In this paper, we study the harmlessness alignment problem of multimodal large language models~(MLLMs). We conduct a systematic empirical analysis of the harmlessness performance of representative MLLMs and reveal that the image input poses the alignment vulnerability of MLLMs. Inspired by this, we propose a novel jailbreak method named HADES, which hides and amplifies the harmfulness of the malicious intent within the text input, using meticulously crafted images. Experimental results show that HADES can effectively jailbreak existing MLLMs, which achieves an average Attack Success Rate~(ASR) of 90.26% for LLaVA-1.5 and 71.60% for Gemini Pro Vision. Our code and data will be publicly released.
Recent transformer-based architectures have shown impressive results in the field of image segmentation. Thanks to their flexibility, they obtain outstanding performance in multiple segmentation tasks, such as semantic and panoptic, under a single unified framework. To achieve such impressive performance, these architectures employ intensive operations and require substantial computational resources, which are often not available, especially on edge devices. To fill this gap, we propose Prototype-based Efficient MaskFormer (PEM), an efficient transformer-based architecture that can operate in multiple segmentation tasks. PEM proposes a novel prototype-based cross-attention which leverages the redundancy of visual features to restrict the computation and improve the efficiency without harming the performance. In addition, PEM introduces an efficient multi-scale feature pyramid network, capable of extracting features that have high semantic content in an efficient way, thanks to the combination of deformable convolutions and context-based self-modulation. We benchmark the proposed PEM architecture on two tasks, semantic and panoptic segmentation, evaluated on two different datasets, Cityscapes and ADE20K. PEM demonstrates outstanding performance on every task and dataset, outperforming task-specific architectures while being comparable and even better than computationally-expensive baselines.
Monocular Semantic Occupancy Prediction aims to infer the complete 3D geometry and semantic information of scenes from only 2D images. It has garnered significant attention, particularly due to its potential to enhance the 3D perception of autonomous vehicles. However, existing methods rely on a complex cascaded framework with relatively limited information to restore 3D scenes, including a dependency on supervision solely on the whole network's output, single-frame input, and the utilization of a small backbone. These challenges, in turn, hinder the optimization of the framework and yield inferior prediction results, particularly concerning smaller and long-tailed objects. To address these issues, we propose MonoOcc. In particular, we (i) improve the monocular occupancy prediction framework by proposing an auxiliary semantic loss as supervision to the shallow layers of the framework and an image-conditioned cross-attention module to refine voxel features with visual clues, and (ii) employ a distillation module that transfers temporal information and richer knowledge from a larger image backbone to the monocular semantic occupancy prediction framework with low cost of hardware. With these advantages, our method yields state-of-the-art performance on the camera-based SemanticKITTI Scene Completion benchmark. Codes and models can be accessed at https://github.com/ucaszyp/MonoOcc
In image-based robot manipulation tasks with large observation and action spaces, reinforcement learning struggles with low sample efficiency, slow training speed, and uncertain convergence. As an alternative, large pre-trained foundation models have shown promise in robotic manipulation, particularly in zero-shot and few-shot applications. However, using these models directly is unreliable due to limited reasoning capabilities and challenges in understanding physical and spatial contexts. This paper introduces ExploRLLM, a novel approach that leverages the inductive bias of foundation models (e.g. Large Language Models) to guide exploration in reinforcement learning. We also exploit these foundation models to reformulate the action and observation spaces to enhance the training efficiency in reinforcement learning. Our experiments demonstrate that guided exploration enables much quicker convergence than training without it. Additionally, we validate that ExploRLLM outperforms vanilla foundation model baselines and that the policy trained in simulation can be applied in real-world settings without additional training.
In the realm of medical imaging, particularly for COVID-19 detection, deep learning models face substantial challenges such as the necessity for extensive computational resources, the paucity of well-annotated datasets, and a significant amount of unlabeled data. In this work, we introduce the first lightweight detector designed to overcome these obstacles, leveraging a frozen CLIP image encoder and a trainable multilayer perception (MLP). Enhanced with Conditional Value at Risk (CVaR) for robustness and a loss landscape flattening strategy for improved generalization, our model is tailored for high efficacy in COVID-19 detection. Furthermore, we integrate a teacher-student framework to capitalize on the vast amounts of unlabeled data, enabling our model to achieve superior performance despite the inherent data limitations. Experimental results on the COV19-CT-DB dataset demonstrate the effectiveness of our approach, surpassing baseline by up to 10.6% in `macro' F1 score in supervised learning. The code is available at https://github.com/Purdue-M2/COVID-19_Detection_M2_PURDUE.
This research paper presents an innovative multi-task learning framework that allows concurrent depth estimation and semantic segmentation using a single camera. The proposed approach is based on a shared encoder-decoder architecture, which integrates various techniques to improve the accuracy of the depth estimation and semantic segmentation task without compromising computational efficiency. Additionally, the paper incorporates an adversarial training component, employing a Wasserstein GAN framework with a critic network, to refine model's predictions. The framework is thoroughly evaluated on two datasets - the outdoor Cityscapes dataset and the indoor NYU Depth V2 dataset - and it outperforms existing state-of-the-art methods in both segmentation and depth estimation tasks. We also conducted ablation studies to analyze the contributions of different components, including pre-training strategies, the inclusion of critics, the use of logarithmic depth scaling, and advanced image augmentations, to provide a better understanding of the proposed framework. The accompanying source code is accessible at \url{https://github.com/PardisTaghavi/SwinMTL}.
Performance disparities across sub-populations are known to exist in deep learning-based vision recognition models, but previous work has largely addressed such fairness concerns assuming knowledge of sensitive attribute labels. To overcome this reliance, previous strategies have involved separate learning structures to expose and adjust for disparities. In this work, we explore a new paradigm that does not require sensitive attribute labels, and evades the need for extra training by leveraging the vision-language model, CLIP, as a rich knowledge source to infer sensitive information. We present sample clustering based on similarity derived from image and attribute-specified language embeddings and assess their correspondence to true attribute distribution. We train a target model by re-sampling and augmenting under-performed clusters. Extensive experiments on multiple benchmark bias datasets show clear fairness gains of the model over existing baselines, which indicate that CLIP can extract discriminative sensitive information prompted by language, and used to promote model fairness.
User Interface (UI) understanding has been an increasingly popular topic over the last few years. So far, there has been a vast focus solely on web and mobile applications. In this paper, we introduce the harder task of computer UI understanding. With the goal of enabling research in this field, we have generated a dataset with a set of videos where a user is performing a sequence of actions and each image shows the desktop contents at that time point. We also present a framework that is composed of a synthetic sample generation pipeline to augment the dataset with relevant characteristics, and a contrastive learning method to classify images in the videos. We take advantage of the natural conditional, tree-like, relationship of the images' characteristics to regularize the learning of the representations by dealing with multiple partial tasks simultaneously. Experimental results show that the proposed framework outperforms previously proposed hierarchical multi-label contrastive losses in fine-grain UI classification.
Deep learning (DL) models have been advancing automatic medical image analysis on various modalities, including echocardiography, by offering a comprehensive end-to-end training pipeline. This approach enables DL models to regress ejection fraction (EF) directly from 2D+time echocardiograms, resulting in superior performance. However, the end-to-end training pipeline makes the learned representations less explainable. The representations may also fail to capture the continuous relation among echocardiogram clips, indicating the existence of spurious correlations, which can negatively affect the generalization. To mitigate this issue, we propose CoReEcho, a novel training framework emphasizing continuous representations tailored for direct EF regression. Our extensive experiments demonstrate that CoReEcho: 1) outperforms the current state-of-the-art (SOTA) on the largest echocardiography dataset (EchoNet-Dynamic) with MAE of 3.90 & R2 of 82.44, and 2) provides robust and generalizable features that transfer more effectively in related downstream tasks. The code is publicly available at https://github.com/fadamsyah/CoReEcho.
Semi-supervised learning (SSL) seeks to enhance task performance by training on both labeled and unlabeled data. Mainstream SSL image classification methods mostly optimize a loss that additively combines a supervised classification objective with a regularization term derived solely from unlabeled data. This formulation neglects the potential for interaction between labeled and unlabeled images. In this paper, we introduce InterLUDE, a new approach to enhance SSL made of two parts that each benefit from labeled-unlabeled interaction. The first part, embedding fusion, interpolates between labeled and unlabeled embeddings to improve representation learning. The second part is a new loss, grounded in the principle of consistency regularization, that aims to minimize discrepancies in the model's predictions between labeled versus unlabeled inputs. Experiments on standard closed-set SSL benchmarks and a medical SSL task with an uncurated unlabeled set show clear benefits to our approach. On the STL-10 dataset with only 40 labels, InterLUDE achieves 3.2% error rate, while the best previous method reports 14.9%.