AAAI 2024 Papers

This research focuses on improving the robustness of machine learning systems to natural variations and distribution shifts. A design trade space is presented, and various methods are compared, including adversarial training, data augmentation techniques, and novel approaches inspired by model-based robust optimization formulations.

Models that learn from data are widely and rapidly being deployed today for real-world use, but they suffer from unforeseen failures due to distribution shift, adversarial attacks, noise and corruption, and data scarcity. But many failures also occur because many modern AI tasks require reasoning beyond pattern matching -- and such reasoning abilities are difficult to formulate as data-based input-output function fitting. The reliability problem has become increasingly important under the new paradigm of semantic ``multimodal'' learning. My research provides avenues to develop robust and reliable computer vision systems, particularly by leveraging the interactions between vision and language. In this AAAI New Faculty highlights talk, I will cover three thematic areas of my research, ranging from robustness in computer vision, open-domain reliability in visual reasoning, and challenges and opportunities in evaluation of generative models. Readers are encouraged to refer to my website (www.tejasgokhale.com) for more details and updates from my lab's activities towards the goal of robust visual understanding.

Image stitching seamlessly integrates images captured from varying perspectives into a single wide field-of-view image. Such integration not only broadens the captured scene but also augments holistic perception in computer vision applications. Given a pair of captured images, subtle perturbations and distortions which go unnoticed by the human visual system tend to attack the correspondence matching, impairing the performance of image stitching algorithms. In light of this challenge, this paper presents the first attempt to improve the robustness of image stitching against adversarial attacks. Specifically, we introduce a stitching-oriented attack (SoA), tailored to amplify the alignment loss within overlapping regions, thereby targeting the feature matching procedure. To establish an attack resistant model, we delve into the robustness of stitching architecture and develop an adaptive adversarial training (AAT) to balance attack resistance with stitching precision. In this way, we relieve the gap between the routine adversarial training and benign models, ensuring resilience without quality compromise. Comprehensive evaluation across real-world and synthetic datasets validate the deterioration of SoA on stitching performance. Furthermore, AAT emerges as a more robust solution against adversarial perturbations, delivering superior stitching results. Code is available at: https://github.com/Jzy2017/TRIS.

Anomaly detection (AD), often termed outlier detection, is a key machine learning (ML) task, aiming to identify uncommon yet crucial patterns in data. With the increasing complexity of the modern world, the applications of AD span wide—from NASA's spacecraft monitoring to early patient prioritization at University of Pittsburgh Medical Center. Technology giants like Google and Amazon also leverage AD for service disruption identification. Here, I will traverse my AD works with promising new directions, particularly emphasizing reproducible benchmarks (Part 1), automated algorithms (Part 2), and scalable systems (Part 3).

The real-world deployment of machine learning algorithms often poses challenges due to shifts in data distributions and tasks. These shifts can lead to a degradation in model performance, as the model may not have encountered such changes during training. Additionally, they can make it difficult for the model to generalize to new scenarios and can result in poor performance in real-world applications. In this talk, I will present our research on building machine learning models that are highly generalizable and easily adaptable to different shifts. Specifically, I will first discuss our approach to improving out-of-distribution robustness and mitigating spurious correlations by training environment-invariant models through selective augmentation and post-hoc rectification. Second, I will present our techniques for continuous and rapid adaptation of models to new tasks and environments. This includes methods to facilitate compositional generalization and adaptation by extracting relationships from historical observations and to enhance reliable adaptation even in the face of imperfect observations. Additionally, I will showcase our successful practices for addressing shifts in real-world applications, such as in the healthcare, e-commerce, and transportation industries. The talk will also touch upon the remaining challenges and outline future research directions in this area.

Test-Time Adaptation aims to adapt source domain model to testing data at inference stage with success demonstrated in adapting to unseen corruptions. However, these attempts may fail under more challenging real-world scenarios. Existing works mainly consider real-world test-time adaptation under non-i.i.d. data stream and continual domain shift. In this work, we first complement the existing real-world TTA protocol with a globally class imbalanced testing set. We demonstrate that combining all settings together poses new challenges to existing methods. We argue the failure of state-of-the-art methods is first caused by indiscriminately adapting normalization layers to imbalanced testing data. To remedy this shortcoming, we propose a balanced batchnorm layer to swap out the regular batchnorm at inference stage. The new batchnorm layer is capable of adapting without biasing towards majority classes. We are further inspired by the success of self-training (ST) in learning from unlabeled data and adapt ST for test-time adaptation. However, ST alone is prone to over adaption which is responsible for the poor performance under continual domain shift. Hence, we propose to improve self-training under continual domain shift by regularizing model updates with an anchored loss. The final TTA model, termed as TRIBE, is built upon a tri-net architecture with balanced batchnorm layers. We evaluate TRIBE on four datasets representing real-world TTA settings. TRIBE consistently achieves the state-of-the-art performance across multiple evaluation protocols. The code is available at https://github.com/Gorilla-Lab-SCUT/TRIBE.

We study the fair division of indivisible items with subsidies among n agents, where the absolute marginal valuation of each item is at most one. Under monotone valuations (where each item is a good), it is known that a maximum subsidy of 2(n-1) and a total subsidy of 2(n-1)² are sufficient to guarantee the existence of an envy-freeable allocation. In this paper, we improve upon these bounds, even in a wider model. Namely, we show that, given an EF1 allocation, we can compute in polynomial time an envy-free allocation with a subsidy of at most n-1 per agent and a total subsidy of at most n(n-1)/2. Moreover, we present further improved bounds for monotone valuations.

Conventional Outlier Robust Tensor Decomposition (ORTD) approaches generally represent sparse outlier corruption within a specific mode. However, such an assumption, which may hold for matrices, proves inadequate when applied to high-order tensors. In the tensor domain, the outliers are prone to be corrupted in multiple modes simultaneously. Addressing this limitation, this study proposes a novel ORTD approach by recovering low-rank tensors contaminated by outliers spanning multiple modes. In particular, we conceptualize outliers within high-order tensors as latent tensor group sparsity by decomposing the corrupted tensor into a sum of multiple latent components, where each latent component is exclusive to outliers within a particular direction. Thus, it can effectively mitigate the outlier corruptions prevalent in high-order tensors across multiple modes. To theoretically guarantee recovery performance, we rigorously analyze a non-asymptotic upper bound of the estimation error for the proposed ORTD approach. In the optimization process, we develop an efficient alternate direction method of multipliers (ADMM) algorithm. Empirical validation of the approach's efficacy is undertaken through comprehensive experimentation.

Multi-Intent spoken language understanding (SLU) can handle complicated utterances expressing multiple intents, which has attracted increasing attention from researchers. Although existing models have achieved promising performance, most of them still suffer from two leading problems: (1) each intent has its specific scope and the semantic information outside the scope might potentially hinder accurate predictions, i.e. scope barrier; (2) only the guidance from intent to slot is modeled but the guidance from slot to intent is often neglected, i.e. unidirectional guidance. In this paper, we propose a novel Multi-Intent SLU framework termed HAOT, which utilizes hierarchical attention to divide the scopes of each intent and applies optimal transport to achieve the mutual guidance between slot and intent. Experiments demonstrate that our model achieves state-of-the-art performance on two public Multi-Intent SLU datasets, obtaining the 3.4 improvement on MixATIS dataset compared to the previous best models in overall accuracy.

Natural language explanation in visual question answer (VQA-NLE) aims to explain the decision-making process of models by generating natural language sentences to increase users' trust in the black-box systems. Existing post-hoc methods have achieved significant progress in obtaining a plausible explanation. However, such post-hoc explanations are not always aligned with human logical inference, suffering from the issues on: 1) Deductive unsatisfiability, the generated explanations do not logically lead to the answer; 2) Factual inconsistency, the model falsifies its counterfactual explanation for answers without considering the facts in images; and 3) Semantic perturbation insensitivity, the model can not recognize the semantic changes caused by small perturbations. These problems reduce the faithfulness of explanations generated by models. To address the above issues, we propose a novel self-supervised Multi-level Contrastive Learning based natural language Explanation model (MCLE) for VQA with semantic-level, image-level, and instance-level factual and counterfactual samples. MCLE extracts discriminative features and aligns the feature spaces from explanations with visual question and answer to generate more consistent explanations. We conduct extensive experiments, ablation analysis, and case study to demonstrate the effectiveness of our method on two VQA-NLE benchmarks.

Despite the recent progress in deep neural networks (DNNs), it remains challenging to explain the predictions made by DNNs. Existing explanation methods for DNNs mainly focus on post-hoc explanations where another explanatory model is employed to provide explanations. The fact that post-hoc methods can fail to reveal the actual original reasoning process of DNNs raises the need to build DNNs with built-in interpretability. Motivated by this, many self-explaining neural networks have been proposed to generate not only accurate predictions but also clear and intuitive insights into why a particular decision was made. However, existing self-explaining networks are limited in providing distribution-free uncertainty quantification for the two simultaneously generated prediction outcomes (i.e., a sample's final prediction and its corresponding explanations for interpreting that prediction). Importantly, they also fail to establish a connection between the confidence values assigned to the generated explanations in the interpretation layer and those allocated to the final predictions in the ultimate prediction layer. To tackle the aforementioned challenges, in this paper, we design a novel uncertainty modeling framework for self-explaining networks, which not only demonstrates strong distribution-free uncertainty modeling performance for the generated explanations in the interpretation layer but also excels in producing efficient and effective prediction sets for the final predictions based on the informative high-level basis explanations. We perform the theoretical analysis for the proposed framework. Extensive experimental evaluation demonstrates the effectiveness of the proposed uncertainty framework.

Model extraction attacks (MEAs) enable an attacker to replicate the functionality of a victim deep neural network (DNN) model by only querying its API service remotely, posing a severe threat to the security and integrity of pay-per-query DNN-based services. Although the majority of current research on MEAs has primarily concentrated on neural classifiers, there is a growing prevalence of image-to-image translation (I2IT) tasks in our everyday activities. However, techniques developed for MEA of DNN classifiers cannot be directly transferred to the case of I2IT, rendering the vulnerability of I2IT models to MEA attacks often underestimated. This paper unveils the threat of MEA in I2IT tasks from a new perspective. Diverging from the traditional approach of bridging the distribution gap between attacker queries and victim training samples, we opt to mitigate the effect caused by the different distributions, known as the domain shift. This is achieved by introducing a new regularization term that penalizes high-frequency noise, and seeking a flatter minimum to avoid overfitting to the shifted distribution. Extensive experiments on different image translation tasks, including image super-resolution and style transfer, are performed on different backbone victim models, and the new design consistently outperforms the baseline by a large margin across all metrics. A few real-life I2IT APIs are also verified to be extremely vulnerable to our attack, emphasizing the need for enhanced defenses and potentially revised API publishing policies.

The learnware paradigm aims to establish a market of numerous well-performed machine learning models, enabling users to leverage existing helpful models for their tasks instead of starting from scratch. Each learnware in the market is a model submitted by its developer, associated with a specification generated with the help of learnware market, representing the model's specialty and utility and enabling it to be identified for new user tasks. As the market continuously scales up, accommodating an ever-increasing number of learnwares, the critical challenge of the learnware paradigm is to effectively and efficiently identify the most helpful learnware(s) for a new user task without accessing the user's raw data. In this paper, to achieve increasingly accurate learnware characterization and identification along with a growing number of learnwares in the market, we propose an approach called Evolvable Learnware Specification with Index (ELSI). Specifically, based on the key idea of leveraging the task information within learnware specifications, we tackle the challenge of ascertaining the capabilities of models beyond their original training tasks, thereby enabling learnware specifications and the entire market to evolve continuously. Furthermore, through organizing learnwares and constructing specification indexes, we design a practical procedure to accurately and efficiently identify helpful learnwares without examining the entire market. Theoretical analysis and extensive experiments on a learnware market prototype encompassing thousands of models and covering six real-world scenarios validate the effectiveness and efficiency of our approach.

Recently, there has been a growing interest in learning and explaining causal effects within Neural Network (NN) models. By virtue of NN architectures, previous approaches consider only direct and total causal effects assuming independence among input variables. We view an NN as a structural causal model (SCM) and extend our focus to include indirect causal effects by introducing feedforward connections among input neurons. We propose an ante-hoc method that captures and maintains direct, indirect, and total causal effects during NN model training. We also propose an algorithm for quantifying learned causal effects in an NN model and efficient approximation strategies for quantifying causal effects in high-dimensional data. Extensive experiments conducted on synthetic and real-world datasets demonstrate that the causal effects learned by our ante-hoc method better approximate the ground truth effects compared to existing methods.

Randomized smoothing is currently the state-of-the-art method that provides certified robustness for deep neural networks. However, due to its excessively conservative nature, this method of incomplete verification often cannot achieve an adequate certified radius on real-world datasets. One way to obtain a larger certified radius is to use an input-specific algorithm instead of using a fixed Gaussian filter for all data points. Several methods based on this idea have been proposed, but they either suffer from high computational costs or gain marginal improvement in certified radius. In this work, we show that by exploiting the quasiconvex problem structure, we can find the optimal certified radii for most data points with slight computational overhead. This observation leads to an efficient and effective input-specific randomized smoothing algorithm. We conduct extensive experiments and empirical analysis on CIFAR-10 and ImageNet. The results show that the proposed method significantly enhances the certified radii with low computational overhead.

Graph neural networks (GNNs) have recently been shown to be vulnerable to adversarial attacks, where slight perturbations in the graph structure can lead to erroneous predictions. However, current robust models for defending against such attacks inherit the transductive limitations of graph convolutional networks (GCNs). As a result, they are constrained by fixed structures and do not naturally generalize to unseen nodes. Here, we discover that transductive GCNs inherently possess a distillable robustness, achieved through a wave-induced resonance process. Based on this, we foster this resonance to facilitate inductive and robust learning. Specifically, we first prove that the signal formed by GCN-driven message passing (MP) is equivalent to the edge-based Laplacian wave, where, within a wave system, resonance can naturally emerge between the signal and its transmitting medium. This resonance provides inherent resistance to malicious perturbations inflicted on the signal system. We then prove that merely three MP iterations within GCNs can induce signal resonance between nodes and edges, manifesting as a coupling between nodes and their distillable surrounding local subgraph. Consequently, we present Graph Resonance-fostering Network (GRN) to foster this resonance via learning node representations from their distilled resonating subgraphs. By capturing the edge-transmitted signals within this subgraph and integrating them with the node signal, GRN embeds these combined signals into the central node's representation. This node-wise embedding approach allows for generalization to unseen nodes. We validate our theoretical findings with experiments, and demonstrate that GRN generalizes robustness to unseen nodes, whilst maintaining state-of-the-art classification accuracy on perturbed graphs. Appendices can be found on arXiv version: https://arxiv.org/abs/2312.08651

Deep Metric Learning (DML) plays an important role in modern computer vision research, where we learn a distance metric for a set of image representations. Recent DML techniques utilize the proxy to interact with the corresponding image samples in the embedding space. However, existing proxy-based DML methods focus on learning individual proxy-to-sample distance, while the overall distribution of samples and proxies lacks attention. In this paper, we present a novel proxy-based DML framework that focuses on aligning the sample and proxy distributions to improve the efficiency of proxy-based DML losses. Specifically, we propose the Data-Augmented Domain Adaptation (DADA) method to adapt the domain gap between the group of samples and proxies. To the best of our knowledge, we are the first to leverage domain adaptation to boost the performance of proxy-based DML. We show that our method can be easily plugged into existing proxy-based DML losses. Our experiments on benchmarks, including the popular CUB-200-2011, CARS196, Stanford Online Products, and In-Shop Clothes Retrieval, show that our learning algorithm significantly improves the existing proxy losses and achieves superior results compared to the existing methods. The code and Appendix are available at: https://github.com/Noahsark/DADA

Relational structured data is a way of representing knowledge using nodes and edges, while also capturing the meaning of that knowledge in a structured form that can be used for machine learning. Compared with vision and natural language data, relational structured data represents and manipulates structured knowledge, which can be beneficial for tasks that involve reasoning or inference. On the other hand, vision and NLP deal more with unstructured data (like images and text), and they often require different types of models and algorithms to extract useful information or features from the data. Human-like Learning develops methods that can harness relational structures and learning-to-learn to rapidly acquire and generalize knowledge to new tasks and situations. With Human-like Learning, the learning algorithm is efficient and can adapt to new or unseen situations, which is crucial in real-world applications where environments may change unpredictably. Moreover, the models are easier for humans to understand and interpret, which is important for transparency and trust in AI systems. In this talk, we present our recent attempts towards human-like learning from relational structured data.

Language acquisition and utilization transcend the mere exchange of lexical units. Visual cues, prosody, gestures, body movements, and context play an undeniably crucial role. Humans naturally communicate multimodally, employing multiple channels and synthesizing information from diverse modalities. My research delves into the characterization and construction of multimodal models that seamlessly integrate data from multiple independent modalities. I will cover recent work that highlights the challenges, achievements, and opportunities towards developing capable multimodal discursive models.