AAAI 2024 Papers

Humans have been using stories to entertain, educate, and persuade audiences for centuries. The advent of modern AI tools in the form of Large Language Models (LLMs) such as chatGPT continues to fulfill this purpose. However while recent work has shown that LLMs can successfully be used for narrative generation, they lack coherence and can be prone to repetition and stilted language. Automated Planning can therefore be combined with Natural Language text generation to create narratives (stories) that are logical, coherent, and believable. A planning model provides scaffolding to an LLM so that the LLM's language generation is context-dependent, in order to allow users to create more coherent, logical, and believable stories in a variety of domains.

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.

Artificial Intelligence is undoubtedly becoming pervasive in everyday life of everyone. In this setting, developing correct AI conception since childhood is not only a need to be addressed in educational curricula, but is also a children right. Accordingly, several initiatives at national and international levels aim at promoting AI and emerging technology literacy, supported also by a proliferation in the literature of learning courses covering a variety of topics, learning objectives and targeted ages. Schools are therefore pushed to introduce innovative activities for children in their curricula. In this paper, we report the results of a case study where we tested the contribution of an AI block-based course in developing computational thinking, and human and AI minds understanding in fifth and sixth grade children.

Document Image Enhancement (DIE) remains challenging due to the prevalence of multiple degradations in document images captured by cameras. In this paper, we respond an interesting question: can the performance of pre-trained models and downstream DIE models be improved if they are bootstrapped using different degradation types of the same semantic samples and their high-dimensional features with ambiguous inter-class distance? To this end, we propose an effective contrastive learning paradigm for DIE — a Document image enhancement framework with Normalization and Latent Contrast (DocNLC). While existing DIE methods focus on eliminating one type of degradation, DocNLC considers the relationship between different types of degradation while utilizing both direct and latent contrasts to constrain content consistency, thus achieving a unified treatment of multiple types of degradation. Specifically, we devise a latent contrastive learning module to enforce explicit decorrelation of the normalized representations of different degradation types and to minimize the redundancy between them. Comprehensive experiments show that our method outperforms state-of-the-art DIE models in both pre-training and fine-tuning stages on four publicly available independent datasets. In addition, we discuss the potential benefits of DocNLC for downstream tasks. Our code is released at https://github.com/RylonW/DocNLC

Multimodal Sarcasm Understanding (MSU) has a wide range of applications in the news field such as public opinion analysis and forgery detection. However, existing MSU benchmarks and approaches usually focus on sentence-level MSU. In document-level news, sarcasm clues are sparse or small and are often concealed in long text. Moreover, compared to sentence-level comments like tweets, which mainly focus on only a few trends or hot topics (e.g., sports events), content in the news is considerably diverse. Models created for sentence-level MSU may fail to capture sarcasm clues in document-level news. To fill this gap, we present a comprehensive benchmark for Document-level Multimodal Sarcasm Understanding (DocMSU). Our dataset contains 102,588 pieces of news with text-image pairs, covering 9 diverse topics such as health, business, etc. The proposed large-scale and diverse DocMSU significantly facilitates the research of document-level MSU in real-world scenarios. To take on the new challenges posed by DocMSU, we introduce a fine-grained sarcasm comprehension method to properly align the pixel-level image features with word-level textual features in documents. Experiments demonstrate the effectiveness of our method, showing that it can serve as a baseline approach to the challenging DocMSU.

We propose DocFormerv2, a multi-modal transformer for Visual Document Understanding (VDU). The VDU domain entails understanding documents (beyond mere OCR predictions) e.g., extracting information from a form, VQA for documents and other tasks. VDU is challenging as it needs a model to make sense of multiple modalities (visual, language and spatial) to make a prediction. Our approach, termed DocFormerv2 is an encoder-decoder transformer which takes as input - vision, language and spatial features. DocFormerv2 is pre-trained with unsupervised tasks employed asymmetrically i.e., two novel document tasks on encoder and one on the auto-regressive decoder. The unsupervised tasks have been carefully designed to ensure that the pre-training encourages local-feature alignment between multiple modalities. DocFormerv2 when evaluated on nine challenging datasets shows state-of-the-art performance on all over strong baselines - On TabFact (+4.3%), InfoVQA (+1.4%), FUNSD (+1.0%). Furthermore, to show generalization capabilities, on three VQA tasks involving scene-text, DocFormerv2 outperforms previous comparably-sized models and even does better than much larger models (such as GIT2, PaLI and Flamingo) on these tasks. Extensive ablations show that due to its novel pre-training tasks, DocFormerv2 understands multiple modalities better than prior-art in VDU.

Artificial Intelligence (AI) models have become an integral part of modern society, significantly improving human lives. However, ensuring the reliability and safety of these models is of paramount importance. One critical aspect is the continuous monitoring and verification of model performance to prevent any potential risks. Real-time online evaluation of AI models is necessary to maintain their effectiveness and mitigate any harm caused by performance degradation. The traditional approach to model evaluation involves supervised methods that rely on manual labeling to compare results with model predictions. Unfortunately, this method is not suitable for online model monitoring due to its inherent lag and high cost. While there have been attempts to explore free-label model evaluation, these approaches often consider only the global features of the entire dataset. Additionally, they can only perform model evaluation based on a single dimension of model confidence or features. In this paper, we propose a novel approach called Divide-and-Aggregate Learning (DAL) for unsupervised model evaluation. Our method addresses the limitations of previous approaches by dividing the output of the model into buckets, capturing local information of the distribution. We then aggregate this local information to obtain global information and further represent the relationship between the distribution and model performance. Importantly, our method can simultaneously handle the confidence distribution and feature distribution of the model output. Extensive experiments have been conducted to demonstrate the effectiveness of our DAL model. The results show that our approach outperforms previous methods on four widely used datasets. We will make our source code publicly available.

Large-scale pre-training has proven to be an effective method for improving performance across different tasks. Current person search methods use ImageNet pre-trained models for feature extraction, yet it is not an optimal solution due to the gap between the pre-training task and person search task (as a downstream task). Therefore, in this paper, we focus on pre-training for person search, which involves detecting and re-identifying individuals simultaneously. Although labeled data for person search is scarce, datasets for two sub-tasks person detection and re-identification are relatively abundant. To this end, we propose a hybrid pre-training framework specifically designed for person search using sub-task data only. It consists of a hybrid learning paradigm that handles data with different kinds of supervisions, and an intra-task alignment module that alleviates domain discrepancy under limited resources. To the best of our knowledge, this is the first work that investigates how to support full-task pre-training using sub-task data. Extensive experiments demonstrate that our pre-trained model can achieve significant improvements across diverse protocols, such as person search method, fine-tuning data, pre-training data and model backbone. For example, our model improves ResNet50 based NAE by 10.3% relative improvement w.r.t. mAP. Our code and pre-trained models are released for plug-and-play usage to the person search community (https://github.com/personsearch/PretrainPS).

This paper tackles the problem of federated domain generalization in person re-identification (FedDG re-ID), aiming to learn a model generalizable to unseen domains with decentralized source domains. Previous methods mainly focus on preventing local overfitting. However, the direction of diversifying local data through stylization for model training is largely overlooked. This direction is popular in domain generalization but will encounter two issues under federated scenario: (1) Most stylization methods require the centralization of multiple domains to generate novel styles but this is not applicable under decentralized constraint. (2) The authenticity of generated data cannot be ensured especially given limited local data, which may impair the model optimization. To solve these two problems, we propose the Diversity-Authenticity Co-constrained Stylization (DACS), which can generate diverse and authentic data for learning robust local model. Specifically, we deploy a style transformation model on each domain to generate novel data with two constraints: (1) A diversity constraint is designed to increase data diversity, which enlarges the Wasserstein distance between the original and transformed data; (2) An authenticity constraint is proposed to ensure data authenticity, which enforces the transformed data to be easily/hardly recognized by the local-side global/local model. Extensive experiments demonstrate the effectiveness of the proposed DACS and show that DACS achieves state-of-the-art performance for FedDG re-ID.

In recent years, following the success of text guided image generation, text guided 3D generation has gained increasing attention among researchers. Dreamfusion is a notable approach that enhances generation quality by utilizing 2D text guided diffusion models and introducing SDS loss, a technique for distilling 2D diffusion model information to train 3D models. However, the SDS loss has two major limitations that hinder its effectiveness. Firstly, when given a text prompt, the SDS loss struggles to produce diverse content. Secondly, during training, SDS loss may cause the generated content to overfit and collapse, limiting the model's ability to learn intricate texture details. To overcome these challenges, we propose a novel approach called Noise Recalibration algorithm. By incorporating this technique, we can generate 3D content with significantly greater diversity and stunning details. Our approach offers a promising solution to the limitations of SDS loss.

We consider the task of generating diverse and realistic videos guided by natural audio samples from a wide variety of semantic classes. For this task, the videos are required to be aligned both globally and temporally with the input audio: globally, the input audio is semantically associated with the entire output video, and temporally, each segment of the input audio is associated with a corresponding segment of that video. We utilize an existing text-conditioned video generation model and a pre-trained audio encoder model. The proposed method is based on a lightweight adaptor network, which learns to map the audio-based representation to the input representation expected by the text-to-video generation model. As such, it also enables video generation conditioned on text, audio, and, for the first time as far as we can ascertain, on both text and audio. We validate our method extensively on three datasets demonstrating significant semantic diversity of audio-video samples and further propose a novel evaluation metric (AV-Align) to assess the alignment of generated videos with input audio samples. AV-Align is based on the detection and comparison of energy peaks in both modalities. In comparison to recent state-of-the-art approaches, our method generates videos that are better aligned with the input sound, both with respect to content and temporal axis. We also show that videos produced by our method present higher visual quality and are more diverse. Code and samples are available at: https://pages.cs.huji.ac.il/adiyoss-lab/TempoTokens/.

Generative Artificial Intelligence (AI) has garnered significant attention for its remarkable ability to generate text, images, and other forms of content. However, an inherent and increasingly concerning issue within generative AI systems is bias. These AI models often exhibit an Anglo-centric bias and tend to overlook the importance of diversity. This can be attributed to their training on extensive datasets sourced from the internet, which inevitably inherit the biases present in those data sources. Employing these datasets leads to AI-generated content that mirrors and perpetuates existing biases, encompassing various aspects such as gender, ethnic and cultural stereotypes. Addressing bias in generative AI is a complex challenge that necessitates substantial efforts. In order to tackle this issue, we propose a methodology for constructing moderately sized datasets with a social inclination. These datasets can be employed to rectify existing imbalances in datasets or to train models to generate socially inclusive material. Additionally, we present preliminary findings derived from training our model on these socially inclined datasets.

Text-based person search is a challenging task aimed at locating specific target pedestrians through text descriptions. Recent advancements have been made in this field, but there remains a deficiency in datasets tailored for text-based person search. The creation of new, real-world datasets is hindered by concerns such as the risk of pedestrian privacy leakage and the substantial costs of annotation. In this paper, we introduce a framework, named Diverse Person (DP), to achieve efficient and high-quality text-based person search data generation without involving privacy concerns. Specifically, we propose to leverage available images of clothing and accessories as reference attribute images to edit the original dataset images through diffusion models. Additionally, we employ a Large Language Model (LLM) to produce annotations that are both high in quality and stylistically consistent with those found in real-world datasets. Extensive experimental results demonstrate that the baseline models trained with our DP can achieve new state-of-the-art results on three public datasets, with performance improvements up to 4.82%, 2.15%, and 2.28% on CUHK-PEDES, ICFG-PEDES, and RSTPReid in terms of Rank-1 accuracy, respectively.

To achieve high-quality translation with low latency, a Simultaneous Speech Translation (SimulST) system relies on a policy module to decide whether to translate immediately or wait for additional streaming input, along with a translation model capable of effectively handling partial speech input. Prior research has tackled these components separately, either using ``wait-k'' policies based on fixed-length segments or detected word boundaries, or dynamic policies based on different strategies (e.g., meaningful units), while employing offline models for prefix-to-prefix translation. In this paper, we propose Divergence-Guided Simultaneous Speech Translation (DiG-SST), a tightly integrated approach focusing on both translation quality and latency for streaming input. Specifically, we introduce a simple yet effective prefix-based strategy for training translation models with partial speech input, and develop an adaptive policy that makes read/write decisions for the translation model based on the expected divergence in translation distributions resulting from future input. Our experiments on multiple translation directions of the MuST-C benchmark demonstrate that our approach achieves a better trade-off between translation quality and latency compared to existing methods.

Recommendation strategies are typically evaluated by using previously logged data, employing off-policy evaluation methods to estimate their expected performance. However, for strategies that present users with slates of multiple items, the resulting combinatorial action space renders many of these methods impractical. Prior work has developed estimators that leverage the structure in slates to estimate the expected off-policy performance, but the estimation of the entire performance distribution remains elusive. Estimating the complete distribution allows for a more comprehensive evaluation of recommendation strategies, particularly along the axes of risk and fairness that employ metrics computable from the distribution. In this paper, we propose an estimator for the complete off-policy performance distribution for slates and establish conditions under which the estimator is unbiased and consistent. This builds upon prior work on off-policy evaluation for slates and off-policy distribution estimation in reinforcement learning. We validate the efficacy of our method empirically on synthetic data as well as on a slate recommendation simulator constructed from real-world data (MovieLens-20M). Our results show a significant reduction in estimation variance and improved sample efficiency over prior work across a range of slate structures.

Instrumental variables (IVs), widely applied in economics and healthcare, enable consistent counterfactual prediction in the presence of hidden confounding factors, effectively addressing endogeneity issues. The prevailing IV-based counterfactual prediction methods typically rely on the availability of valid IVs (satisfying Relevance, Exclusivity, and Exogeneity), a requirement which often proves elusive in real-world scenarios. Various data-driven techniques are being developed to create valid IVs (or representations of IVs) from a pool of IV candidates. However, most of these techniques still necessitate the inclusion of valid IVs within the set of candidates. This paper proposes a distribution-conditioned adversarial variational autoencoder to tackle this challenge. Specifically: 1) for Relevance and Exclusivity, we deduce the corresponding evidence lower bound following the Bayesian network structure and build the variational autoencoder; accordingly, 2) for Exogeneity , we design an adversarial game to encourage latent factors originating from the marginal distribution, compelling the independence between IVs and other outcome-related factors. Extensive experimental results validate the effectiveness, stability and generality of our proposed model in generating valid IV factors in the absence of valid IV candidates.

Multi-Task Learning (MTL) is a framework, where multiple related tasks are learned jointly and benefit from a shared representation space, or parameter transfer. To provide sufficient learning support, modern MTL uses annotated data with full, or sufficiently large overlap across tasks, i.e., each input sample is annotated for all, or most of the tasks. However, collecting such annotations is prohibitive in many real applications, and cannot benefit from datasets available for individual tasks. In this work, we challenge this setup and show that MTL can be successful with classification tasks with little, or non-overlapping annotations, or when there is big discrepancy in the size of labeled data per task. We explore task-relatedness for co-annotation and co-training, and propose a novel approach, where knowledge exchange is enabled between the tasks via distribution matching. To demonstrate the general applicability of our method, we conducted diverse case studies in the domains of affective computing, face recognition, species recognition, and shopping item classification using nine datasets. Our large-scale study of affective tasks for basic expression recognition and facial action unit detection illustrates that our approach is network agnostic and brings large performance improvements compared to the state-of-the-art in both tasks and across all studied databases. In all case studies, we show that co-training via task-relatedness is advantageous and prevents negative transfer (which occurs when MT model's performance is worse than that of at least one single-task model).

Conventional image set methods typically learn from image sets stored in one location. However, in real-world applications, image sets are often distributed or collected across different positions. Learning from such distributed image sets presents a challenge that has not been studied thus far. Moreover, efficiency is seldom addressed in large-scale image set applications. To fulfill these gaps, this paper proposes Distributed Manifold Hashing (DMH), which models distributed image sets as a connected graph. DMH employs Riemannian manifold to effectively represent each image set and further suggests learning hash code for each image set to achieve efficient computation and storage. DMH is formally formulated as a distributed learning problem with local consistency constraint on global variables among neighbor nodes, and can be optimized in parallel. Extensive experiments on three benchmark datasets demonstrate that DMH achieves highly competitive accuracies in a distributed setting and provides faster classification and retrieval than state-of-the-arts.

Partial label learning (PLL) refers to the classification task where each training instance is ambiguously annotated with a set of candidate labels. Despite substantial advancements in tackling this challenge, limited attention has been devoted to a more specific and realistic setting, denoted as instance-dependent partial label learning (IDPLL). Within this contex, the assignment of partial labels depends on the distinct features of individual instances, rather than being random. In this paper, we initiate an exploration into a self-distillation framework for this problem, driven by the proven effectiveness and stability of this framework. Nonetheless, a crucial shortfall is identified: the foundational assumption central to IDPLL, involving what we term as partial label knowledge stipulating that candidate labels should exhibit superior confidence compared to non-candidates, is not fully upheld within the distillation process. To address this challenge, we introduce DIRK, a novel distillation approach that leverages a rectification process to DIstill Reliable Knowledge, while concurrently preserves informative fine-grained label confidence. In addition, to harness the rectified confidence to its fullest potential, we propose a knowledge-based representation refinement module, seamlessly integrated into the DIRK framework. This module effectively transmits the essence of similarity knowledge from the label space to the feature space, thereby amplifying representation learning and subsequently engendering marked improvements in model performance. Experiments and analysis on multiple datasets validate the rationality and superiority of our proposed approach.