AAAI 2023 Papers

It has become a common practice for many perceptual systems to integrate information from multiple sensors to improve the accuracy of object detection. For example, autonomous vehicles use visible light, and infrared (IR) information to ensure that the car can cope with complex weather conditions. However, the accuracy of the algorithm is usually a trade-off between the computational complexity and memory consumption. In this study, we evaluate the performance and complexity of different fusion operators in multi-modal object detection tasks. On top of that, a Poolformer-based fusion operator (PoolFuser) is proposed to enhance the accuracy of detecting targets without compromising the efficiency of the detection framework.

Dependency stochastic Boolean satisfiability (DSSAT) generalizes stochastic Boolean satisfiability (SSAT) in existential variables being Henkinized allowing their dependencies on randomized variables to be explicitly specified. It allows NEXPTIME problems of reasoning under uncertainty and partial information to be compactly encoded. To date, no decision procedure has been implemented for solving DSSAT formulas. This work provides the first such tool by converting DSSAT into SSAT with dependency elimination, similar to converting dependency quantified Boolean formula (DQBF) to quantified Boolean formula (QBF). Moreover, we extend (D)QBF preprocessing techniques and implement the first standalone (D)SSAT preprocessor. Experimental results show that solving DSSAT via dependency elimination is highly applicable and that existing SSAT solvers may benefit from preprocessing.

We consider the task of weighted first-order model counting (WFOMC) used for probabilistic inference in the area of statistical relational learning. Given a formula φ, domain size n and a pair of weight functions, what is the weighted sum of all models of φ over a domain of size n? It was shown that computing WFOMC of any logical sentence with at most two logical variables can be done in time polynomial in n. However, it was also shown that the task is #P1-complete once we add the third variable, which inspired the search for extensions of the two-variable fragment that would still permit a running time polynomial in n. One of such extension is the two-variable fragment with counting quantifiers. In this paper, we prove that adding a linear order axiom (which forces one of the predicates in φ to introduce a linear ordering of the domain elements in each model of φ) on top of the counting quantifiers still permits a computation time polynomial in the domain size. We present a new dynamic programming-based algorithm which can compute WFOMC with linear order in time polynomial in n, thus proving our primary claim.

The Variational Autoencoder (VAE) suffers from a significant loss of information when trained on a non-stationary data distribution. This loss in VAE models, called catastrophic forgetting, has not been studied theoretically before. We analyse the forgetting behaviour of a VAE in continual generative modelling by developing a new lower bound on the data likelihood, which interprets the forgetting process as an increase in the probability distance between the generator's distribution and the evolved data distribution. The proposed bound shows that a VAE-based dynamic expansion model can achieve better performance if its capacity increases appropriately considering the shift in the data distribution. Based on this analysis, we propose a novel expansion criterion that aims to preserve the information diversity among the VAE components, while ensuring that it acquires more knowledge with fewer parameters. Specifically, we implement this expansion criterion from the perspective of a multi-player game and propose the Online Adversarial Expansion Strategy (OAES), which considers all previously learned components as well as the currently updated component as multiple players in a game, while an adversary model evaluates their performance. The proposed OAES can dynamically estimate the discrepancy between each player and the adversary without accessing task information. This leads to the gradual addition of new components while ensuring the knowledge diversity among all of them. We show theoretically and empirically that the proposed extension strategy can enable a VAE model to achieve the best performance given an appropriate model size.

Lifelong person re-identification (LReID) is in significant demand for real-world development as a large amount of ReID data is captured from diverse locations over time and cannot be accessed at once inherently. However, a key challenge for LReID is how to incrementally preserve old knowledge and gradually add new capabilities to the system. Unlike most existing LReID methods, which mainly focus on dealing with catastrophic forgetting, our focus is on a more challenging problem, which is, not only trying to reduce the forgetting on old tasks but also aiming to improve the model performance on both new and old tasks during the lifelong learning process. Inspired by the biological process of human cognition where the somatosensory neocortex and the hippocampus work together in memory consolidation, we formulated a model called Knowledge Refreshing and Consolidation (KRC) that achieves both positive forward and backward transfer. More specifically, a knowledge refreshing scheme is incorporated with the knowledge rehearsal mechanism to enable bi-directional knowledge transfer by introducing a dynamic memory model and an adaptive working model. Moreover, a knowledge consolidation scheme operating on the dual space further improves model stability over the long-term. Extensive evaluations show KRC’s superiority over the state-of-the-art LReID methods with challenging pedestrian benchmarks. Code is available at https://github.com/cly234/LReID-KRKC.

Existing knowledge graph (KG) embedding models have primarily focused on static KGs. However, real-world KGs do not remain static, but rather evolve and grow in tandem with the development of KG applications. Consequently, new facts and previously unseen entities and relations continually emerge, necessitating an embedding model that can quickly learn and transfer new knowledge through growth. Motivated by this, we delve into an expanding field of KG embedding in this paper, i.e., lifelong KG embedding. We consider knowledge transfer and retention of the learning on growing snapshots of a KG without having to learn embeddings from scratch. The proposed model includes a masked KG autoencoder for embedding learning and update, with an embedding transfer strategy to inject the learned knowledge into the new entity and relation embeddings, and an embedding regularization method to avoid catastrophic forgetting. To investigate the impacts of different aspects of KG growth, we construct four datasets to evaluate the performance of lifelong KG embedding. Experimental results show that the proposed model outperforms the state-of-the-art inductive and lifelong embedding baselines.

Task-Free Continual Learning (TFCL) represents a challenging scenario for lifelong learning because the model, under this paradigm, does not access any task information. The Dynamic Expansion Model (DEM) has shown promising results in this scenario due to its scalability and generalisation power. However, DEM focuses only on addressing forgetting and ignores minimizing the model size, which limits its deployment in practical systems. In this work, we aim to simultaneously address network forgetting and model size optimization by developing the Lifelong Compression Mixture Model (LGMM) equipped with the Maximum Mean Discrepancy (MMD) based expansion criterion for model expansion. A diversity-aware sample selection approach is proposed to selectively store a variety of samples to promote information diversity among the components of the LGMM, which allows more knowledge to be captured with an appropriate model size. In order to avoid having multiple components with similar knowledge in the LGMM, we propose a data-free component discarding mechanism that evaluates a knowledge relation graph matrix describing the relevance between each pair of components. A greedy selection procedure is proposed to identify and remove the redundant components from the LGMM. The proposed discarding mechanism can be performed during or after the training. Experiments on different datasets show that LGMM achieves the best performance for TFCL.

LiDAR-based 3D object detection, semantic segmentation, and panoptic segmentation are usually implemented in specialized networks with distinctive architectures that are difficult to adapt to each other. This paper presents LidarMultiNet, a LiDAR-based multi-task network that unifies these three major LiDAR perception tasks. Among its many benefits, a multi-task network can reduce the overall cost by sharing weights and computation among multiple tasks. However, it typically underperforms compared to independently combined single-task models. The proposed LidarMultiNet aims to bridge the performance gap between the multi-task network and multiple single-task networks. At the core of LidarMultiNet is a strong 3D voxel-based encoder-decoder architecture with a Global Context Pooling (GCP) module extracting global contextual features from a LiDAR frame. Task-specific heads are added on top of the network to perform the three LiDAR perception tasks. More tasks can be implemented simply by adding new task-specific heads while introducing little additional cost. A second stage is also proposed to refine the first-stage segmentation and generate accurate panoptic segmentation results. LidarMultiNet is extensively tested on both Waymo Open Dataset and nuScenes dataset, demonstrating for the first time that major LiDAR perception tasks can be unified in a single strong network that is trained end-to-end and achieves state-of-the-art performance. Notably, LidarMultiNet reaches the official 1 place in the Waymo Open Dataset 3D semantic segmentation challenge 2022 with the highest mIoU and the best accuracy for most of the 22 classes on the test set, using only LiDAR points as input. It also sets the new state-of-the-art for a single model on the Waymo 3D object detection benchmark and three nuScenes benchmarks.

Video captioning has become a broad and interesting research area. Attention-based encoder-decoder methods are extensively used for caption generation. However, these methods mostly utilize the visual attentive feature to highlight the video regions while overlooked the semantic features of the available captions. These semantic features contain significant information that helps to generate highly informative human description-like captions. Therefore, we propose a novel visual and semantic enhanced video captioning network, named as VSVCap, that efficiently utilizes multiple ground truth captions. We aim to generate captions that are visually and semantically enhanced by exploiting both video and text modalities. To achieve this, we propose a fine-grained cross-graph attention mechanism that captures detailed graph embedding correspondence between visual graphs and textual knowledge graphs. We have performed node-level matching and structure-level reasoning between the weighted regional graph and knowledge graph. The proposed network achieves promising results on three benchmark datasets, i.e., YouTube2Text, MSR-VTT, and VATEX. The experimental results show that our network accurately captures all key objects, relationships, and semantically enhanced events of a video to generate human annotation-like captions.

Compositional Zero-Shot Learning (CZSL) aims at identifying unseen compositions composed of previously seen attributes and objects during the test phase. In real images, the visual appearances of attributes and objects (primitive concepts) generally interact with each other. Namely, the visual appearances of an attribute may change when composed with different objects, and vice versa. But previous works overlook this important property. In this paper, we introduce a simple yet effective approach with leveraging sub-class discrimination. Specifically, we define the primitive concepts in different compositions as sub-classes, and then maintain the sub-class discrimination to address the above challenge. More specifically, inspired by the observation that the composed recognition models could account for the differences across sub-classes, we first propose to impose the embedding alignment between the composed and disentangled recognition to incorporate sub-class discrimination at the feature level. Then we develop the prototype modulator networks to adjust the class prototypes w.r.t. the composition information, which can enhance sub-class discrimination at the classifier level. We conduct extensive experiments on the challenging benchmark datasets, and the considerable performance improvement over state-of-the-art approaches is achieved, which indicates the effectiveness of our method. Our code is available at https://github.com/hxm97/SCD-CZSL.

We consider semi-supervised binary classification for applications in which data points are naturally grouped (e.g., survey responses grouped by state) and the labeled data is biased (e.g., survey respondents are not representative of the population). The groups overlap in the feature space and consequently the input-output patterns are related across the groups. To model the inherent structure in such data, we assume the partition-projected class-conditional invariance across groups, defined in terms of the group-agnostic feature space. We demonstrate that under this assumption, the group carries additional information about the class, over the group-agnostic features, with provably improved area under the ROC curve. Further assuming invariance of partition-projected class-conditional distributions across both labeled and unlabeled data, we derive a semi-supervised algorithm that explicitly leverages the structure to learn an optimal, group-aware, probability-calibrated classifier, despite the bias in the labeled data. Experiments on synthetic and real data demonstrate the efficacy of our algorithm over suitable baselines and ablative models, spanning standard supervised and semi-supervised learning approaches, with and without incorporating the group directly as a feature.

Class-incremental continual learning is a core step towards developing artificial intelligence systems that can continuously adapt to changes in the environment by learning new concepts without forgetting those previously learned. This is especially needed in the medical domain where continually learning from new incoming data is required to classify an expanded set of diseases. In this work, we focus on how old knowledge can be leveraged to learn new classes without catastrophic forgetting. We propose a framework that comprises of two main components: (1) a dynamic architecture with expanding representations to preserve previously learned features and accommodate new features; and (2) a training procedure alternating between two objectives to balance the learning of new features while maintaining the model’s performance on old classes. Experiment results on multiple medical datasets show that our solution is able to achieve superior performance over state-of-the-art baselines in terms of class accuracy and forgetting.

Audio-visual speech recognition (AVSR) has gained remarkable success for ameliorating the noise-robustness of speech recognition. Mainstream methods focus on fusing audio and visual inputs to obtain modality-invariant representations. However, such representations are prone to over-reliance on audio modality as it is much easier to recognize than video modality in clean conditions. As a result, the AVSR model underestimates the importance of visual stream in face of noise corruption. To this end, we leverage visual modality-specific representations to provide stable complementary information for the AVSR task. Specifically, we propose a reinforcement learning (RL) based framework called MSRL, where the agent dynamically harmonizes modality-invariant and modality-specific representations in the auto-regressive decoding process. We customize a reward function directly related to task-specific metrics (i.e., word error rate), which encourages the MSRL to effectively explore the optimal integration strategy. Experimental results on the LRS3 dataset show that the proposed method achieves state-of-the-art in both clean and various noisy conditions. Furthermore, we demonstrate the better generality of MSRL system than other baselines when test set contains unseen noises.

Generative adversarial networks (GANs) are known for their strong abilities on capturing the underlying distribution of training instances. Since the seminal work of GAN, many variants of GAN have been proposed. However, existing GANs are almost established on the assumption that the training dataset is clean. But in many real-world applications, this may not hold, that is, the training dataset may be contaminated by a proportion of undesired instances. When training on such datasets, existing GANs will learn a mixture distribution of desired and contaminated instances, rather than the desired distribution of desired data only (target distribution). To learn the target distribution from contaminated datasets, two purified generative adversarial networks (PuriGAN) are developed, in which the discriminators are augmented with the capability to distinguish between target and contaminated instances by leveraging an extra dataset solely composed of contamination instances. We prove that under some mild conditions, the proposed PuriGANs are guaranteed to converge to the distribution of desired instances. Experimental results on several datasets demonstrate that the proposed PuriGANs are able to generate much better images from the desired distribution than comparable baselines when trained on contaminated datasets. In addition, we also demonstrate the usefulness of PuriGAN on downstream applications by applying it to the tasks of semi-supervised anomaly detection on contaminated datasets and PU-learning. Experimental results show that PuriGAN is able to deliver the best performance over comparable baselines on both tasks.

In the past few years, numerous multi-view graph clustering algorithms have been proposed to enhance the clustering performance by exploring information from multiple views. Despite the superior performance, the high time and space expenditures limit their scalability. Accordingly, anchor graph learning has been introduced to alleviate the computational complexity. However, existing approaches can be further improved by the following considerations: (i) Existing anchor-based methods share the same number of anchors across views. This strategy violates the diversity and flexibility of multi-view data distribution. (ii) Searching for the optimal anchor number within hyper-parameters takes much extra tuning time, which makes existing methods impractical. (iii) How to flexibly fuse multi-view anchor graphs of diverse sizes has not been well explored in existing literature. To address the above issues, we propose a novel anchor-based method termed Flexible and Diverse Anchor Graph Fusion for Scalable Multi-view Clustering (FDAGF) in this paper. Instead of manually tuning optimal anchor with massive hyper-parameters, we propose to optimize the contribution weights of a group of pre-defined anchor numbers to avoid extra time expenditure among views. Most importantly, we propose a novel hybrid fusion strategy for multi-size anchor graphs with theoretical proof, which allows flexible and diverse anchor graph fusion. Then, an efficient linear optimization algorithm is proposed to solve the resultant problem. Comprehensive experimental results demonstrate the effectiveness and efficiency of our proposed framework. The source code is available at https://github.com/Jeaninezpp/FDAGF.

Graph Neural Networks (GNNs) have drawn significant attentions over the years and been broadly applied to essential applications requiring solid robustness or vigorous security standards, such as product recommendation and user behavior modeling. Under these scenarios, exploiting GNN's vulnerabilities and further downgrading its performance become extremely incentive for adversaries. Previous attackers mainly focus on structural perturbations or node injections to the existing graphs, guided by gradients from the surrogate models. Although they deliver promising results, several limitations still exist. For the structural perturbation attack, to launch a proposed attack, adversaries need to manipulate the existing graph topology, which is impractical in most circumstances. Whereas for the node injection attack, though being more practical, current approaches require training surrogate models to simulate a white-box setting, which results in significant performance downgrade when the surrogate architecture diverges from the actual victim model. To bridge these gaps, in this paper, we study the problem of black-box node injection attack, without training a potentially misleading surrogate model. Specifically, we model the node injection attack as a Markov decision process and propose Gradient-free Graph Advantage Actor Critic, namely G2A2C, a reinforcement learning framework in the fashion of advantage actor critic. By directly querying the victim model, G2A2C learns to inject highly malicious nodes with extremely limited attacking budgets, while maintaining a similar node feature distribution. Through our comprehensive experiments over eight acknowledged benchmark datasets with different characteristics, we demonstrate the superior performance of our proposed G2A2C over the existing state-of-the-art attackers. Source code is publicly available at: https://github.com/jumxglhf/G2A2C.

Earnings conference calls are indicative information events for volatility forecasting, which is essential for financial risk management and asset pricing. Although recent volatility forecasting models have explored the textual content of conference calls for prediction, they suffer from modeling the long-text and representing the risk-relevant information. This work proposes to identify key sentences for robust and interpretable transcript representation learning based on the cognitive theory. Specifically, we introduce TextRank to find key sentences and leverage attention mechanism to screen out the candidates by modeling the semantic correlations. Upon on the structural information of earning conference calls, we propose a structure-based contrastive learning method to facilitate the effective transcript representation. Empirical results on the benchmark dataset demonstrate the superiority of our model over competitive baselines in volatility forecasting.

Attention modules, which adaptively weight and refine features according to the importance of the input, have become a critical technique to boost the capability of convolutional neural networks. However, most existing attention modules are heuristic without a sound interpretation, and thus, require empirical engineering to design structure and operators within the modules. To handle the above issue, based on our 'less is more important' observation, we propose an Attention Module guided by Probability Density Function (PDF), dubbed PdfAM, which enjoys a rational motivation and requires few empirical structure designs. Concretely, we observe that pixels with less occurrence are prone to be textural details or foreground objects with much importance to aid vision tasks. Thus, with PDF values adopted as a smooth and anti-noise alternative to the pixel occurrence frequency, we design our PdfAM by first estimating the PDF based on some distribution assumption, and then predicting a 3D attention map via applying a negative correlation between the attention weights and the estimated PDF values. Furthermore, we develop learnable PDF-rescale parameters so as to adaptively transform the estimated PDF and predict a customized negative correlation. Experiments show that our PdfAM consistently boosts various networks under both high- and low-level vision tasks, and also performs favorably against other attention modules in terms of accuracy and convergence.

We study the online Traveling Salesman Problem (TSP) on the line augmented with machine-learned predictions. In the classical problem, there is a stream of requests released over time along the real line. The goal is to minimize the makespan of the algorithm. We distinguish between the open variant and the closed one, in which we additionally require the algorithm to return to the origin after serving all requests. The state of the art is a 1.64-competitive algorithm and a 2.04-competitive algorithm for the closed and open variants, respectively. In both cases, a tight lower bound is known. In both variants, our primary prediction model involves predicted positions of the requests. We introduce algorithms that (i) obtain a tight 1.5 competitive ratio for the closed variant and a 1.66 competitive ratio for the open variant in the case of perfect predictions, (ii) are robust against unbounded prediction error, and (iii) are smooth, i.e., their performance degrades gracefully as the prediction error increases. Moreover, we further investigate the learning-augmented setting in the open variant by additionally considering a prediction for the last request served by the optimal offline algorithm. Our algorithm for this enhanced setting obtains a 1.33 competitive ratio with perfect predictions while also being smooth and robust, beating the lower bound of 1.44 we show for our original prediction setting for the open variant. Also, we provide a lower bound of 1.25 for this enhanced setting.