Revisiting Multi-Modal LLM Evaluation

With the advent of multi-modal large language models (MLLMs), datasets used for visual question answering (VQA) and referring expression comprehension have seen a resurgence. However, the most popular datasets used to evaluate MLLMs are some of the earliest ones created, and they have many known problems, including extreme bias, spurious correlations, and an inability to permit fine-grained analysis. In this paper, we pioneer evaluating recent MLLMs (LLaVA 1.5, LLaVA-NeXT, BLIP2, InstructBLIP, GPT-4V, and GPT-4o) on datasets designed to address weaknesses in earlier ones. We assess three VQA datasets: 1) TDIUC, which permits fine-grained analysis on 12 question types; 2) TallyQA, which has simple and complex counting questions; and 3) DVQA, which requires optical character recognition for chart understanding. We also study VQDv1, a dataset that requires identifying all image regions that satisfy a given query. Our experiments reveal the weaknesses of many MLLMs that have not previously been reported. Our code is integrated into the widely used LAVIS framework for MLLM evaluation, enabling the rapid assessment of future MLLMs. Project webpage: https://kevinlujian.github.io/MLLM_Evaluations/

They're All Doctors: Synthesizing Diverse Counterfactuals to Mitigate Associative Bias

Vision Language Models (VLMs) such as CLIP are powerful models; however they can exhibit unwanted biases, making them less safe when deployed directly in applications such as text-to-image, text-to-video retrievals, reverse search, or classification tasks. In this work, we propose a novel framework to generate synthetic counterfactual images to create a diverse and balanced dataset that can be used to fine-tune CLIP. Given a set of diverse synthetic base images from text-to-image models, we leverage off-the-shelf segmentation and inpainting models to place humans with diverse visual appearances in context. We show that CLIP trained on such datasets learns to disentangle the human appearance from the context of an image, i.e., what makes a doctor is not correlated to the person's visual appearance, like skin color or body type, but to the context, such as background, the attire they are wearing, or the objects they are holding. We demonstrate that our fine-tuned CLIP model, $CF_\alpha$, improves key fairness metrics such as MaxSkew, MinSkew, and NDKL by 40-66\% for image retrieval tasks, while still achieving similar levels of performance in downstream tasks. We show that, by design, our model retains maximal compatibility with the original CLIP models, and can be easily controlled to support different accuracy versus fairness trade-offs in a plug-n-play fashion.

FairDeDup: Detecting and Mitigating Vision-Language Fairness Disparities in Semantic Dataset Deduplication

Recent dataset deduplication techniques have demonstrated that content-aware dataset pruning can dramatically reduce the cost of training Vision-Language Pretrained (VLP) models without significant performance losses compared to training on the original dataset. These results have been based on pruning commonly used image-caption datasets collected from the web -- datasets that are known to harbor harmful social biases that may then be codified in trained models. In this work, we evaluate how deduplication affects the prevalence of these biases in the resulting trained models and introduce an easy-to-implement modification to the recent SemDeDup algorithm that can reduce the negative effects that we observe. When examining CLIP-style models trained on deduplicated variants of LAION-400M, we find our proposed FairDeDup algorithm consistently leads to improved fairness metrics over SemDeDup on the FairFace and FACET datasets while maintaining zero-shot performance on CLIP benchmarks.

FINEMATCH: Aspect-based Fine-grained Image and Text Mismatch Detection and Correction

Recent progress in large-scale pre-training has led to the development of advanced vision-language models (VLMs) with remarkable proficiency in comprehending and generating multimodal content. Despite the impressive ability to perform complex reasoning for VLMs, current models often struggle to effectively and precisely capture the compositional information on both the image and text sides. To address this, we propose FineMatch, a new aspect-based fine-grained text and image matching benchmark, focusing on text and image mismatch detection and correction. This benchmark introduces a novel task for boosting and evaluating the VLMs' compositionality for aspect-based fine-grained text and image matching. In this task, models are required to identify mismatched aspect phrases within a caption, determine the aspect's class, and propose corrections for an image-text pair that may contain between 0 and 3 mismatches. To evaluate the models' performance on this new task, we propose a new evaluation metric named ITM-IoU for which our experiments show a high correlation to human evaluation. In addition, we also provide a comprehensive experimental analysis of existing mainstream VLMs, including fully supervised learning and in-context learning settings. We have found that models trained on FineMatch demonstrate enhanced proficiency in detecting fine-grained text and image mismatches. Moreover, models (e.g., GPT-4V, Gemini Pro Vision) with strong abilities to perform multimodal in-context learning are not as skilled at fine-grained compositional image and text matching analysis. With FineMatch, we are able to build a system for text-to-image generation hallucination detection and correction.

SCoRD: Subject-Conditional Relation Detection with Text-Augmented Data

We propose Subject-Conditional Relation Detection SCoRD, where conditioned on an input subject, the goal is to predict all its relations to other objects in a scene along with their locations. Based on the Open Images dataset, we propose a challenging OIv6-SCoRD benchmark such that the training and testing splits have a distribution shift in terms of the occurrence statistics of $\langle$subject, relation, object$\rangle$ triplets. To solve this problem, we propose an auto-regressive model that given a subject, it predicts its relations, objects, and object locations by casting this output as a sequence of tokens. First, we show that previous scene-graph prediction methods fail to produce as exhaustive an enumeration of relation-object pairs when conditioned on a subject on this benchmark. Particularly, we obtain a recall@3 of 83.8% for our relation-object predictions compared to the 49.75% obtained by a recent scene graph detector. Then, we show improved generalization on both relation-object and object-box predictions by leveraging during training relation-object pairs obtained automatically from textual captions and for which no object-box annotations are available. Particularly, for $\langle$subject, relation, object$\rangle$ triplets for which no object locations are available during training, we are able to obtain a recall@3 of 42.59% for relation-object pairs and 32.27% for their box locations.

Improving Visual Grounding by Encouraging Consistent Gradient-based Explanations

We propose a margin-based loss for vision-language model pretraining that encourages gradient-based explanations that are consistent with region-level annotations. We refer to this objective as Attention Mask Consistency (AMC) and demonstrate that it produces superior visual grounding performance compared to models that rely instead on region-level annotations for explicitly training an object detector such as Faster R-CNN. AMC works by encouraging gradient-based explanation masks that focus their attention scores mostly within annotated regions of interest for images that contain such annotations. Particularly, a model trained with AMC on top of standard vision-language modeling objectives obtains a state-of-the-art accuracy of 86.59% in the Flickr30k visual grounding benchmark, an absolute improvement of 5.48% when compared to the best previous model. Our approach also performs exceedingly well on established benchmarks for referring expression comprehension and offers the added benefit by design of gradient-based explanations that better align with human annotations.

OccamNets: Mitigating Dataset Bias by Favoring Simpler Hypotheses

Dataset bias and spurious correlations can significantly impair generalization in deep neural networks. Many prior efforts have addressed this problem using either alternative loss functions or sampling strategies that focus on rare patterns. We propose a new direction: modifying the network architecture to impose inductive biases that make the network robust to dataset bias. Specifically, we propose OccamNets, which are biased to favor simpler solutions by design. OccamNets have two inductive biases. First, they are biased to use as little network depth as needed for an individual example. Second, they are biased toward using fewer image locations for prediction. While OccamNets are biased toward simpler hypotheses, they can learn more complex hypotheses if necessary. In experiments, OccamNets outperform or rival state-of-the-art methods run on architectures that do not incorporate these inductive biases. Furthermore, we demonstrate that when the state-of-the-art debiasing methods are combined with OccamNets results further improve.

Learning to Predict Visual Attributes in the Wild

Visual attributes constitute a large portion of information contained in a scene. Objects can be described using a wide variety of attributes which portray their visual appearance (color, texture), geometry (shape, size, posture), and other intrinsic properties (state, action). Existing work is mostly limited to study of attribute prediction in specific domains. In this paper, we introduce a large-scale in-the-wild visual attribute prediction dataset consisting of over 927K attribute annotations for over 260K object instances. Formally, object attribute prediction is a multi-label classification problem where all attributes that apply to an object must be predicted. Our dataset poses significant challenges to existing methods due to large number of attributes, label sparsity, data imbalance, and object occlusion. To this end, we propose several techniques that systematically tackle these challenges, including a base model that utilizes both low- and high-level CNN features with multi-hop attention, reweighting and resampling techniques, a novel negative label expansion scheme, and a novel supervised attribute-aware contrastive learning algorithm. Using these techniques, we achieve near 3.7 mAP and 5.7 overall F1 points improvement over the current state of the art. Further details about the VAW dataset can be found at http://vawdataset.com/.

An Investigation of Critical Issues in Bias Mitigation Techniques

A critical problem in deep learning is that systems learn inappropriate biases, resulting in their inability to perform well on minority groups. This has led to the creation of multiple algorithms that endeavor to mitigate bias. However, it is not clear how effective these methods are. This is because study protocols differ among papers, systems are tested on datasets that fail to test many forms of bias, and systems have access to hidden knowledge or are tuned specifically to the test set. To address this, we introduce an improved evaluation protocol, sensible metrics, and a new dataset, which enables us to ask and answer critical questions about bias mitigation algorithms. We evaluate seven state-of-the-art algorithms using the same network architecture and hyperparameter selection policy across three benchmark datasets. We introduce a new dataset called Biased MNIST that enables assessment of robustness to multiple bias sources. We use Biased MNIST and a visual question answering (VQA) benchmark to assess robustness to hidden biases. Rather than only tuning to the test set distribution, we study robustness across different tuning distributions, which is critical because for many applications the test distribution may not be known during development. We find that algorithms exploit hidden biases, are unable to scale to multiple forms of bias, and are highly sensitive to the choice of tuning set. Based on our findings, we implore the community to adopt more rigorous assessment of future bias mitigation methods. All data, code, and results are publicly available at: https://github.com/erobic/bias-mitigators.

On the Value of Out-of-Distribution Testing: An Example of Goodhart's Law

Out-of-distribution (OOD) testing is increasingly popular for evaluating a machine learning system's ability to generalize beyond the biases of a training set. OOD benchmarks are designed to present a different joint distribution of data and labels between training and test time. VQA-CP has become the standard OOD benchmark for visual question answering, but we discovered three troubling practices in its current use. First, most published methods rely on explicit knowledge of the construction of the OOD splits. They often rely on ``inverting'' the distribution of labels, e.g. answering mostly 'yes' when the common training answer is 'no'. Second, the OOD test set is used for model selection. Third, a model's in-domain performance is assessed after retraining it on in-domain splits (VQA v2) that exhibit a more balanced distribution of labels. These three practices defeat the objective of evaluating generalization, and put into question the value of methods specifically designed for this dataset. We show that embarrassingly-simple methods, including one that generates answers at random, surpass the state of the art on some question types. We provide short- and long-term solutions to avoid these pitfalls and realize the benefits of OOD evaluation.