analysis of performance shows mixed effects: skipping pretraining tends to reduce performance, larger vision models sometimes improve performance, and increasing language model size has inconsistent effects.

MixSyn: Learning Composition and Style for Multi-Source Image Synthesis

Synthetic images created by generative models increase in quality and expressiveness as newer models utilize larger datasets and novel architectures. Although this photorealism is a positive side-effect from a creative standpoint, it becomes problematic when such generative models are used for impersonation without consent. Most of these approaches are built on the partial transfer between source and target pairs, or they generate completely new samples based on an ideal distribution, still resembling the closest real sample in the dataset. We propose MixSyn for learning novel fuzzy compositions from multiple sources and creating novel images as a mix of image regions corresponding to the compositions. MixSyn not only combines uncorrelated regions from multiple source masks into a coherent semantic composition, but also generates mask-aware high-quality reconstructions of non-existing images. We compare MixSyn to state-of-the-art single-source sequential generation and collage generation approaches in terms of quality, diversity, realism, and expressive power, while also showcasing interactive synthesis, mix & match, and edit propagation tasks, with no mask dependency.

My Body My Choice: Human-Centric Full-Body Anonymization

In an era of increasing privacy concerns for our online presence, we propose that the decision to appear in a piece of content should only belong to the owner of the body. Although some automatic approaches for fullbody anonymization have been proposed, human-guided anonymization can adapt to various contexts, such as cultural norms, personal relations, esthetic concerns, and security issues. “My Body My Choice” (MBMC) enables physical and adversarial anonymization by removal and swapping approaches aimed for four tasks, designed by single or multi, ControlNet or GAN modules, combining several diffusion models. We evaluate anonymization on seven datasets; compare with SOTA inpainting and anonymization methods; evaluate by image, adversarial, and generative metrics; and conduct reidentification experiments.

NTO3D: Neural Target Object 3D Reconstruction with Segment Anything

Neural 3D reconstruction from multi-view images has recently attracted increasing attention from the community. Existing methods normally learn a neural field for the whole scene, while it is still under-explored how to reconstruct a target object indicated by users. Considering the Segment Anything Model (SAM) has shown effectiveness in segmenting any 2D images, in this paper, we propose NTO3D, a novel high-quality Neural Target Object 3D (NTO3D) reconstruction method, which leverages the benefits of both neural field and SAM. We first propose a novel strategy to lift the multi-view 2D segmentation masks of SAM into a unified 3D occupancy field. The 3D occupancy field is then projected into 2D space and generates the new prompts for SAM. This process is iterative until convergence to separate the target object from the scene. After this, we then lift the 2D features of the SAM encoder into a 3D feature field in order to improve the reconstruction quality of the target object. NTO3D lifts the 2D masks and features of SAM into the 3D neural field for high-quality neural target object 3D reconstruction. We conduct detailed experiments on several benchmark datasets to demonstrate the advantages of our method.

Parameter-Efficient Active Learning for Foundation Models

We present a novel investigation into the application of parameter efficient fine-tuning methods within an active learning (AL) framework, to advance the sampling selection process in extremely budget constrained classification tasks. The focus on image datasets, known for their out-of-distribution characteristics, adds a layer of complexity and relevance to our study. Through a detailed evaluation, we illustrate the improved AL performance on these challenging datasets, highlighting the strategic advantage of merging parameter efficient fine-tuning methods with foundation models. This contributes to the broader discourse on optimizing AL strategies, presenting a promising avenue for future exploration in leveraging foundation models for efficient and effective data annotation in specialized domains.

Situation Monitor: Diversity-Driven Zero-Shot Out-of-Distribution Using Budding Ensemble Architecture in Object Detection

Aiming at facilitating the safety-critical machine learning applications, we present a novel zero shot out-of-distribution (OOD) detection — Situation Monitor for the task of object detection using transformers. This monitor uses the Diversity-based Budding Ensemble Architecture (DBEA) and the monitor aims to enhance the OOD performance by integrating diversity loss into the training process on top of budding ensemble architecture, detecting far-OOD samples, and minimizing false positives on near-OOD samples. Moreover, the utilization of the DBEA loss not only enhances the model’s OOD performance but also improves the calibration of confidence scores, particularly in relation to the intersection of union of the objects. The DBEA model achieves these advancements with a 14% reduction in parameters compared to the vanilla model. This signifies a substantial improvement in efficiency without compromising the model’s ability to accurately detect OOD instances and calibrate the confidence scores.

SocialCounterfactuals: Probing and Mitigating Intersectional Social Biases in Vision-Language Models with Counterfactual Examples

While vision-language models (VLMs) have achieved remarkable performance improvements recently, there is growing evidence that these models also possess harmful biases with respect to social attributes such as gender and race. Prior studies have primarily focused on probing such bias attributes individually while ignoring biases associated with intersections between social attributes. This could be due to the difficulty of collecting an exhaustive set of image-text pairs for various combinations of social attributes. To address this challenge, we employ text-to-image diffusion models to produce counterfactual examples for probing intersectional social biases at scale. Our approach utilizes Stable Diffusion with cross attention control to produce sets of counterfactual image-text pairs that are highly similar in their depiction of a subject (e.g., a given occupation) while differing only in their depiction of intersectional social attributes (e.g., race and gender). Through our over-generate-then-filter methodology, we produce SocialCounterfactuals, a high-quality dataset containing 171,000 image-text pairs for probing intersectional biases related to gender, race, and physical characteristics. We conduct extensive experiments to demonstrate the usefulness of our generated dataset for probing and mitigating intersectional social biases in state-of-the-art VLMs.

Uncovering Bias in Large Vision-Language Models at Scale with Counterfactuals

With the advent of large language models (LLMs) possessing increasingly impressive capabilities, a number of large vision-language models have been proposed to augment LLMs with visual inputs. Such models condition generated text on both an input image and a text prompt, enabling a variety of use cases such as visual question answering and multimodal chat. While prior studies have examined the social biases contained in text generated by LLMs, this topic has been relatively unexplored in LVLMs. Examining social biases in LVLMs is particularly challenging due to the confounding contributions of bias induced by information contained across the text and visual modalities. To address this challenging problem, we conduct a large-scale study of text generated by different LVLMs under counterfactual changes to input images. Specifically, we present LVLMs with identical open-ended text prompts while conditioning on images from different counterfactual sets, where each set contains images which are largely identical in their depiction of a common subject (e.g., a doctor), but vary only in terms of intersectional social attributes (e.g., race and gender). We comprehensively evaluate the text produced by different models under this counterfactual generation setting at scale, producing over 57 million responses from popular LVLMs. Our multi-dimensional analysis reveals that social attributes such as race, gender, and physical characteristics depicted in input images can significantly influence the generation of toxic content, competency-associated words, harmful stereotypes, and numerical ratings of depicted individuals. We additionally explore the relationship between social bias in LVLMs and their corresponding LLMs, as well as inference-time strategies to mitigate bias.

Demonstrations

CLIP-InterpreT: An interpretability tool for CLIP-like models

We investigate the CLIP image encoder by analyzing how individual model components affect the final representation. We decompose the image representation as a sum across individual image patches, model layers, and attention heads, and use CLIP's text representation to interpret the summands. Interpreting the attention heads, we characterize each head's role by automatically finding text representations that span its output space, which reveals property-specific roles for many heads (e.g., location or shape). Next, interpreting the image patches, we uncover an emergent spatial localization within CLIP. Finally, we use this understanding to remove spurious features from CLIP and to create a strong zero-shot image segmenter. Our results indicate that a scalable understanding of transformer models is attainable and can be used to repair and improve models.