OASIS Uncovers: High-Quality T2I Models, Same Old Stereotypes

Images generated by text-to-image (T2I) models often exhibit visual biases and stereotypes of concepts such as culture and profession. Existing quantitative measures of stereotypes are based on statistical parity that does not align with the sociological definition of stereotypes and, therefore, incorrectly categorizes biases as stereotypes. Instead of oversimplifying stereotypes as biases, we propose a quantitative measure of stereotypes that aligns with its sociological definition. We then propose OASIS to measure the stereotypes in a generated dataset and understand their origins within the T2I model. OASIS includes two scores to measure stereotypes from a generated image dataset: (M1) Stereotype Score to measure the distributional violation of stereotypical attributes, and (M2) WALS to measure spectral variance in the images along a stereotypical attribute. OASIS also includes two methods to understand the origins of stereotypes in T2I models: (U1) StOP to discover attributes that the T2I model internally associates with a given concept, and (U2) SPI to quantify the emergence of stereotypical attributes in the latent space of the T2I model during image generation. Despite the considerable progress in image fidelity, using OASIS, we conclude that newer T2I models such as FLUX.1 and SDv3 contain strong stereotypical predispositions about concepts and still generate images with widespread stereotypical attributes. Additionally, the quantity of stereotypes worsens for nationalities with lower Internet footprints.

Fairness and Bias Mitigation in Computer Vision: A Survey

Computer vision systems have witnessed rapid progress over the past two decades due to multiple advances in the field. As these systems are increasingly being deployed in high-stakes real-world applications, there is a dire need to ensure that they do not propagate or amplify any discriminatory tendencies in historical or human-curated data or inadvertently learn biases from spurious correlations. This paper presents a comprehensive survey on fairness that summarizes and sheds light on ongoing trends and successes in the context of computer vision. The topics we discuss include 1) The origin and technical definitions of fairness drawn from the wider fair machine learning literature and adjacent disciplines. 2) Work that sought to discover and analyze biases in computer vision systems. 3) A summary of methods proposed to mitigate bias in computer vision systems in recent years. 4) A comprehensive summary of resources and datasets produced by researchers to measure, analyze, and mitigate bias and enhance fairness. 5) Discussion of the field's success, continuing trends in the context of multimodal foundation and generative models, and gaps that still need to be addressed. The presented characterization should help researchers understand the importance of identifying and mitigating bias in computer vision and the state of the field and identify potential directions for future research.

Utility-Fairness Trade-Offs and How to Find Them

When building classification systems with demographic fairness considerations, there are two objectives to satisfy: 1) maximizing utility for the specific task and 2) ensuring fairness w.r.t. a known demographic attribute. These objectives often compete, so optimizing both can lead to a trade-off between utility and fairness. While existing works acknowledge the trade-offs and study their limits, two questions remain unanswered: 1) What are the optimal trade-offs between utility and fairness? and 2) How can we numerically quantify these trade-offs from data for a desired prediction task and demographic attribute of interest? This paper addresses these questions. We introduce two utility-fairness trade-offs: the Data-Space and Label-Space Trade-off. The trade-offs reveal three regions within the utility-fairness plane, delineating what is fully and partially possible and impossible. We propose U-FaTE, a method to numerically quantify the trade-offs for a given prediction task and group fairness definition from data samples. Based on the trade-offs, we introduce a new scheme for evaluating representations. An extensive evaluation of fair representation learning methods and representations from over 1000 pre-trained models revealed that most current approaches are far from the estimated and achievable fairness-utility trade-offs across multiple datasets and prediction tasks.

FairerCLIP: Debiasing CLIP's Zero-Shot Predictions using Functions in RKHSs

Large pre-trained vision-language models such as CLIP provide compact and general-purpose representations of text and images that are demonstrably effective across multiple downstream zero-shot prediction tasks. However, owing to the nature of their training process, these models have the potential to 1) propagate or amplify societal biases in the training data and 2) learn to rely on spurious features. This paper proposes FairerCLIP, a general approach for making zero-shot predictions of CLIP more fair and robust to spurious correlations. We formulate the problem of jointly debiasing CLIP's image and text representations in reproducing kernel Hilbert spaces (RKHSs), which affords multiple benefits: 1) Flexibility: Unlike existing approaches, which are specialized to either learn with or without ground-truth labels, FairerCLIP is adaptable to learning in both scenarios. 2) Ease of Optimization: FairerCLIP lends itself to an iterative optimization involving closed-form solvers, which leads to $4\times$-$10\times$ faster training than the existing methods. 3) Sample Efficiency: Under sample-limited conditions, FairerCLIP significantly outperforms baselines when they fail entirely. And, 4) Performance: Empirically, FairerCLIP achieves appreciable accuracy gains on benchmark fairness and spurious correlation datasets over their respective baselines.

Mechanics-Informed Autoencoder Enables Automated Detection and Localization of Unforeseen Structural Damage

Structural health monitoring (SHM) is vital for ensuring the safety and longevity of structures like buildings and bridges. As the volume and scale of structures and the impact of their failure continue to grow, there is a dire need for SHM techniques that are scalable, inexpensive, operate passively without human intervention, and customized for each mechanical structure without the need for complex baseline models. We present a novel "deploy-and-forget" approach for automated detection and localization of damages in structures. It is based on a synergistic combination of fully passive measurements from inexpensive sensors and a mechanics-informed autoencoder. Once deployed, our solution continuously learns and adapts a bespoke baseline model for each structure, learning from its undamaged state's response characteristics. After learning from just 3 hours of data, it can autonomously detect and localize different types of unforeseen damage. Results from numerical simulations and experiments indicate that incorporating the mechanical characteristics into the variational autoencoder allows for up to 35\% earlier detection and localization of damage over a standard autoencoder. Our approach holds substantial promise for a significant reduction in human intervention and inspection costs and enables proactive and preventive maintenance strategies, thus extending the lifespan, reliability, and sustainability of civil infrastructures.

AutoFHE: Automated Adaption of CNNs for Efficient Evaluation over FHE

Secure inference of deep convolutional neural networks (CNNs) under RNS-CKKS involves polynomial approximation of unsupported non-linear activation functions. However, existing approaches have three main limitations: 1) Inflexibility: The polynomial approximation and associated homomorphic evaluation architecture are customized manually for each CNN architecture and do not generalize to other networks. 2) Suboptimal Approximation: Each activation function is approximated instead of the function represented by the CNN. 3) Restricted Design: Either high-degree or low-degree polynomial approximations are used. The former retains high accuracy but slows down inference due to bootstrapping operations, while the latter accelerates ciphertext inference but compromises accuracy. To address these limitations, we present AutoFHE, which automatically adapts standard CNNs for secure inference under RNS-CKKS. The key idea is to adopt layerwise mixed-degree polynomial activation functions, which are optimized jointly with the homomorphic evaluation architecture in terms of the placement of bootstrapping operations. The problem is modeled within a multi-objective optimization framework to maximize accuracy and minimize the number of bootstrapping operations. AutoFHE can be applied flexibly on any CNN architecture, and it provides diverse solutions that span the trade-off between accuracy and latency. Experimental evaluation over RNS-CKKS encrypted CIFAR datasets shows that AutoFHE accelerates secure inference by $1.32\times$ to $1.8\times$ compared to methods employing high-degree polynomials. It also improves accuracy by up to 2.56% compared to methods using low-degree polynomials. Lastly, AutoFHE accelerates inference and improves accuracy by $103\times$ and 3.46%, respectively, compared to CNNs under TFHE.

Spurious Correlations and Where to Find Them

Spurious correlations occur when a model learns unreliable features from the data and are a well-known drawback of data-driven learning. Although there are several algorithms proposed to mitigate it, we are yet to jointly derive the indicators of spurious correlations. As a result, the solutions built upon standalone hypotheses fail to beat simple ERM baselines. We collect some of the commonly studied hypotheses behind the occurrence of spurious correlations and investigate their influence on standard ERM baselines using synthetic datasets generated from causal graphs. Subsequently, we observe patterns connecting these hypotheses and model design choices.

Seed Feature Maps-based CNN Models for LEO Satellite Remote Sensing Services

Deploying high-performance convolutional neural network (CNN) models on low-earth orbit (LEO) satellites for rapid remote sensing image processing has attracted significant interest from industry and academia. However, the limited resources available on LEO satellites contrast with the demands of resource-intensive CNN models, necessitating the adoption of ground-station server assistance for training and updating these models. Existing approaches often require large floating-point operations (FLOPs) and substantial model parameter transmissions, presenting considerable challenges. To address these issues, this paper introduces a ground-station server-assisted framework. With the proposed framework, each layer of the CNN model contains only one learnable feature map (called the seed feature map) from which other feature maps are generated based on specific rules. The hyperparameters of these rules are randomly generated instead of being trained, thus enabling the generation of multiple feature maps from the seed feature map and significantly reducing FLOPs. Furthermore, since the random hyperparameters can be saved using a few random seeds, the ground station server assistance can be facilitated in updating the CNN model deployed on the LEO satellite. Experimental results on the ISPRS Vaihingen, ISPRS Potsdam, UAVid, and LoveDA datasets for semantic segmentation services demonstrate that the proposed framework outperforms existing state-of-the-art approaches. In particular, the SineFM-based model achieves a higher mIoU than the UNetFormer on the UAVid dataset, with 3.3x fewer parameters and 2.2x fewer FLOPs.

On the Biometric Capacity of Generative Face Models

There has been tremendous progress in generating realistic faces with high fidelity over the past few years. Despite this progress, a crucial question remains unanswered: "Given a generative face model, how many unique identities can it generate?" In other words, what is the biometric capacity of the generative face model? A scientific basis for answering this question will benefit evaluating and comparing different generative face models and establish an upper bound on their scalability. This paper proposes a statistical approach to estimate the biometric capacity of generated face images in a hyperspherical feature space. We employ our approach on multiple generative models, including unconditional generators like StyleGAN, Latent Diffusion Model, and "Generated Photos," as well as DCFace, a class-conditional generator. We also estimate capacity w.r.t. demographic attributes such as gender and age. Our capacity estimates indicate that (a) under ArcFace representation at a false acceptance rate (FAR) of 0.1%, StyleGAN3 and DCFace have a capacity upper bound of $1.43\times10^6$ and $1.190\times10^4$, respectively; (b) the capacity reduces drastically as we lower the desired FAR with an estimate of $1.796\times10^4$ and $562$ at FAR of 1% and 10%, respectively, for StyleGAN3; (c) there is no discernible disparity in the capacity w.r.t gender; and (d) for some generative models, there is an appreciable disparity in the capacity w.r.t age. Code is available at https://github.com/human-analysis/capacity-generative-face-models.

Mitigating Task Interference in Multi-Task Learning via Explicit Task Routing with Non-Learnable Primitives

Multi-task learning (MTL) seeks to learn a single model to accomplish multiple tasks by leveraging shared information among the tasks. Existing MTL models, however, have been known to suffer from negative interference among tasks. Efforts to mitigate task interference have focused on either loss/gradient balancing or implicit parameter partitioning with partial overlaps among the tasks. In this paper, we propose ETR-NLP to mitigate task interference through a synergistic combination of non-learnable primitives (NLPs) and explicit task routing (ETR). Our key idea is to employ non-learnable primitives to extract a diverse set of task-agnostic features and recombine them into a shared branch common to all tasks and explicit task-specific branches reserved for each task. The non-learnable primitives and the explicit decoupling of learnable parameters into shared and task-specific ones afford the flexibility needed for minimizing task interference. We evaluate the efficacy of ETR-NLP networks for both image-level classification and pixel-level dense prediction MTL problems. Experimental results indicate that ETR-NLP significantly outperforms state-of-the-art baselines with fewer learnable parameters and similar FLOPs across all datasets. Code is available at this \href{https://github.com/zhichao-lu/etr-nlp-mtl}.