Large Language Models (LLMs) have emerged as a new information channel. Meanwhile, one critical but under-explored question is: Is it possible to bypass the safety alignment and inject harmful information into LLMs stealthily? In this paper, we propose to reformulate knowledge editing as a new type of safety threat for LLMs, namely Editing Attack, and conduct a systematic investigation with a newly constructed dataset EditAttack. Specifically, we focus on two typical safety risks of Editing Attack including Misinformation Injection and Bias Injection. For the first risk, we find that editing attacks can inject both commonsense and long-tail misinformation into LLMs, and the effectiveness for the former one is particularly high. For the second risk, we discover that not only can biased sentences be injected into LLMs with high effectiveness, but also one single biased sentence injection can degrade the overall fairness. Then, we further illustrate the high stealthiness of editing attacks. Our discoveries demonstrate the emerging misuse risks of knowledge editing techniques on compromising the safety alignment of LLMs and the feasibility of disseminating misinformation or bias with LLMs as new channels.

Multimodal Large Language Models (MLLMs) achieve impressive performance once optimized on massive datasets. Such datasets often contain sensitive or copyrighted content, raising significant data privacy concerns. Regulatory frameworks mandating the 'right to be forgotten' drive the need for machine unlearning. This technique allows for the removal of target data without resource-consuming retraining. However, while well-studied for text, visual concept unlearning in MLLMs remains underexplored. A primary challenge is precisely removing a target visual concept without disrupting model performance on related entities. To address this, we introduce AUVIC, a novel visual concept unlearning framework for MLLMs. AUVIC applies adversarial perturbations to enable precise forgetting. This approach effectively isolates the target concept while avoiding unintended effects on similar entities. To evaluate our method, we construct VCUBench. It is the first benchmark designed to assess visual concept unlearning in group contexts. Experimental results demonstrate that AUVIC achieves state-of-the-art target forgetting rates while incurs minimal performance degradation on non-target concepts.

Composed image retrieval aims to retrieve images based on a query that consists of a reference image and text describing desired modifications to that image. It has recently attracted attention for its ability to tailor image retrieval to user intentions by combining information-rich reference images with concise natural language instructions. Despite its current success, the robustness of composed image retrieval methods to either (1) common corruptions or (2) variations of the textual descriptions have never been systematically evaluated. In this paper, we perform the first robustness study of composed image retrieval, establishing three new benchmarks for a systematic evaluation of robustness to common corruption (in both the textual and visual domains) and robustness in text understanding. For analysis of natural image corruption, we introduce two new large-scale benchmark datasets, CIRR-C and FashionIQ-C, for the open domains and fashion domains respectively–both of which feature 75 visual corruptions and 35 textual corruptions. To facilitate robust evaluation of text understanding, we introduce a new diagnostic dataset CIRR-D by expanding the CIRR dataset with synthetic data, specifically probing text understanding across variations in: numerical, attribute, object removal, and background. We introduce BenchCIR, a testbed for evaluating composed image retrieval model robustness with standardized evaluation protocols. Through benchmarking ten published models in the testbed, we reveal insights into how the composition of visual and textual modalities affects model robustness. The code is in https://suntongtongtong.github.io/BenchCIR/

Federated Learning (FL) is a distributed machine learning (ML) paradigm, in which multiple clients collaboratively train ML models without centralizing their local data. Similar to conventional ML pipelines, the client local optimization and server aggregation procedure in FL are sensitive to the hyperparameter (HP) selection. Despite extensive research on tuning HPs for centralized ML, these methods yield suboptimal results when employed in FL. This is mainly because their "training-after-tuning" framework is unsuitable for FL with limited client computation power. While some approaches have been proposed for HP-Tuning in FL, they are limited to the HPs for client local updates. In this work, we propose a novel HP-tuning algorithm, called Federated Population-based Hyperparameter Tuning (FedPop), to address this vital yet challenging problem. FedPop employs population-based evolutionary algorithms to optimize the HPs, which accommodates various HP types at both the client and server sides. Compared with prior tuning methods, FedPop employs an online "tuning-while-training" framework, offering computational efficiency and enabling the exploration of a broader HP search space. Our empirical validation on the common FL benchmarks and complex real-world FL datasets, including full-sized Non-IID ImageNet-1K, demonstrates the effectiveness of the proposed method, which substantially outperforms the concurrent state-of-the-art HP-tuning methods in FL.

Tong Liu, Zhixin Lai, Jiawen Wang, Gengyuan Zhang, Shuo Chen, Philip Torr, Vera Demberg, Volker Tresp, Jindong Gu. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Large Audio-Language Models (LALMs), such as GPT-4o, have recently unlocked audio dialogue capabilities, enabling direct spoken exchanges with humans.The potential of LALMs broadens their applicability across a wide range of practical scenarios supported by audio dialogues.However, given these advancements, a comprehensive benchmark to evaluate the performance of LALMs in the open-ended audio dialogue understanding remains absent currently.To address this gap, we propose an Audio Dialogue Understanding Benchmark (ADU-Bench), which consists of 4 benchmark datasets.They assess the open-ended audio dialogue ability for LALMs in 3 general scenarios, 12 skills, 9 multilingual languages, and 4 categories of ambiguity handling.Notably, we firstly propose the evaluation of ambiguity handling in audio dialogues that expresses different intentions beyond the same literal meaning of sentences, e.g., "Really!?" with different intonations.In summary, ADU-Bench includes over 20,000 open-ended audio dialogues for the assessment of LALMs.Through extensive experiments on 16 LALMs, our analysis reveals that existing LALMs struggle with mathematical symbols and formulas, understanding human behavior such as roleplay, comprehending multiple languages, and handling audio dialogue ambiguities from different phonetic elements, such as intonations, pause positions, and homophones.The benchmark is available at https://adu-bench.github.io/.

Fan Yin, Zifeng Wang, I-Hung Hsu, Jun Yan, Ke Jiang, Yanfei Chen, Jindong Gu, Long Le, Kai-Wei Chang, Chen-Yu Lee, Hamid Palangi, Tomas Pfister. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Predicting in-hospital mortality for ICU patients is critical for improving clinical outcomes. Deep learning models have achieved remarkable accuracy but often lack explainability. To address this, we propose$X-M M P$, an eXplainable Multimodal Mortality Predictor based on transformers for heterogeneous ICU data. X-MMP integrates tabular time-series, vital signs, and clinical notes, and introduces a Layer-Wise Relevance Propagation (LRP) method to generate interpretable saliency maps across modalities. Evaluated on the multimodal MIMICIII dataset, X-MMP achieves competitive predictive performance and generates clinically meaningful explanations. Code: https://github.com/lixingqiao/XAI-ICU.

One-Shot Federated Learning (OSFL), a special decentralized machine learning paradigm, has recently gained significant attention. OSFL requires only a single round of client data or model upload, which reduces communication costs and mitigates privacy threats compared to traditional FL. Despite these promising prospects, existing methods face challenges due to client data heterogeneity and limited data quantity when applied to real-world OSFL systems. Recently, Latent Diffusion Models (LDM) have shown remarkable advancements in synthesizing high-quality images through pretraining on large-scale datasets, thereby presenting a potential solution to overcome these issues. However, directly applying pretrained LDM to heterogeneous OSFL results in significant distribution shifts in synthetic data, leading to performance degradation in classification models trained on such data. This issue is particularly pronounced in rare domains, such as medical imaging, which are underrepresented in LDM’s pretraining data. To address this challenge, we propose Federated Bi-Level Personalization (FedBiP), which personalizes the pretrained LDM at both instance-level and concept-level. Hereby, FedBiP synthesizes images following the client’s local data distribution without compromising the privacy regulations. FedBiP is also the first approach to simultaneously address feature space heterogeneity and client data scarcity in OSFL. Our method is validated through extensive experiments on three OSFL benchmarks with feature space heterogeneity, as well as on challenging medical and satellite image datasets with label heterogeneity. The results demonstrate the effectiveness of FedBiP, which substantially outperforms other OSFL methods. Our code is available at https://github.com/HaokunChen245/FedBiP.

Current image generation models can effortlessly produce high-quality, highly realistic images, but this also increases the risk of misuse. In various Text-to-Image or Image-to-Image tasks, attackers can generate a series of images containing inappropriate content by simply editing the language modality input. To mitigate this security concern, numerous guarding or defensive strategies have been proposed, with a particular emphasis on safeguarding language modality. However, in practical applications, threats in the vision modality, particularly in tasks involving the editing of real-world images, present heightened security risks as they can easily infringe upon the rights of the image owner. Therefore, this paper employs a method named typographic attack to reveal that various image generation models are also susceptible to threats within the vision modality. Furthermore, we also evaluate the defense performance of various existing methods when facing threats in the vision modality and uncover their ineffectiveness. Finally, we propose the Vision Modal Threats in Image Generation Models (VMT-IGMs) dataset, which would serve as a baseline for evaluating the vision modality vulnerability of various image generation models.Warning: This paper includes content that may cause discomfort or distress. Potentially disturbing content has been blocked and blurred.

Large language models (LLMs) excel at retrieving information from lengthy text, but their vision-language counterparts (VLMs) face difficulties with hour-long videos, especially for temporal grounding. Specifically, these VLMs are constrained by frame limitations, often losing essential temporal details needed for accurate event localization in extended video content. We propose ReVisionLLM, a recursive vision-language model designed to locate events in hour-long videos. Inspired by human search strategies, our model initially targets broad segments of interest, progressively revising its focus to pinpoint exact temporal boundaries. Our model can seamlessly handle videos of vastly different lengths—from minutes to hours. We also introduce a hierarchical training strategy that starts with short clips to capture distinct events and progressively extends to longer videos. To our knowledge, ReVision-LLM is the first VLM capable of temporal grounding in hour-long videos, outperforming previous state-of-the-art methods across multiple datasets by a significant margin (e.g., +2.6% [email protected] on MAD). The code is available at https://github.com/Tanveer81/ReVisionLLM

Localizing events in videos based on semantic queries is a pivotal task in video understanding research and user-oriented applications like video search. Yet, current research predominantly relies on natural language queries (NLQs), overlooking the potential of using multimodal queries (MQs) that incorporate images to flexibly represent semantic queries, particularly when it is difficult to express non-verbal or unfamiliar concepts in words. To bridge this gap, we introduce ICQ, a new benchmark designed for localizing events in videos with MQs, alongside an evaluation dataset ICQ-Highlight. To adapt and reevaluate existing video localization models for this new task, we propose 3 Multimodal Query Adaptation methods and a novel Surrogate Fine-Tuning strategy, serving as strong baseline methods. ICQ systematically benchmarks 12 state-of-the-art backbone models, spanning from specialized video localization models to Video Large Language Models. Our extensive experiments highlight the high potential of using MQs in real-world applications. We believe this is a first step toward video event localization with MQs1.

Prompt sensitivity, referring to the phenomenon where paraphrasing (i.e., repeating something written or spoken using different words) leads to significant changes in large language model (LLM) performance, has been widely accepted as a core limitation of LLMs.In this work, we revisit this issue and ask: Is the widely reported high prompt sensitivity truly an inherent weakness of LLMs, or is it largely an artifact of evaluation processes?To answer this question, we systematically evaluate 7 LLMs (e.g., GPT and Gemini family) across 6 benchmarks, including both multiple-choice and open-ended tasks on 12 diverse prompt templates.We find that much of the prompt sensitivity stems from heuristic evaluation methods, including log-likelihood scoring and rigid answer matching, which often overlook semantically correct responses expressed through alternative phrasings, such as synonyms or paraphrases.When we adopt LLM-as-a-Judge evaluations, we observe a substantial reduction in performance variance and a consistently higher correlation in model rankings across prompts.Our findings suggest that modern LLMs are more robust to prompt templates than previously believed, and that prompt sensitivity may be more an artifact of evaluation than a flaw in the models.

Individual Large Language Models (LLMs) have demonstrated significant capabilities across various domains, such as healthcare and law.Recent studies also show that coordinated multi-agent systems exhibit enhanced decisionmaking and reasoning abilities through collaboration.However, due to the vulnerabilities of individual LLMs and the difficulty of accessing all agents in a multi-agent system, a key question arises: If attackers only know one agent, could they still generate adversarial samples capable of misleading the collective decision?To explore this question, we formulate it as a game with incomplete information, where attackers know only one target agent and lack knowledge of the other agents in the system.With this formulation, we propose M-Spoiler, a framework that simulates agent interactions within a multi-agent system to generate adversarial samples.These samples are then used to manipulate the target agent in the target system, misleading the system's collaborative decisionmaking process.More specifically, M-Spoiler introduces a stubborn agent that actively aids in optimizing adversarial samples by simulating potential stubborn responses from agents in the target system.This enhances the effectiveness of the generated adversarial samples in misleading the system.Through extensive experiments across various tasks, our findings confirm the risks posed by the knowledge of an individual agent in multi-agent systems and demonstrate the effectiveness of our framework.We also explore several defense mechanisms, showing that our proposed attack framework remains more potent than baselines, underscoring the need for further research into defensive strategies.Our source code is available at here.

Mihir Parmar, Xin Liu, Palash Goyal, Yanfei Chen, Long Le, Swaroop Mishra, Hossein Mobahi, Jindong Gu, Zifeng Wang, Hootan Nakhost, Chitta Baral, Chen-Yu Lee, Tomas Pfister, Hamid Palangi. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Large language models (LLMs) excel in diverse applications but face dual challenges: generating harmful content under jailbreak attacks and over-refusing benign queries due to rigid safety mechanisms.These issues severely affect the application of LLMs.Existing approaches can be divided into three types: contrastive decoding, activation manipulation, and prompting strategies.However, all these approaches face challenges like inefficiency, fragility, or architectural constraints, ultimately failing to strike a balance between safety and usability.These problems are more obvious in multimodal large language models (MLLMs), especially in terms of heightened over-refusal in cross-modal tasks and new security risks arising from expanded attack surfaces.We propose Magic Image 1 , an optimization-driven visual prompt framework that enhances security and reduces over-refusal at the same time.The Magic Image is optimized using gradients derived from harmful/benign training samples.Using the magic image can modify the model's original safety alignment, maintaining robust safety while reducing unnecessary denials.Experiments demonstrate its effectiveness in preserving model performance and improving safety-responsiveness balance across datasets, including unseen data, offering a practical solution for reliable MLLM deployment.

Recent advancements in diffusion-based text-to-image (T2I) models have enabled the generation of high-quality and photorealistic images from text. However, they often exhibit societal biases related to gender, race, and socioeconomic status, thereby potentially reinforcing harmful stereotypes and shaping public perception in unintended ways. While existing bias mitigation methods demonstrate effectiveness, they often encounter attribute entanglement, where adjustments to attributes relevant to the bias (i.e., target attributes) unintentionally alter attributes unassociated with the bias (i.e., non-target attributes), causing undesirable distribution shifts. To address this challenge, we introduce Entanglement-Free Attention (EFA), a method that accurately incorporates target attributes (e.g., White, Black, and Asian) while preserving non-target attributes (e.g., background) during bias mitigation. At inference time, EFA randomly samples a target attribute with equal probability and adjusts the cross-attention in selected layers to incorporate the sampled attribute, achieving a fair distribution of target attributes. Extensive experiments demonstrate that EFA outperforms existing methods in mitigating bias while preserving non-target attributes, thereby maintaining the original model's output distribution and generative capacity.

Large language models (LLMs) are being rapidly developed, and a key component of their widespread deployment is their safety-related alignment. Many red-teaming efforts aim to jailbreak LLMs, where among these efforts, the Greedy Coordinate Gradient (GCG) attack's success has led to a growing interest in the study of optimization-based jailbreaking techniques. Although GCG is a significant milestone, its attacking efficiency remains unsatisfactory. In this paper, we present several improved (empirical) techniques for optimization-based jailbreaks like GCG. We first observe that the single target template of ”Sure'' largely limits the attacking performance of GCG; given this, we propose to apply diverse target templates containing harmful self-suggestion and/or guidance to mislead LLMs. Besides, from the optimization aspects, we propose an automatic multi-coordinate updating strategy in GCG (i.e., adaptively deciding how many tokens to replace in each step) to accelerate convergence, as well as tricks like easy-to-hard initialization. Then, we combine these improved technologies to develop an efficient jailbreak method, dubbed $\mathcal{I}$-GCG. In our experiments, we evaluate our $\mathcal{I}$-GCG on a series of benchmarks (such as NeurIPS 2023 Red Teaming Track). The results demonstrate that our improved techniques can help GCG outperform state-of-the-art jailbreaking attacks and achieve a nearly 100\% attack success rate. The code is released at https://github.com/jiaxiaojunQAQ/I-GCG.

Mathematical diagrams have a distinctive structure. Standard feature transforms designed for natural images (e.g., CLIP) fail to process them effectively, limiting their utility in multimodal large language models (MLLMs). Current efforts to improve MLLMs have primarily focused on scaling mathematical visual instruction datasets and strengthening LLM backbones, yet fine-grained visual recognition errors remain unaddressed. Our systematic evaluation on the visual grounding capabilities of state-of-the-art MLLMs highlights that fine-grained visual understanding remains a crucial bottleneck in visual mathematical reasoning (GPT-4o exhibits a 70% grounding error rate, and correcting these errors improves reasoning accuracy by 12%). We thus propose a novel approach featuring a geometrically-grounded vision encoder and a feature router that dynamically selects between hierarchical visual feature maps. Our model accurately recognizes visual primitives and generates precise visual prompts aligned with the language model’s reasoning needs. In experiments, PRIMITIVE-Qwen2.5-7B outperforms other 7B models by 12% on MathVerse and is on par with GPT-4V on MathVista. Our findings highlight the need for better fine-grained visual integration in MLLMs. Code is available at github.com/AI4Math-ShanZhang/SVE-Math.

Multimodal Large Language Models (MLLMs) demonstrate exceptional performance in cross-modality interaction, yet they also suffer adversarial vulnerabilities. In particular, the transferability of adversarial examples remains an ongoing challenge. In this paper, we specifically analyze the manifestation of adversarial transferability among MLLMs and identify the key factors that influence this characteristic. We discover that the transferability of MLLMs exists in cross-LLM scenarios with the same vision encoder and indicate two key Factors that may influence transferability. We provide two semantic-level data augmentation methods, Adding Image Patch (AIP) and Typography Augment Transferability Method (TATM), which boost the transferability of adversarial examples across MLLMs. To explore the potential impact in the real world, we utilize two tasks that can have both negative and positive societal impacts: 1. Harmful Content Insertion and 2. Information Protection.

Large Language Models (LLMs) demonstrate impressive zero-shot performance across a wide range of natural language processing tasks. Integrating various modality encoders further expands their capabilities, giving rise to Multimodal Large Language Models (MLLMs) that process not only text but also visual and auditory modality inputs. However, these advanced capabilities may also pose significant safety problems, as models can be exploited to generate harmful or inappropriate content through jailbreak attack. While prior work has extensively explored how manipulating textual or visual modality inputs can circumvent safeguards in LLMs and MLLMs, the vulnerability of audio-specific Jailbreak on Large Audio-Language Models (LALMs) remains largely underexplored. To address this gap, we introduce \textbf{Jailbreak-AudioBench}, which consists of the Toolbox, curated Dataset, and comprehensive Benchmark. The Toolbox supports not only text-to-audio conversion but also various editing techniques for injecting audio hidden semantics. The curated Dataset provides diverse explicit and implicit jailbreak audio examples in both original and edited forms. Utilizing this dataset, we evaluate multiple state-of-the-art LALMs and establish the most comprehensive Jailbreak benchmark to date for audio modality. Finally, Jailbreak-AudioBench establishes a foundation for advancing future research on LALMs safety alignment by enabling the in-depth exposure of more powerful jailbreak threats, such as query-based audio editing, and by facilitating the development of effective defense mechanisms.

We introduce REAL, a benchmark and framework for multi-turn agent evaluations on deterministic simulations of real-world websites. REAL comprises high-fidelity, deterministic replicas of 11 widely-used websites across domains such as e-commerce, travel, communication, and professional networking. We also release a benchmark consisting of 112 practical tasks that mirror everyday complex user interactions requiring both accurate information retrieval and state-changing actions. All interactions occur within this fully controlled setting, eliminating safety risks and enabling robust, reproducible evaluation of agent capability and reliability. Our novel evaluation framework combines programmatic checks of website state for action-based tasks with rubric-guided LLM-based judgments for information retrieval. The framework supports both open-source and proprietary agent systems through a flexible evaluation harness that accommodates black-box commands within browser environments, allowing research labs to test agentic systems without modification. Our empirical results show that frontier language models achieve at most a 41% success rate on REAL, highlighting critical gaps in autonomous web navigation and task completion capabilities. Our framework supports easy integration of new tasks, reproducible evaluation, and scalable post-training data generation, marking a significant step forward in evaluating and advancing agent capabilities.

Large Vision-Language Models (LVLMs) with discrete image tokenizers unify multimodal representations by encoding visual inputs into a finite set of tokens. Despite their effectiveness, we find that these models still hallucinate non-existent objects. We hypothesize that one reason is due to visual priors induced during training: when certain image tokens frequently co-occur in the same spatial regions and represent shared objects, they become strongly associated with the verbalizations of those objects. As a result, the model may hallucinate by evoking visually absent tokens that often co-occur with present ones. To test this assumption, we construct a co-occurrence graph of image tokens using a segmentation dataset and employ a Graph Neural Network (GNN) with contrastive learning followed by a clustering method to group tokens that frequently co-occur in similar visual contexts. We find that hallucinations predominantly correspond to clusters whose tokens dominate the input, and more specifically, that the visually absent tokens in those clusters show much higher correlation with hallucinated objects compared to tokens present in the image. Based on this observation, we propose a hallucination mitigation method that suppresses the influence of visually absent tokens by modifying latent image embeddings during generation. Experiments show our method reduces hallucinations while preserving expressivity.

Retrieval-Augmented Generation (RAG) based on knowledge graphs (KGs) enhances large language models (LLMs) by providing structured and interpretable external knowledge. However, existing KG-based RAG methods struggle to retrieve accurate and diverse information from text-rich KGs for complex real-world queries. Process Reward Models (PRMs) offer a way to align the retrieval process of KG-based RAG with query-specific knowledge requirements, but they heavily rely on process-level supervision signals that are expensive and hard to obtain on KGs. To address this challenge, we propose GraphFlow, a framework that efficiently retrieves accurate and diverse knowledge required for real-world queries from text-rich KGs. GraphFlow employs a transition-based flow matching objective to jointly optimize a retrieval policy and a flow estimator. The flow estimator factorizes the reward of the retrieval outcome into the intermediate retrieval states. Such reward factorization guides the retrieval policy to retrieve candidates from KGs in proportion to their reward. This allows GraphFlow to explore high-quality regions of KGs that yield diverse and relevant results. We evaluate GraphFlow on the STaRK benchmark, which includes real-world queries from multiple domains over text-rich KGs. GraphFlow outperforms strong KG-RAG baselines, including GPT-4o, by 10\% on average in hit rate and recall. It also shows strong generalization to unseen KGs, demonstrating its effectiveness and robustness.

Large Language Models (LLMs) with in-context learning (ICL) ability can quickly adapt to a specific context given a few demonstrations (demos). Recently, Multimodal Large Language Models (MLLMs) built upon LLMs have also shown multimodal ICL ability, i.e., responding to queries given a few multimodal demos, including images, queries, and answers. While ICL has been extensively studied on LLMs, its research on MLLMs remains limited. One essential question is whether these MLLMs can truly conduct multimodal ICL, or if only the textual modality is necessary. We investigate this question by examining two primary factors that influence ICL: 1) Demo content, i.e., understanding the influences of demo content in different modalities. 2) Demo selection strategy, i.e., how to select better multimodal demos for improved performance. Experiments revealed that multimodal ICL is predominantly driven by the textual content whereas the visual information in the demos has little influence. Interestingly, visual content is still necessary and useful for selecting demos to increase performance. Motivated by our analysis, we propose a simple yet effective approach, termed Mixed Modality In-Context Example Selection (MMICES), which considers both visual and language modalities when selecting demos. Extensive experiments are conducted to support our findings and verify the improvement brought by our method.

Visual Question Answering (VQA) systems witnessed a significant advance in recent years due to the development of large-scale Vision-Language Pre-trained Models (VLPMs). As the application scenario and user demand change over time, an advanced VQA system is expected to be capable of continuously expanding its knowledge and capabilities over time, not only to handle new tasks (i.e., new question types or visual scenes) but also to answer questions in new specialized domains without forgetting previously acquired knowledge and skills. Existing works studying CL on VQA tasks primarily consider answer-and question-type incremental learning or sceneand function-incremental learning, whereas how VQA systems perform when they encounter new domains and increasing user demands has not been studied. Motivated by this, we introduce CL-CrossVQA, a rigorous Continual Learning benchmark for Cross-domain Visual Question Answering, through which we conduct extensive experiments on 4 VLPMs, 5 CL approaches, and 5 VQA datasets from different domains. In addition, by probing the forgetting phenomenon of the intermediate layers, we provide insights into how model architecture affects CL performance, why CL approaches can help mitigate forgetting in VLPMs, and how to design CL approaches suitable for VLPMs in this challenging continual learning environment. To facilitate future work on developing an advanced All-in-One VQA system, we will release our datasets and code.

With the increasing penetration of machine learning applications in critical decision-making areas, calls for algorithmic fairness are more prominent. Although there have been various modalities to improve algorithmic fairness through learning with fairness constraints, their performance does not generalize well in the test set. A performance-promising fair algorithm with better generalizability is needed. This article proposes a novel adaptive reweighing method to eliminate the impact of the distribution shifts between training and test data on model generalizability. Most previous reweighing methods propose to assign a unified weight for each (sub)group. Rather, our method granularly models the distance from the sample predictions to the decision boundary. Our adaptive reweighing method prioritizes samples closer to the decision boundary and assigns a higher weight to improve the generalizability of fair classifiers. Extensive experiments are performed to validate the generalizability of our adaptive priority reweighing method for accuracy and fairness measures (i.e., equal opportunity, equalized odds, and demographic parity) in tabular benchmarks. We also highlight the performance of our method in improving the fairness of language and vision models. The code is available at https://github.com/che2198/APW .

Recently, foundation models have exhibited remarkable advancements in multi-modal learning. These models, equipped with millions (or billions) of parameters, typically require a substantial amount of data for finetuning. However, collecting and centralizing training data from diverse sectors becomes challenging due to distinct privacy regulations. Federated Learning (FL) emerges as a promising solution, enabling multiple clients to collaboratively train neural networks without centralizing their local data. To alleviate client computation burdens and communication overheads, previous works have adapted Parameter-efficient Finetuning (PEFT) methods for FL. Hereby, only a small fraction of the model parameters are optimized and communicated during federated communications. Nevertheless, most previous works have focused on a single modality and neglected one common phenomenon, i.e., the presence of data heterogeneity across the clients. Therefore, in this work, we propose a finetuning framework tailored to heterogeneous multi-modal FL, called Federated Dual-Aadapter Teacher (FedDAT). Specifically, our approach leverages a Dual-Adapter Teacher (DAT) to address data heterogeneity by regularizing the client local updates and applying Mutual Knowledge Distillation (MKD) for an efficient knowledge transfer. FedDAT is the first approach that enables an efficient distributed finetuning of foundation models for a variety of heterogeneous Vision-Language tasks. To demonstrate its effectiveness, we conduct extensive experiments on four multi-modality FL benchmarks with different types of data heterogeneity, where FedDAT substantially outperforms the existing centralized PEFT methods adapted for FL.

The field of few-shot learning (FSL) has shown promising results in scenarios where training data is limited, but its vulnerability to backdoor attacks remains largely unexplored. We first explore this topic by first evaluating the performance of the existing backdoor attack methods on few-shot learning scenarios. Unlike in standard supervised learning, existing backdoor attack methods failed to perform an effective attack in FSL due to two main issues. Firstly, the model tends to overfit to either benign features or trigger features, causing a tough trade-off between attack success rate and benign accuracy. Secondly, due to the small number of training samples, the dirty label or visible trigger in the support set can be easily detected by victims, which reduces the stealthiness of attacks. It seemed that FSL could survive from backdoor attacks. However, in this paper, we propose the Few-shot Learning Backdoor Attack (FLBA) to show that FSL can still be vulnerable to backdoor attacks. Specifically, we first generate a trigger to maximize the gap between poisoned and benign features. It enables the model to learn both benign and trigger features, which solves the problem of overfitting. To make it more stealthy, we hide the trigger by optimizing two types of imperceptible perturbation, namely attractive and repulsive perturbation, instead of attaching the trigger directly. Once we obtain the perturbations, we can poison all samples in the benign support set into a hidden poisoned support set and fine-tune the model on it. Our method demonstrates a high Attack Success Rate (ASR) in FSL tasks with different few-shot learning paradigms while preserving clean accuracy and maintaining stealthiness. This study reveals that few-shot learning still suffers from backdoor attacks, and its security should be given attention.

Uncertainty quantification is critical for deploying deep neural networks (DNNs) in real-world applications. An Auxiliary Uncertainty Estimator (AuxUE) is one of the most effective means to estimate the uncertainty of the main task prediction without modifying the main task model. To be considered robust, an AuxUE must be capable of maintaining its performance and triggering higher uncertainties while encountering Out-of-Distribution (OOD) inputs, i.e., to provide robust aleatoric and epistemic uncertainty. However, for vision regression tasks, current AuxUE designs are mainly adopted for aleatoric uncertainty estimates, and AuxUE robustness has not been explored. In this work, we propose a generalized AuxUE scheme for more robust uncertainty quantification on regression tasks. Concretely, to achieve a more robust aleatoric uncertainty estimation, different distribution assumptions are considered for heteroscedastic noise, and Laplace distribution is finally chosen to approximate the prediction error. For epistemic uncertainty, we propose a novel solution named Discretization-Induced Dirichlet pOsterior (DIDO), which models the Dirichlet posterior on the discretized prediction error. Extensive experiments on age estimation, monocular depth estimation, and super-resolution tasks show that our proposed method can provide robust uncertainty estimates in the face of noisy inputs and that it can be scalable to both image-level and pixel-wise tasks.

Diffusion-based models have gained significant popularity for text-to-image generation due to their exceptional image-generation capabilities. A risk with these models is the potential generation of inappropriate content, such as biased or harmful images. However, the underlying reasons for generating such undesired content from the perspective of the diffusion model's internal representation remain unclear. Previous work interprets vectors in an interpretable latent space of diffusion models as seman-tic concepts. However, existing approaches cannot discover directions for arbitrary concepts, such as those related to inappropriate concepts. In this work, we propose a novel self-supervised approach to find interpretable latent directions for a given concept. With the discovered vectors, we further propose a simple approach to mitigate inappropriate generation. Extensive experiments have been conducted to verify the effectiveness of our mitigation approach, namely, for fair generation, safe generation, and responsible text-enhancing generation. Project page: https://interpretdiffusion.github.io.

With the prevalence of the Pretraining-Finetuning paradigm in transfer learning, the robustness of down-stream tasks has become a critical concern. In this work, we delve into adversarial robustness in transfer learning and reveal the critical role of initialization, including both the pretrained model and the linear head. First, we discover the necessity of an adversarially robust pre-trained model. Specifically, we reveal that with a standard pretrained model, Parameter-Efficient Finetuning (PEFT) methods either fail to be adversarially robust or continue to exhibit significantly degraded adversarial robustness on downstream tasks, even with adversarial training during finetuning. Leveraging a robust pretrained model, surprisingly, we observe that a simple linear probing can outper-form full finetuning and other PEFT methods with random initialization on certain datasets. We further identify that linear probing excels in preserving robustness from the robust pretraining. Based on this, we propose Robust Linear Initialization (RoLl) for adversarial finetuning, which initializes the linear head with the weights obtained by ad-versarial linear probing to maximally inherit the robust-ness from pretraining. Across five different image classification datasets, we demonstrate the effectiveness of RoLI and achieve new state-of-the-art results. Our code is available at https://github.com/DongXzz/RoLI.

Adversarial attack methods based on point manipulation for 3D point cloud classification have revealed the fragility of 3D models, yet the adversarial examples they produce are easily perceived or defended against. The tradeoff between the imperceptibility and adversarial strength leads most point attack methods to inevitably introduce easily detectable outlier points upon a successful attack. An-other promising strategy, shape-based attack, can effectively eliminate outliers, but existing methods often suffer significant reductions in imperceptibility due to irrational deformations. We find that concealing deformation perturbations in areas insensitive to human eyes can achieve a better tradeoff between imperceptibility and adversarial strength, specifically in parts of the object surface that are complex and exhibit drastic curvature changes. Therefore, we propose a novel shape-based adversarial attack method, HiT-ADV, which initially conducts a two-stage search for attack regions based on saliency and imperceptibility scores, and then adds deformation perturbations in each attack region using Gaussian kernel functions. Additionally, HiT-ADV is extendable to physical attack. We propose that by employing benign resampling and benign rigid transformations, we can further enhance physical adversarial strength with little sacrifice to imperceptibility. Extensive experiments have validated the superiority of our method in terms of adversarial and imperceptible properties in both digital and physical spaces. Our code is avaliable at: https://github.com/TRLou/HiT-ADV.

Large vision-language models (VLMs) such as GPT-4 have achieved exceptional performance across various multi-modal tasks. However, the deployment of VLMs necessitates substantial energy consumption and computational resources. Once attackers maliciously induce high energy consumption and latency time (energy-latency cost) during inference of VLMs, it will exhaust computational resources. In this paper, we explore this attack surface about availability of VLMs and aim to induce high energy-latency cost during inference of VLMs. We find that high energy-latency cost during inference of VLMs can be manipulated by maximizing the length of generated sequences. To this end, we propose verbose images, with the goal of crafting an imperceptible perturbation to induce VLMs to generate long sentences during inference. Concretely, we design three loss objectives. First, a loss is proposed to delay the occurrence of end-of-sequence (EOS) token, where EOS token is a signal for VLMs to stop generating further tokens. Moreover, an uncertainty loss and a token diversity loss are proposed to increase the uncertainty over each generated token and the diversity among all tokens of the whole generated sequence, respectively, which can break output dependency at token-level and sequence-level. Furthermore, a temporal weight adjustment algorithm is proposed, which can effectively balance these losses. Extensive experiments demonstrate that our verbose images can increase the length of generated sequences by 7.87× and 8.56× compared to original images on MS-COCO and ImageNet datasets, which presents potential challenges for various applications.

When a small number of poisoned samples are injected into the training dataset of a deep neural network, the network can be induced to exhibit malicious behavior during inferences, which poses potential threats to real-world applications. While they have been intensively studied in classification, backdoor attacks on semantic segmentation have been largely overlooked. Unlike classification, semantic segmentation aims to classify every pixel within a given image. In this work, we explore backdoor attacks on segmentation models to misclassify all pixels of a victim class by injecting a specific trigger on non-victim pixels during inferences, which is dubbed Influencer Backdoor Attack (IBA). IBA is expected to maintain the classification accuracy of non-victim pixels and mislead classifications of all victim pixels in every single inference and could be easily applied to real-world scenes. Based on the context aggregation ability of segmentation models, we proposed a simple, yet effective, Nearest-Neighbor trigger injection strategy. We also introduce an innovative Pixel Random Labeling strategy which maintains optimal performance even when the trigger is placed far from the victim pixels. Our extensive experiments reveal that current segmentation models do suffer from backdoor attacks, demonstrate IBA real-world applicability, and show that our proposed techniques can further increase attack performance.

Different from traditional task-specific vision models, recent large VLMs can readily adapt to different vision tasks by simply using different textual instructions, i.e., prompts. However, a well-known concern about traditional task-specific vision models is that they can be misled by imperceptible adversarial perturbations. Furthermore, the concern is exacerbated by the phenomenon that the same adversarial perturbations can fool different task-specific models. Given that VLMs rely on prompts to adapt to different tasks, an intriguing question emerges: Can a single adversarial image mislead all predictions of VLMs when a thousand different prompts are given? This question essentially introduces a novel perspective on adversarial transferability: cross-prompt adversarial transferability. In this work, we propose the Cross-Prompt Attack (CroPA). This proposed method updates the visual adversarial perturbation with learnable textual prompts, which are designed to counteract the misleading effects of the adversarial image. By doing this, CroPA significantly improves the transferability of adversarial examples across prompts. Extensive experiments are conducted to verify the strong cross-prompt adversarial transferability of CroPA with prevalent VLMs including Flamingo, BLIP-2, and InstructBLIP in various different tasks.

Using feature attributions for post-hoc explanations is a common practice to understand and verify the predictions of opaque machine learning models. Despite the numerous techniques available, individual methods often produce inconsistent and unstable results, putting their overall reliability into question. In this work, we aim to systematically improve the quality of feature attributions by combining multiple explanations across distinct methods or their variations. For this purpose, we propose a novel approach to derive optimal convex combinations of feature attributions that yield provable improvements of desired quality criteria such as robustness or faithfulness to the model behavior. Through extensive experiments involving various model architectures and popular feature attribution techniques, we demonstrate that our combination strategy consistently outperforms individual methods and existing baselines.

Large Language Model (LLM) agents have been increasingly adopted as simulation tools to model humans in social science and role-playing applications. However, one fundamental question remains: can LLM agents really simulate human behavior? In this paper, we focus on one critical and elemental behavior in human interactions, trust, and investigate whether LLM agents can simulate human trust behavior. We first find that LLM agents generally exhibit trust behavior, referred to as agent trust, under the framework of Trust Games, which are widely recognized in behavioral economics. Then, we discover that GPT-4 agents manifest high behavioral alignment with humans in terms of trust behavior, indicating the feasibility of simulating human trust behavior with LLM agents. In addition, we probe the biases of agent trust and differences in agent trust towards other LLM agents and humans. We also explore the intrinsic properties of agent trust under conditions including external manipulations and advanced reasoning strategies. Our study provides new insights into the behaviors of LLM agents and the fundamental analogy between LLMs and humans beyond value alignment. We further illustrate broader implications of our discoveries for applications where trust is paramount.

Fast Adversarial Training (FAT) not only improves the model robustness but also reduces the training cost of standard adversarial training. However, fast adversarial training often suffers from Catastrophic Overfitting (CO), which results in poor robustness performance. Catastrophic Overfitting describes the phenomenon of a sudden and significant decrease in robust accuracy during the training of fast adversarial training. Many effective techniques have been developed to prevent Catastrophic Overfitting and improve the model robustness from different perspectives. However, these techniques adopt inconsistent training settings and require different training costs, i.e., training time and memory costs, leading to unfair comparisons. In this paper, we conduct a comprehensive study of over 10 fast adversarial training methods in terms of adversarial robustness and training costs. We revisit the effectiveness and efficiency of fast adversarial training techniques in preventing Catastrophic Overfitting from the perspective of model local nonlinearity and propose an effective Lipschitz regularization method for fast adversarial training. Furthermore, we explore the effect of data augmentation and weight averaging in fast adversarial training and propose a simple yet effective auto weight averaging method to improve robustness further. By assembling these techniques, we propose an effective FGSM-based fast adversarial training method equipped with Lipschitz regularization and Auto Weight averaging, abbreviated as FGSM-LAW. Experimental evaluations on four benchmark databases demonstrate the superiority of our method over state-of-the-art fast adversarial training methods and the advanced standard adversarial training methods.

Adversarial training has shown promise in building robust models against adversarial examples. A major drawback of adversarial training is the computational overhead introduced by the generation of adversarial examples. To overcome this limitation, adversarial training based on single-step attacks has been explored. Previous work improves the single-step adversarial training from different perspectives, e.g., sample initialization, loss regularization, and training strategy. Almost all of them treat the underlying model as a black box. In this work, we propose to exploit the interior building blocks of the model to improve efficiency. Specifically, we propose to dynamically sample lightweight subnetworks as a surrogate model during training. By doing this, both the forward and backward passes can be accelerated for efficient adversarial training. Besides, we provide theoretical analysis to show the model robustness can be improved by the single-step adversarial training with sampled subnetworks. Furthermore, we propose a novel sampling strategy where the sampling varies from layer to layer and from iteration to iteration. Compared with previous methods, our method not only reduces the training cost but also achieves better model robustness. Evaluations on a series of popular datasets demonstrate the effectiveness of the proposed FB-Better. Our code has been released at https://github.com/jiaxiaojunQAQ/FP-Better.

Given the effectiveness of deep neural networks in various fields, the security of neural networks has received great attention. The backdoor attack, which induces malicious behaviors of models by poisoning part of the training set, still remains a challenging problem. Many recent efforts have proposed different ways of embedding backdoors to improve the stealthiness of backdoor attacks. Yet, lowering the percentage of poisoned samples is one of the most direct ways to increase stealthiness. A recent study (Filtering-and-Updating strategy, FUS) has revealed that the sample selection for poisoning is also crucial, as different samples contribute differently to the final decision boundary of the network. Concretely, they utilize each sample’s forgetting events during the training stage to identify which samples will contribute more to the network’s prediction. The training phase of their search method, however, is computationally expensive and slow. To overcome this, in this paper, we propose an efficient sample selection strategy based on the high-frequency energy (HFE) of training samples with a global screening and updating strategy, which can not only achieve a higher backdoor-attack success rate but also reduce the searching time by a factor of 4320 compared to FUS (12 hours vs 10 seconds). The extensive experiment results on CIFAR-10, CIFAR-100, and ImageNet-10 have shown that our proposed method is much simpler, faster, and more efficient.

Backdoor defenses have been studied to alleviate the threat of deep neural networks (DNNs) being backdoor attacked and thus maliciously altered. Since DNNs usually adopt some external training data from an untrusted third party, a robust backdoor defense strategy during the training stage is of importance. We argue that the core of training-time defense is to select poisoned samples and to handle them properly. In this work, we summarize the training-time defenses from a unified framework as splitting the poisoned dataset into two data pools. Under our framework, we propose an adaptively splitting dataset-based defense (ASD). Concretely, we apply loss-guided split and meta-learning-inspired split to dynamically update two data pools. With the split clean data pool and polluted data pool, ASD successfully defends against backdoor attacks during training. Extensive experiments on multiple benchmark datasets and DNN models against six state-of-the-art backdoor attacks demonstrate the superiority of our ASD. Our code is available at https://github.com/KuofengGao/ASD.

The field of multimodal research focusing on the comprehension and creation of both images and text has witnessed significant strides. This progress is exemplified by the emergence of sophisticated models dedicated to image captioning at scale, such as the notable Flamingo model and text-to-image generative models, with DALL-E serving as a prominent example. An interesting question worth exploring in this domain is whether Flamingo and DALL-E understand each other. To study this question, we propose a reconstruction task where Flamingo generates a description for a given image and DALL-E uses this description as input to synthesize a new image. We argue that these models understand each other if the generated image is similar to the given image. Specifically, we study the relationship between the quality of the image reconstruction and that of the text generation. We find that an optimal description of an image is one that gives rise to a generated image similar to the original one. The finding motivates us to propose a unified framework to finetune the text-to-image and image-to-text models. Concretely, the reconstruction part forms a regularization loss to guide the tuning of the models. Extensive experiments on multiple datasets with different image captioning and image generation models validate our findings and demonstrate the effectiveness of our proposed unified framework. As DALL-E and Flamingo are not publicly available, we use Stable Diffusion and BLIP in the remaining work. Project website: https://dalleflamingo.github.io.

Federated Learning (FL) is a decentralized machine learning paradigm, in which multiple clients collaboratively train neural networks without centralizing their local data, and hence preserve data privacy. However, real-world FL applications usually encounter challenges arising from distribution shifts across the local datasets of individual clients. These shifts may drift the global model aggregation or result in convergence to deflected local optimum. While existing efforts have addressed distribution shifts in the label space, an equally important challenge remains relatively unexplored. This challenge involves situations where the local data of different clients indicate identical label distributions but exhibit divergent feature distributions. This issue can significantly impact the global model performance in the FL framework. In this work, we propose Federated Representation Augmentation (FRAug) to resolve this practical and challenging problem. FRAug optimizes a shared embedding generator to capture client consensus. Its output synthetic embeddings are transformed into client-specific by a locally optimized RTNet to augment the training space of each client. Our empirical evaluation on three public benchmarks and a real-world medical dataset demonstrates the effectiveness of the proposed method, which substantially outperforms the current state-of-the-art FL methods for feature distribution shifts, including PartialFed and FedBN.

Video-Text Retrieval (VTR) is a crucial multi-modal task in an era of massive video-text data on the Internet. A plethora of work characterized by using a two-stream Vision-Language model architecture that learns a joint representation of video-text pairs has become a prominent approach for the VTR task. However, these models operate under the assumption of bijective video-text correspondences and neglect a more practical scenario where video content usually encompasses multiple events, while texts like user queries or webpage metadata tend to be specific and correspond to single events. This establishes a gap between the previous training objective and real-world applications, leading to the potential performance degradation of earlier models during inference. In this study, we introduce the Multi-event Video-Text Retrieval (MeVTR) task, addressing scenarios in which each video contains multiple different events, as a niche scenario of the conventional Video-Text Retrieval Task. We present a simple model, Me-Retriever, which incorporates key event video representation and a new MeVTR loss for the MeVTR task. Comprehensive experiments show that this straightforward framework outperforms other models in the Video-to-Text and Text-to-Video tasks, effectively establishing a robust baseline for the MeVTR task. We believe this work serves as a strong foundation for future studies. Code is available at https://github.com/gengyuanmax/MeVTR.

Various adaptation methods, such as LoRA, prompts, and adapters, have been proposed to enhance the performance of pre-trained vision-language models in specific domains. As test samples in real-world applications usually differ from adaptation data, the robustness of these adaptation methods against distribution shifts are essential. In this study, we assess the robustness of 11 widely-used adaptation methods across 4 vision-language datasets under multimodal corruptions. Concretely, we introduce 7 benchmark datasets, including 96 visual and 87 textual corruptions, to investigate the robustness of different adaptation methods, the impact of available adaptation examples, and the influence of trainable parameter size during adaptation. Our analysis reveals that: 1) Adaptation methods are more sensitive to text corruptions than visual corruptions. 2) Full fine-tuning does not consistently provide the highest robustness; instead, adapters can achieve better robustness with comparable clean performance. 3) Contrary to expectations, our findings indicate that increasing the number of adaptation data and parameters does not guarantee enhanced robustness; instead, it results in even lower robustness. We hope this study could benefit future research in the development of robust multimodal adaptation methods. The benchmark, code, and dataset used in this study can be accessed at https://adarobustness.github.io.

Capsule Networks, as alternatives to Convolutional Neural Networks, have been proposed to recognize objects from images. The current literature demonstrates many advantages of CapsNets over CNNs. However, how to create explanations for individual classifications of CapsNets has not been well explored. The widely used saliency methods are mainly proposed for explaining CNN-based classifications; they create saliency map explanations by combining activation values and the corresponding gradients, e.g., Grad-CAM. These saliency methods require a specific architecture of the underlying classifiers and cannot be trivially applied to CapsNets due to the iterative routing mechanism therein. To overcome the lack of interpretability, we can either propose new post-hoc interpretation methods for CapsNets or modifying the model to have build-in explanations. In this work, we explore the latter. Specifically, we propose interpretable Graph Capsule Networks (GraCapsNets), where we replace the routing part with a multi-head attention-based Graph Pooling approach. In the proposed model, individual classification explanations can be created effectively and efficiently. Our model also demonstrates some unexpected benefits, even though it replaces the fundamental part of CapsNets. Our GraCapsNets achieve better classification performance with fewer parameters and better adversarial robustness, when compared to CapsNets. Besides, GraCapsNets also keep other advantages of CapsNets, namely, disentangled representations and affine transformation robustness.

The Capsule Network is widely believed to be more robust than Convolutional Networks. However, there are no comprehensive comparisons between these two networks, and it is also unknown which components in the CapsNet affect its robustness. In this paper, we first carefully examine the special designs in CapsNet that differ from that of a ConvNet commonly used for image classification. The examination reveals five major new/different components in CapsNet: a transformation process, a dynamic routing layer, a squashing function, a marginal loss other than cross-entropy loss, and an additional class-conditional reconstruction loss for regularization. Along with these major differences, we conduct comprehensive ablation studies on three kinds of robustness, including affine transformation, overlapping digits, and semantic representation. The study reveals that some designs, which are thought critical to CapsNet, actually can harm its robustness, i.e., the dynamic routing layer and the transformation process, while others are beneficial for the robustness. Based on these findings, we propose enhanced ConvNets simply by introducing the essential components behind the CapsNet’s success. The proposed simple ConvNets can achieve better robustness than the CapsNet.

A large number of saliency methods have been proposed to explain individual decisions of deep convolutional neural networks (DCNNs). They work by identifying the relevance of each input feature to the predicted output class. However, the feature representations of hidden layers are difficult to interpret semantically. In this work, human-interpretable semantic concepts are associated with vectors in feature space. The association process is mathematically formulated as an optimization problem. The semantic vectors obtained from the optimal solution are applied to interpret deep neural networks globally and locally. The global interpretations are useful to understand the knowledge learned by DCNNs. The interpretation of local behaviors can help to gain a better understanding of the individual decisions made by DCNNs. The empirical experiments demonstrate how to use identified semantics to interpret the existing DCNNs.

Convolutional neural networks (CNNs) achieve translational invariance by using pooling operations. However, the operations do not preserve the spatial relationships in the learned representations. Hence, CNNs cannot extrapolate to various geometric transformations of inputs. Recently, Capsule Networks (CapsNets) have been proposed to tackle this problem. In CapsNets, each entity is represented by a vector and routed to high-level entity representations by a dynamic routing algorithm. CapsNets have been shown to be more robust than CNNs to affine transformations of inputs. However, there is still a huge gap between their performance on transformed inputs compared to untransformed versions. In this work, we first revisit the routing procedure by (un)rolling its forward and backward passes. Our investigation reveals that the routing procedure contributes neither to the generalization ability nor to the affine robustness of the CapsNets. Furthermore, we explore the limitations of capsule transformations and propose affine CapsNets (Aff-CapsNets), which are more robust to affine transformations. On our benchmark task, where models are trained on the MNIST dataset and tested on the AffNIST dataset, our Aff-CapsNets improve the benchmark performance by a large margin (from 79% to 93.21%), without using any routing mechanism.

Knowledge Distillation, as a model compression technique, has received great attention. The knowledge of a well-performed teacher is distilled to a student with a small architecture. The architecture of the small student is often chosen to be similar to their teacher's, with fewer layers or fewer channels, or both. However, even with the same number of FLOPs or parameters, the students with different architecture can achieve different generalization ability. The configuration of a student architecture requires intensive network architecture engineering. In this work, instead of designing a good student architecture manually, we propose to search for the optimal student automatically. Based on L1-norm optimization, a subgraph from the teacher network topology graph is selected as a student, the goal of which is to minimize the KL-divergence between student's and teacher's outputs. We verify the proposal on CIFAR10 and CIFAR100 datasets. The empirical experiments show that the learned student architecture achieves better performance than ones specified manually. We also visualize and understand the architecture of the found student.