Multipanel images, commonly seen as web screenshots, posters, etc., pervade our daily lives. These images, characterized by their composition of multiple subfigures in distinct layouts, effectively convey information to people. Toward building advanced multimodal AI applications, such as agents that understand complex scenes and navigate through webpages, the skill of multipanel visual reasoning is essential, and a comprehensive evaluation of models in this regard is important. Therefore, our paper introduces Multipanel Visual Question Answering (MultipanelVQA), a novel benchmark that specifically challenges models in comprehending multipanel images. The benchmark comprises 6,600 questions and answers related to multipanel images. While these questions are straightforward for average humans, achieving nearly perfect correctness, they pose significant challenges to the state-of-the-art Large Vision Language Models (LVLMs) we tested. In our study, we utilized synthetically curated multipanel images specifically designed to isolate and evaluate the impact of diverse factors on model performance, revealing the sensitivity of LVLMs to various interferences in multipanel images, such as adjacent subfigures and layout complexity. As a result, MultipanelVQA highlights the need and direction for improving LVLMs' ability to understand complex visual-language contexts. Code and data are released at https://sites.google.com/view/multipanelvqa/home.
Domain generalization is a critical challenge for machine learning systems. Prior domain generalization methods focus on extracting domain-invariant features across several stationary domains to enable generalization to new domains. However, in non-stationary tasks where new domains evolve in an underlying continuous structure, such as time, merely extracting the invariant features is insufficient for generalization to the evolving new domains. Nevertheless, it is non-trivial to learn both evolving and invariant features within a single model due to their conflicts. To bridge this gap, we build causal models to characterize the distribution shifts concerning the two patterns, and propose to learn both dynamic and invariant features via a new framework called Mutual Information-Based Sequential Autoencoders (MISTS). MISTS adopts information theoretic constraints onto sequential autoencoders to disentangle the dynamic and invariant features, and leverage a domain adaptive classifier to make predictions based on both evolving and invariant information. Our experimental results on both synthetic and real-world datasets demonstrate that MISTS succeeds in capturing both evolving and invariant information, and present promising results in evolving domain generalization tasks.
In-context learning (ICL) refers to the ability of a model to condition on a few in-context demonstrations (input-output examples of the underlying task) to generate the answer for a new query input, without updating parameters. Despite the impressive ICL ability of LLMs, it has also been found that ICL in LLMs is sensitive to input demonstrations and limited to short context lengths. To understand the limitations and principles for successful ICL, we conduct an investigation with ICL linear regression of transformers. We characterize several Out-of-Distribution (OOD) cases for ICL inspired by realistic LLM ICL failures and compare transformers with DeepSet, a simple yet powerful architecture for ICL. Surprisingly, DeepSet outperforms transformers across a variety of distribution shifts, implying that preserving permutation invariance symmetry to input demonstrations is crucial for OOD ICL. The phenomenon specifies a fundamental requirement by ICL, which we termed as ICL invariance. Nevertheless, the positional encodings in LLMs will break ICL invariance. To this end, we further evaluate transformers with identical positional encodings and find preserving ICL invariance in transformers achieves state-of-the-art performance across various ICL distribution shifts
Invariant graph representation learning aims to learn the invariance among data from different environments for out-of-distribution generalization on graphs. As the graph environment partitions are usually expensive to obtain, augmenting the environment information has become the de facto approach. However, the usefulness of the augmented environment information has never been verified. In this work, we find that it is fundamentally impossible to learn invariant graph representations via environment augmentation without additional assumptions. Therefore, we develop a set of minimal assumptions, including variation sufficiency and variation consistency, for feasible invariant graph learning. We then propose a new framework Graph invAriant Learning Assistant (GALA). GALA incorporates an assistant model that needs to be sensitive to graph environment changes or distribution shifts. The correctness of the proxy predictions by the assistant model hence can differentiate the variations in spurious subgraphs. We show that extracting the maximally invariant subgraph to the proxy predictions provably identifies the underlying invariant subgraph for successful OOD generalization under the established minimal assumptions. Extensive experiments on datasets including DrugOOD with various graph distribution shifts confirm the effectiveness of GALA.
In our work, we explore the synergistic capabilities of pre-trained vision-and-language models (VLMs) and large language models (LLMs) for visual commonsense reasoning (VCR). We categorize the problem of VCR into visual commonsense understanding (VCU) and visual commonsense inference (VCI). For VCU, which involves perceiving the literal visual content, pre-trained VLMs exhibit strong cross-dataset generalization. On the other hand, in VCI, where the goal is to infer conclusions beyond image content, VLMs face difficulties. We find that a baseline where VLMs provide perception results (image captions) to LLMs leads to improved performance on VCI. However, we identify a challenge with VLMs' passive perception, which often misses crucial context information, leading to incorrect or uncertain reasoning by LLMs. To mitigate this issue, we suggest a collaborative approach where LLMs, when uncertain about their reasoning, actively direct VLMs to concentrate on and gather relevant visual elements to support potential commonsense inferences. In our method, named ViCor, pre-trained LLMs serve as problem classifiers to analyze the problem category, VLM commanders to leverage VLMs differently based on the problem classification, and visual commonsense reasoners to answer the question. VLMs will perform visual recognition and understanding. We evaluate our framework on two VCR benchmark datasets and outperform all other methods that do not require in-domain supervised fine-tuning.
A common explanation for the failure of out-of-distribution (OOD) generalization is that the model trained with empirical risk minimization (ERM) learns spurious features instead of the desired invariant features. However, several recent studies challenged this explanation and found that deep networks may have already learned sufficiently good features for OOD generalization. The debate extends to the in-distribution and OOD performance correlations along with training or fine-tuning neural nets across a variety of OOD generalization tasks. To understand these seemingly contradicting phenomena, we conduct a theoretical investigation and find that ERM essentially learns both spurious features and invariant features. On the other hand, the quality of learned features during ERM pre-training significantly affects the final OOD performance, as OOD objectives rarely learn new features. Failing to capture all the underlying useful features during pre-training will further limit the final OOD performance. To remedy the issue, we propose Feature Augmented Training (FAT ), to enforce the model to learn all useful features by retaining the already learned features and augmenting new ones by multiple rounds. In each round, the retention and augmentation operations are performed on different subsets of the training data that capture distinct features. Extensive experiments show that FAT effectively learns richer features and consistently improves the OOD performance when applied to various objectives.
The ability to accurately locate and navigate to a specific object is a crucial capability for embodied agents that operate in the real world and interact with objects to complete tasks. Such object navigation tasks usually require large-scale training in visual environments with labeled objects, which generalizes poorly to novel objects in unknown environments. In this work, we present a novel zero-shot object navigation method, Exploration with Soft Commonsense constraints (ESC), that transfers commonsense knowledge in pre-trained models to open-world object navigation without any navigation experience nor any other training on the visual environments. First, ESC leverages a pre-trained vision and language model for open-world prompt-based grounding and a pre-trained commonsense language model for room and object reasoning. Then ESC converts commonsense knowledge into navigation actions by modeling it as soft logic predicates for efficient exploration. Extensive experiments on MP3D, HM3D, and RoboTHOR benchmarks show that our ESC method improves significantly over baselines, and achieves new state-of-the-art results for zero-shot object navigation (e.g., 225\% relative Success Rate improvement than CoW on MP3D).
Federated embodied agent learning protects the data privacy of individual visual environments by keeping data locally at each client (the individual environment) during training. However, since the local data is inaccessible to the server under federated learning, attackers may easily poison the training data of the local client to build a backdoor in the agent without notice. Deploying such an agent raises the risk of potential harm to humans, as the attackers may easily navigate and control the agent as they wish via the backdoor. Towards Byzantine-robust federated embodied agent learning, in this paper, we study the attack and defense for the task of vision-and-language navigation (VLN), where the agent is required to follow natural language instructions to navigate indoor environments. First, we introduce a simple but effective attack strategy, Navigation as Wish (NAW), in which the malicious client manipulates local trajectory data to implant a backdoor into the global model. Results on two VLN datasets (R2R and RxR) show that NAW can easily navigate the deployed VLN agent regardless of the language instruction, without affecting its performance on normal test sets. Then, we propose a new Prompt-Based Aggregation (PBA) to defend against the NAW attack in federated VLN, which provides the server with a ''prompt'' of the vision-and-language alignment variance between the benign and malicious clients so that they can be distinguished during training. We validate the effectiveness of the PBA method on protecting the global model from the NAW attack, which outperforms other state-of-the-art defense methods by a large margin in the defense metrics on R2R and RxR.
Building a conversational embodied agent to execute real-life tasks has been a long-standing yet quite challenging research goal, as it requires effective human-agent communication, multi-modal understanding, long-range sequential decision making, etc. Traditional symbolic methods have scaling and generalization issues, while end-to-end deep learning models suffer from data scarcity and high task complexity, and are often hard to explain. To benefit from both worlds, we propose a Neuro-Symbolic Commonsense Reasoning (JARVIS) framework for modular, generalizable, and interpretable conversational embodied agents. First, it acquires symbolic representations by prompting large language models (LLMs) for language understanding and sub-goal planning, and by constructing semantic maps from visual observations. Then the symbolic module reasons for sub-goal planning and action generation based on task- and action-level common sense. Extensive experiments on the TEACh dataset validate the efficacy and efficiency of our JARVIS framework, which achieves state-of-the-art (SOTA) results on all three dialog-based embodied tasks, including Execution from Dialog History (EDH), Trajectory from Dialog (TfD), and Two-Agent Task Completion (TATC) (e.g., our method boosts the unseen Success Rate on EDH from 6.1\% to 15.8\%). Moreover, we systematically analyze the essential factors that affect the task performance and also demonstrate the superiority of our method in few-shot settings. Our JARVIS model ranks first in the Alexa Prize SimBot Public Benchmark Challenge.
We consider stochastic convex optimization for heavy-tailed data with the guarantee of being differentially private (DP). Prior work on this problem is restricted to the gradient descent (GD) method, which is inefficient for large-scale problems. In this paper, we resolve this issue and derive the first high-probability bounds for private stochastic method with clipping. For general convex problems, we derive excess population risks $\Tilde{O}\left(\frac{d^{1/7}\sqrt{\ln\frac{(n \epsilon)^2}{\beta d}}}{(n\epsilon)^{2/7}}\right)$ and $\Tilde{O}\left(\frac{d^{1/7}\ln\frac{(n\epsilon)^2}{\beta d}}{(n\epsilon)^{2/7}}\right)$ under bounded or unbounded domain assumption, respectively (here $n$ is the sample size, $d$ is the dimension of the data, $\beta$ is the confidence level and $\epsilon$ is the private level). Then, we extend our analysis to the strongly convex case and non-smooth case (which works for generalized smooth objectives with H$\ddot{\text{o}}$lder-continuous gradients). We establish new excess risk bounds without bounded domain assumption. The results above achieve lower excess risks and gradient complexities than existing methods in their corresponding cases. Numerical experiments are conducted to justify the theoretical improvement.