AAAI 2023 Papers

Many deep spatio-temporal learning methods have been proposed for crowd flow modeling in recent years. However, most of them focus on designing a spatial and temporal convolution mechanism to aggregate information from nearby nodes and historical observations for a pre-defined prediction task. Different from the existing research, this paper aims to provide a generic and dynamic representation learning method for crowd flow modeling. The main idea of our method is to maintain a continuous-time representation for each node, and update the representations of all nodes continuously according to the streaming observed data. Along this line, a particular encoder-decoder architecture is proposed, where the encoder converts the newly happened transactions into a timestamped message, and then the representations of related nodes are updated according to the generated message. The role of the decoder is to guide the representation learning process by reconstructing the observed transactions based on the most recent node representations. Moreover, a number of virtual nodes are added to discover macro-level spatial patterns and also share the representations among spatially-interacted stations. Experiments have been conducted on two real-world datasets for four popular prediction tasks in crowd flow modeling. The result demonstrates that our method could achieve better prediction performance for all the tasks than baseline methods.

Computer vision applications for document image understanding (DIU) such as optical character recognition, word spotting, enhancement etc. suffer from structural deformations like strike-outs and unconstrained strokes, to name a few. They also suffer from texture degradation due to blurring, aging, or blotting-spots etc. The DIU applications with deep networks are limited to constrained environment and lack diverse data with text-level and pixel-level annotation simultaneously. In this work, we propose a generative framework to produce realistic synthetic handwritten document images with simultaneous annotation of text and corresponding pixel-level spatial foreground information. The proposed approach generates realistic backgrounds with artificial handwritten texts which supplements data-augmentation in multiple unconstrained DIU systems. The proposed framework is an early work to facilitate DIU system-evaluation in both image quality and recognition performance at a go.

Label distribution learning (LDL) is an effective learning paradigm for dealing with label ambiguity. When applying LDL, the datasets annotated with label distributions (i.e., the real-valued vectors like the probability distribution) are typically required. Unfortunately, most existing datasets only contain the logical labels, and manual annotating with label distributions is costly. To address this problem, we treat the label distribution as a latent vector and infer its posterior by variational Bayes. Specifically, we propose a generative label enhancement model to encode the process of generating feature vectors and logical label vectors from label distributions in a principled way. In terms of features, we assume that the feature vector is generated by a Gaussian mixture dominated by the label distribution, which captures the one-to-many relationship from the label distribution to the feature vector and thus reduces the feature generation error. In terms of logical labels, we design a probability distribution to generate the logical label vector from a label distribution, which captures partial label ranking in the logical label vector and thus provides a more accurate guidance for inferring the label distribution. Besides, to approximate the posterior of the label distribution, we design a inference model, and derive the variational learning objective. Finally, extensive experiments on real-world datasets validate our proposal.

This paper presents a new adversarial training framework for image inpainting with segmentation confusion adversarial training (SCAT) and contrastive learning. SCAT plays an adversarial game between an inpainting generator and a segmentation network, which provides pixel-level local training signals and can adapt to images with free-form holes. By combining SCAT with standard global adversarial training, the new adversarial training framework exhibits the following three advantages simultaneously: (1) the global consistency of the repaired image, (2) the local fine texture details of the repaired image, and (3) the flexibility of handling images with free-form holes. Moreover, we propose the textural and semantic contrastive learning losses to stabilize and improve our inpainting model's training by exploiting the feature representation space of the discriminator, in which the inpainting images are pulled closer to the ground truth images but pushed farther from the corrupted images. The proposed contrastive losses better guide the repaired images to move from the corrupted image data points to the real image data points in the feature representation space, resulting in more realistic completed images. We conduct extensive experiments on two benchmark datasets, demonstrating our model's effectiveness and superiority both qualitatively and quantitatively.

In the fields of natural language processing (NLP) and computer vision (CV), recent advances in generative modeling have led to powerful machine learning systems that can effectively learn from large labeled and unlabeled datasets. These systems, by and large, apply a uniform pretrain-finetune pipeline on sequential data streams and have achieved state-of-the-art-performance across many tasks and benchmarks. In this talk, we will present recent algorithms that extend this paradigm to sequential decision making, by casting it as an inverse problem that can be solved via deep generative models. These generative approaches are stable to train, provide a flexible interface for single- and multi-task inference, and generalize exceedingly well outside their training datasets. We instantiate these algorithms in the context of reinforcement learning and black-box optimization. Empirically, we demonstrate that these approaches perform exceedingly well on high-dimensional benchmarks outperforming the current state-of-the-art approaches based on forward models.

Recent studies have demonstrated that existing deep neural networks (DNNs) on 3D point clouds are vulnerable to adversarial examples, especially under the white-box settings where the adversaries have access to model parameters. However, adversarial 3D point clouds generated by existing white-box methods have limited transferability across different DNN architectures. They have only minor threats in real-world scenarios under the black-box settings where the adversaries can only query the deployed victim model. In this paper, we revisit the transferability of adversarial 3D point clouds. We observe that an adversarial perturbation can be randomly factorized into two sub-perturbations, which are also likely to be adversarial perturbations. It motivates us to consider the effects of the perturbation and its sub-perturbations simultaneously to increase the transferability for sub-perturbations also contain helpful information. In this paper, we propose a simple yet effective attack method to generate more transferable adversarial 3D point clouds. Specifically, rather than simply optimizing the loss of perturbation alone, we combine it with its random factorization. We conduct experiments on benchmark dataset, verifying our method's effectiveness in increasing transferability while preserving high efficiency.

Human guides in museums and galleries are professionally trained to stimulate informal learning in visitors by asking low-risk, open-ended reflective questions that enable them to focus on specific features of artifacts, relate to prior experiences, and elicit curiosity as well as further thought. We present ArtMuse, our AI-powered chatbot for asking reflective questions in context of paintings. Our reflective question generation model in ArtMuse was trained by applying a novel combination of existing models for extractive question answering and open-domain chitchat. User evaluation studies indicate that we are able to generate fluent and specific reflective questions for paintings that are highly-engaging.

Despite advances in generating fluent texts, existing pretraining models tend to attach incoherent event sequences to involved entities when generating narratives such as stories and news. We conjecture that such issues result from representing entities as static embeddings of superficial words, while neglecting to model their ever-changing states, i.e., the information they carry, as the text unfolds. Therefore, we extend the Transformer model to dynamically conduct entity state updates and sentence realization for narrative generation. We propose a contrastive framework to learn the state representations in a discrete space, and insert additional attention layers into the decoder to better exploit these states. Experiments on two narrative datasets show that our model can generate more coherent and diverse narratives than strong baselines with the guidance of meaningful entity states.

Although existing multi-object tracking (MOT) algorithms have obtained competitive performance on various benchmarks, almost all of them train and validate models on the same domain. The domain generalization problem of MOT is hardly studied. To bridge this gap, we first draw the observation that the high-level information contained in natural language is domain invariant to different tracking domains. Based on this observation, we propose to introduce natural language representation into visual MOT models for boosting the domain generalization ability. However, it is infeasible to label every tracking target with a textual description. To tackle this problem, we design two modules, namely visual context prompting (VCP) and visual-language mixing (VLM). Specifically, VCP generates visual prompts based on the input frames. VLM joints the information in the generated visual prompts and the textual prompts from a pre-defined Trackbook to obtain instance-level pseudo textual description, which is domain invariant to different tracking scenes. Through training models on MOT17 and validating them on MOT20, we observe that the pseudo textual descriptions generated by our proposed modules improve the generalization performance of query-based trackers by large margins.

Current math word problem (MWP) solvers are usually Seq2Seq models trained by the (one-problem; one-solution) pairs, each of which is made of a problem description and a solution showing reasoning flow to get the correct answer. However, one MWP problem naturally has multiple solution equations. The training of an MWP solver with (one-problem; one-solution) pairs excludes other correct solutions, and thus limits the generalizability of the MWP solver. One feasible solution to this limitation is to augment multiple solutions to a given problem. However, it is difficult to collect diverse and accurate augment solutions through human efforts. In this paper, we design a new training framework for an MWP solver by introducing a solution buffer and a solution discriminator. The buffer includes solutions generated by an MWP solver to encourage the training data diversity. The discriminator controls the quality of buffered solutions to participate in training. Our framework is flexibly applicable to a wide setting of fully, semi-weakly and weakly supervised training for all Seq2Seq MWP solvers. We conduct extensive experiments on a benchmark dataset Math23k and a new dataset named Weak12k, and show that our framework improves the performance of various MWP solvers under different settings by generating correct and diverse solutions.

Downsampling produces coarsened, multi-resolution representations of data and it is used, for example, to produce lossy compression and visualization of large images, reduce computational costs, and boost deep neural representation learning. Unfortunately, due to their lack of a regular structure, there is still no consensus on how downsampling should apply to graphs and linked data. Indeed reductions in graph data are still needed for the goals described above, but reduction mechanisms do not have the same focus on preserving topological structures and properties, while allowing for resolution-tuning, as is the case in regular data downsampling. In this paper, we take a step in this direction, introducing a unifying interpretation of downsampling in regular and graph data. In particular, we define a graph coarsening mechanism which is a graph-structured counterpart of controllable equispaced coarsening mechanisms in regular data. We prove theoretical guarantees for distortion bounds on path lengths, as well as the ability to preserve key topological properties in the coarsened graphs. We leverage these concepts to define a graph pooling mechanism that we empirically assess in graph classification tasks, providing a greedy algorithm that allows efficient parallel implementation on GPUs, and showing that it compares favorably against pooling methods in literature.

Deep networks trained on the source domain show degraded performance when tested on unseen target domain data. To enhance the model's generalization ability, most existing domain generalization methods learn domain invariant features by suppressing domain sensitive features. Different from them, we propose a Domain Projection and Contrastive Learning (DPCL) approach for generalized semantic segmentation, which includes two modules: Self-supervised Source Domain Projection (SSDP) and Multi-Level Contrastive Learning (MLCL). SSDP aims to reduce domain gap by projecting data to the source domain, while MLCL is a learning scheme to learn discriminative and generalizable features on the projected data. During test time, we first project the target data by SSDP to mitigate domain shift, then generate the segmentation results by the learned segmentation network based on MLCL. At test time, we can update the projected data by minimizing our proposed pixel-to-pixel contrastive loss to obtain better results. Extensive experiments for semantic segmentation demonstrate the favorable generalization capability of our method on benchmark datasets.

In robust optimization, finding a solution that solely respects the constraints is not enough. Usually, the uncertainty and unknown parameters of the model are represented by random variables. In such conditions, a good solution is a solution robust to most-likely assignments of these random variables. Recently, the Confidence constraint has been introduced by Mercier-Aubin et al. in order to enforce this type of robustness in constraint programming. Unfortunately, it is restricted to a conjunction of binary inequalities In this paper, we generalize the Confidence constraint to any constraint and propose an implementation based on Multi-valued Decision Diagrams (MDDs). The Confidence constraint is defined over a vector of random variables. For a given constraint C, and given a threshold, the Confidence constraint ensures that the probability for C to be satisfied by a sample of the random variables is greater than the threshold. We propose to use MDDs to represent the constraints on the random variables. MDDs are an efficient tool for representing combinatorial constraints, thanks to their exponential compression power. Here, both random and decision variables are stored in the MDD, and propagation rules are proposed for removing values of decision variables that cannot lead to robust solutions. Furthermore, for several constraints, we show that decision variables can be omitted from the MDD because lighter filtering algorithms are sufficient. This leads to gain an exponential factor in the MDD size. The experimental results obtained on a chemical deliveries problem in factories – where the chemicals consumption are uncertain – shows the efficiency of the proposed approach.

The rapid development of single-cell RNA sequencing (scRNA-seq) technology allows us to study gene expression heterogeneity at the cellular level. Cell annotation is the basis for subsequent downstream analysis in single-cell data mining. Existing methods rarely explore the fine-grained semantic knowledge of novel cell types absent from the reference data and usually susceptible to batch effects on the classification of seen cell types. Taking into consideration these limitations, this paper proposes a new and practical task called generalized cell type annotation and discovery for scRNA-seq data. In this task, cells of seen cell types are given class labels, while cells of novel cell types are given cluster labels instead of a unified “unassigned” label. To address this problem, we carefully design a comprehensive evaluation benchmark and propose a novel end-to-end algorithm framework called scGAD. Specifically, scGAD first builds the intrinsic correspondence across the reference and target data by retrieving the geometrically and semantically mutual nearest neighbors as anchor pairs. Then we introduce an anchor-based self-supervised learning module with a connectivity-aware attention mechanism to facilitate model prediction capability on unlabeled target data. To enhance the inter-type separation and intra-type compactness, we further propose a confidential prototypical self-supervised learning module to uncover the consensus category structure of the reference and target data. Extensive results on massive real datasets demonstrate the superiority of scGAD over various state-of-the-art clustering and annotation methods.

Generalized Category Discovery (GCD) aims to recognize both known and novel categories from a set of unlabeled data, based on another dataset labeled with only known categories. Without considering differences between known and novel categories, current methods learn about them in a coupled manner, which can hurt model's generalization and discriminative ability. Furthermore, the coupled training approach prevents these models transferring category-specific knowledge explicitly from labeled data to unlabeled data, which can lose high-level semantic information and impair model performance. To mitigate above limitations, we present a novel model called Decoupled Prototypical Network (DPN). By formulating a bipartite matching problem for category prototypes, DPN can not only decouple known and novel categories to achieve different training targets effectively, but also align known categories in labeled and unlabeled data to transfer category-specific knowledge explicitly and capture high-level semantics. Furthermore, DPN can learn more discriminative features for both known and novel categories through our proposed Semantic-aware Prototypical Learning (SPL). Besides capturing meaningful semantic information, SPL can also alleviate the noise of hard pseudo labels through semantic-weighted soft assignment. Extensive experiments show that DPN outperforms state-of-the-art models by a large margin on all evaluation metrics across multiple benchmark datasets. Code and data are available at https://github.com/Lackel/DPN.

We study inductive matrix completion (matrix completion with side information) under an i.i.d. subgaussian noise assumption at a low noise regime, with uniform sampling of the entries. We obtain for the first time generalization bounds with the following three properties: (1) they scale like the standard deviation of the noise and in particular approach zero in the exact recovery case; (2) even in the presence of noise, they converge to zero when the sample size approaches infinity; and (3) for a fixed dimension of the side information, they only have a logarithmic dependence on the size of the matrix. Differently from many works in approximate recovery, we present results both for bounded Lipschitz losses and for the absolute loss, with the latter relying on Talagrand-type inequalities. The proofs create a bridge between two approaches to the theoretical analysis of matrix completion, since they consist in a combination of techniques from both the exact recovery literature and the approximate recovery literature.

The disjunctive skolem chase is a sound, complete, and potentially non-terminating procedure for solving boolean conjunctive query entailment over knowledge bases of disjunctive existential rules. We develop novel acyclicity and cyclicity notions for this procedure; that is, we develop sufficient conditions to determine chase termination and non-termination. Our empirical evaluation shows that our novel notions are significantly more general than existing criteria.

In many applications, we want to influence the decisions of independent agents by designing incentives for their actions. We revisit a fundamental problem in this area, called GAME IMPLEMENTATION: Given a game in standard form and a set of desired strategies, can we design a set of payment promises such that if the players take the payment promises into account, then all undominated strategies are desired? Furthermore, we aim to minimize the cost, that is, the worst-case amount of payments. We study the tractability of computing such payment promises and determine more closely what obstructions we may have to overcome in doing so. We show that GAME IMPLEMENTATION is NP-hard even for two players, solving in particular a long-standing open question and suggesting more restrictions are necessary to obtain tractability results. We thus study the regime in which players have only a small constant number of strategies and obtain the following. First, this case remains NP-hard even if each player’s utility depends only on three others. Second, we repair a flawed efficient algorithm for the case of both small number of strategies and small number of players. Among further results, we characterize sets of desired strategies that can be implemented at zero cost as a generalization of Nash equilibria.

The computation of trajectory similarity is a crucial task in many spatial data analysis applications. However, existing methods have been designed primarily for trajectories in Euclidean space, which overlooks the fact that real-world trajectories are often generated on road networks. This paper addresses this gap by proposing a novel framework, called GRLSTM (Graph-based Residual LSTM). To jointly capture the properties of trajectories and road networks, the proposed framework incorporates knowledge graph embedding (KGE), graph neural network (GNN), and the residual network into the multi-layer LSTM (Residual-LSTM). Specifically, the framework constructs a point knowledge graph to study the multi-relation of points, as points may belong to both the trajectory and the road network. KGE is introduced to learn point embeddings and relation embeddings to build the point fusion graph, while GNN is used to capture the topology structure information of the point fusion graph. Finally, Residual-LSTM is used to learn the trajectory embeddings.To further enhance the accuracy and robustness of the final trajectory embeddings, we introduce two new neighbor-based point loss functions, namely, graph-based point loss function and trajectory-based point loss function. The GRLSTM is evaluated using two real-world trajectory datasets, and the experimental results demonstrate that GRLSTM outperforms all the state-of-the-art methods significantly.