AAAI 2023 论文列表

Autoregressive models have achieved impressive results over a wide range of domains in terms of generation quality and downstream task performance. In the continuous domain, a key factor behind this success is the usage of quantized latent spaces (e.g., obtained via VQ-VAE autoencoders), which allow for dimensionality reduction and faster inference times. However, using existing pre-trained models to perform new non-trivial tasks is difficult since it requires additional fine-tuning or extensive training to elicit prompting. This paper introduces LASS as a way to perform vector-quantized Latent Autoregressive Source Separation (i.e., de-mixing an input signal into its constituent sources) without requiring additional gradient-based optimization or modifications of existing models. Our separation method relies on the Bayesian formulation in which the autoregressive models are the priors, and a discrete (non-parametric) likelihood function is constructed by performing frequency counts over latent sums of addend tokens. We test our method on images and audio with several sampling strategies (e.g., ancestral, beam search) showing competitive results with existing approaches in terms of separation quality while offering at the same time significant speedups in terms of inference time and scalability to higher dimensional data.

Hyper-heuristics are a domain-independent problem solving approach where the main task is to select effective chains of problem-specific low-level heuristics on the fly for an unseen instance. This task can be seen as a reinforcement learning problem, however, the information available to the hyper-heuristic is very limited, usually leading to very limited state representations. In this work, for the first time we use the trajectory of solution changes for a larger set of features for reinforcement learning in the novel hyper-heuristic LAST-RL (Large-State Reinforcement Learning). Further, we introduce a probability distribution for the exploration case in our epsilon-greedy policy that is based on the idea of Iterated Local Search to increase the chance to sample good chains of low-level heuristics. The benefit of the collaboration of our novel components is shown on the academic benchmark of the Cross Domain Heuristic Challenge 2011 consisting of six different problem domains. Our approach can provide state-of-the-art results on this benchmark where it outperforms recent hyper-heuristics based on reinforcement learning, and also demonstrates high performance on a benchmark of complex real-life personnel scheduling domains.

Learning descriptive 3D features is crucial for understanding 3D scenes with diverse objects and complex structures. However, it is usually unknown whether important geometric attributes and scene context obtain enough emphasis in an end-to-end trained 3D scene understanding network. To guide 3D feature learning toward important geometric attributes and scene context, we explore the help of textual scene descriptions. Given some free-form descriptions paired with 3D scenes, we extract the knowledge regarding the object relationships and object attributes. We then inject the knowledge to 3D feature learning through three classification-based auxiliary tasks. This language-assisted training can be combined with modern object detection and instance segmentation methods to promote 3D semantic scene understanding, especially in a label-deficient regime. Moreover, the 3D feature learned with language assistance is better aligned with the language features, which can benefit various 3D-language multimodal tasks. Experiments on several benchmarks of 3D-only and 3D-language tasks demonstrate the effectiveness of our language-assisted 3D feature learning. Code is available at https://github.com/Asterisci/Language-Assisted-3D.

Discriminative pre-trained language models (PrLMs) learn to predict original texts from intentionally corrupted ones. Taking the former text as positive and the latter as negative samples, the PrLM can be trained effectively for contextualized representation. However, the training of such a type of PrLMs highly relies on the quality of the automatically constructed samples. Existing PrLMs simply treat all corrupted texts as equal negative without any examination, which actually lets the resulting model inevitably suffer from the false negative issue where training is carried out on pseudo-negative data and leads to less efficiency and less robustness in the resulting PrLMs. In this work, on the basis of defining the false negative issue in discriminative PrLMs that has been ignored for a long time, we design enhanced pre-training methods to counteract false negative predictions and encourage pre-training language models on true negatives by correcting the harmful gradient updates subject to false negative predictions. Experimental results on GLUE and SQuAD benchmarks show that our counter-false-negative pre-training methods indeed bring about better performance together with stronger robustness.

Molecular representation learning is a fundamental problem in the field of drug discovery and molecular science. Whereas incorporating molecular 3D information in the representations of molecule seems beneficial, which is related to computational chemistry with the basic task of predicting stable 3D structures (conformations) of molecules. Existing machine learning methods either rely on 1D and 2D molecular properties or simulate molecular force field to use additional 3D structure information via Hamiltonian network. The former has the disadvantage of ignoring important 3D structure features, while the latter has the disadvantage that existing Hamiltonian neural network must satisfy the “canonial” constraint, which is difficult to be obeyed in many cases. In this paper, we propose a novel plug-and-play architecture LagNet by simulating molecular force field only with parameterized position coordinates, which implements Lagrangian mechanics to learn molecular representation by preserving 3D conformation without obeying any additional restrictions. LagNet is designed to generate known conformations and generalize for unknown ones from molecular SMILES. Implicit positions in LagNet are learned iteratively using discrete-time Lagrangian equations. Experimental results show that LagNet can well learn 3D molecular structure features, and outperforms previous state-of-the-art baselines related molecular representation by a significant margin.

Multi-label text classification (MLTC) involves tagging a document with its most relevant subset of labels from a label set. In real applications, labels usually follow a long-tailed distribution, where most labels (called as tail-label) only contain a small number of documents and limit the performance of MLTC. To facilitate this low-resource problem, researchers introduced a simple but effective strategy, data augmentation (DA). However, most existing DA approaches struggle in multi-label settings. The main reason is that the augmented documents for one label may inevitably influence the other co-occurring labels and further exaggerate the long-tailed problem. To mitigate this issue, we propose a new pair-level augmentation framework for MLTC, called Label-Specific Feature Augmentation (LSFA), which merely augments positive feature-label pairs for the tail-labels. LSFA contains two main parts. The first is for label-specific document representation learning in the high-level latent space, the second is for augmenting tail-label features in latent space by transferring the documents second-order statistics (intra-class semantic variations) from head labels to tail labels. At last, we design a new loss function for adjusting classifiers based on augmented datasets. The whole learning procedure can be effectively trained. Comprehensive experiments on benchmark datasets have shown that the proposed LSFA outperforms the state-of-the-art counterparts.

Automatically detecting emotions from text has countless applications, ranging from large scale opinion mining to social robots in healthcare and education. However, emotions are subjective in nature and are often expressed in ambiguous ways. At the same time, detecting emotions can also require implicit reasoning, which may not be available as surface- level, lexical information. In this work, we conjecture that the overconfidence of pre-trained language models such as BERT is a critical problem in emotion detection and show that alleviating this problem can considerably improve the generalization performance. We carry out comprehensive experiments on four emotion detection benchmark datasets and show that calibrating our model predictions leads to an average improvement of 1.35% in weighted F1 score.

We propose a novel complete algorithm for multi-agent pathfinding (MAPF) called lazy constraints addition search for MAPF (LaCAM). MAPF is a problem of finding collision-free paths for multiple agents on graphs and is the foundation of multi-robot coordination. LaCAM uses a two-level search to find solutions quickly, even with hundreds of agents or more. At the low-level, it searches constraints about agents' locations. At the high-level, it searches a sequence of all agents' locations, following the constraints specified by the low-level. Our exhaustive experiments reveal that LaCAM is comparable to or outperforms state-of-the-art sub-optimal MAPF algorithms in a variety of scenarios, regarding success rate, planning time, and solution quality of sum-of-costs.

Image instance segmentation is a fundamental research topic in autonomous driving, which is crucial for scene understanding and road safety. Advanced learning-based approaches often rely on the costly 2D mask annotations for training. In this paper, we present a more artful framework, LiDAR-guided Weakly Supervised Instance Segmentation (LWSIS), which leverages the off-the-shelf 3D data, i.e., Point Cloud, together with the 3D boxes, as natural weak supervisions for training the 2D image instance segmentation models. Our LWSIS not only exploits the complementary information in multimodal data during training but also significantly reduces the annotation cost of the dense 2D masks. In detail, LWSIS consists of two crucial modules, Point Label Assignment (PLA) and Graph-based Consistency Regularization (GCR). The former module aims to automatically assign the 3D point cloud as 2D point-wise labels, while the atter further refines the predictions by enforcing geometry and appearance consistency of the multimodal data. Moreover, we conduct a secondary instance segmentation annotation on the nuScenes, named nuInsSeg, to encourage further research on multimodal perception tasks. Extensive experiments on the nuInsSeg, as well as the large-scale Waymo, show that LWSIS can substantially improve existing weakly supervised segmentation models by only involving 3D data during training. Additionally, LWSIS can also be incorporated into 3D object detectors like PointPainting to boost the 3D detection performance for free. The code and dataset are available at https://github.com/Serenos/LWSIS.

Learning to recover clear images from images having a combination of degrading factors is a challenging task. That being said, autonomous surveillance in low visibility conditions caused by high pollution/smoke, poor air quality index, low light, atmospheric scattering, and haze during a blizzard, etc, becomes even more important to prevent accidents. It is thus crucial to form a solution that can not only result in a high-quality image but also which is efficient enough to be deployed for everyday use. However, the lack of proper datasets available to tackle this task limits the performance of the previous methods proposed. To this end, we generate the LowVis-AFO dataset, containing 3647 paired dark-hazy and clear images. We also introduce a new lightweight deep learning model called Low-Visibility Restoration Network (LVRNet). It outperforms previous image restoration methods with low latency, achieving a PSNR value of 25.744 and an SSIM of 0.905, hence making our approach scalable and ready for practical use.

AI systems can create, propagate, support, and automate bias in decision-making processes. To mitigate biased decisions, we both need to understand the origin of the bias and define what it means for an algorithm to make fair decisions. Most group fairness notions assess a model's equality of outcome by computing statistical metrics on the outputs. We argue that these output metrics encounter intrinsic obstacles and present a complementary approach that aligns with the increasing focus on equality of treatment. By Locating Unfairness through Canonical Inverse Design (LUCID), we generate a canonical set that shows the desired inputs for a model given a preferred output. The canonical set reveals the model's internal logic and exposes potential unethical biases by repeatedly interrogating the decision-making process. We evaluate LUCID on the UCI Adult and COMPAS data sets and find that some biases detected by a canonical set differ from those of output metrics. The results show that by shifting the focus towards equality of treatment and looking into the algorithm's internal workings, the canonical sets are a valuable addition to the toolbox of algorithmic fairness evaluation.

Table structure recognition (TSR) aims at extracting tables in images into machine-understandable formats. Recent methods solve this problem by predicting the adjacency relations of detected cell boxes, or learning to generate the corresponding markup sequences from the table images. However, they either count on additional heuristic rules to recover the table structures, or require a huge amount of training data and time-consuming sequential decoders. In this paper, we propose an alternative paradigm. We model TSR as a logical location regression problem and propose a new TSR framework called LORE, standing for LOgical location REgression network, which for the first time combines logical location regression together with spatial location regression of table cells. Our proposed LORE is conceptually simpler, easier to train and more accurate than previous TSR models of other paradigms. Experiments on standard benchmarks demonstrate that LORE consistently outperforms prior arts. Code is available at https:// github.com/AlibabaResearch/AdvancedLiterateMachinery/tree/main/DocumentUnderstanding/LORE-TSR.

Question answering (QA) models often rely on large-scale training datasets, which necessitates the development of a data generation framework to reduce the cost of manual annotations. Although several recent studies have aimed to generate synthetic questions with single-span answers, no study has been conducted on the creation of list questions with multiple, non-contiguous spans as answers. To address this gap, we propose LIQUID, an automated framework for generating list QA datasets from unlabeled corpora. We first convert a passage from Wikipedia or PubMed into a summary and extract named entities from the summarized text as candidate answers. This allows us to select answers that are semantically correlated in context and is, therefore, suitable for constructing list questions. We then create questions using an off-the-shelf question generator with the extracted entities and original passage. Finally, iterative filtering and answer expansion are performed to ensure the accuracy and completeness of the answers. Using our synthetic data, we significantly improve the performance of the previous best list QA models by exact-match F1 scores of 5.0 on MultiSpanQA, 1.9 on Quoref, and 2.8 averaged across three BioASQ benchmarks.

Mixed Integer programs (MIPs) are typically solved by the Branch-and-Bound algorithm. Recently, Learning to imitate fast approximations of the expert strong branching heuristic has gained attention due to its success in reducing the running time for solving MIPs. However, existing learning-to-branch methods assume that the entire training data is available in a single session of training. This assumption is often not true, and if the training data is supplied in continual fashion over time, existing techniques suffer from catastrophic forgetting. In this work, we study the hitherto unexplored paradigm of Lifelong Learning to Branch on Mixed Integer Programs. To mitigate catastrophic forgetting, we propose LIMIP, which is powered by the idea of modeling an MIP instance in the form of a bipartite graph, which we map to an embedding space using a bipartite Graph Attention Network. This rich embedding space avoids catastrophic forgetting through the application of knowledge distillation and elastic weight consolidation, wherein we learn the parameters key towards retaining efficacy and are therefore protected from significant drift. We evaluate LIMIP on a series of NP-hard problems and establish that in comparison to existing baselines, LIMIP is up to 50% better when confronted with lifelong learning

This paper presents an anomaly detection model that combines the strong statistical foundation of density-estimation-based anomaly detection methods with the representation-learning ability of deep-learning models. The method combines an autoencoder, that learns a low-dimensional representation of the data, with a density-estimation model based on density matrices in an end-to-end architecture that can be trained using gradient-based optimization techniques. A systematic experimental evaluation was performed on different benchmark datasets. The experimental results show that the method is able to outperform other state-of-the-art methods.

From the observation that users reading news tend to not click outdated news, we propose the notion of 'lifetime' of news, with two hypotheses: (i) news has a shorter lifetime, compared to other types of items such as movies or e-commerce products; (ii) news only competes with other news whose lifetimes have not ended, and which has an overlapping lifetime (i.e., limited competitions). By further developing the characteristics of the lifetime of news, then we present a novel approach for news recommendation, namely, Lifetime-Aware News reCommEndeR System (LANCER) that carefully exploits the lifetime of news during training and recommendation. Using real-world news datasets (e.g., Adressa and MIND), we successfully demonstrate that state-of-the-art news recommendation models can get significantly benefited by integrating the notion of lifetime and LANCER, by up to about 40% increases in recommendation accuracy.

Recently, word enhancement has become very popular for Chinese Named Entity Recognition (NER), reducing segmentation errors and increasing the semantic and boundary information of Chinese words. However, these methods tend to ignore the semantic relationship before and after the sentence after integrating lexical information. Therefore, the regularity of word length information has not been fully explored in various word-character fusion methods. In this work, we propose a Lexicon-Attention and Data-Augmentation (LADA) method for Chinese NER. We discuss the challenges of using existing methods in incorporating word information for NER and show how our proposed methods could be leveraged to overcome those challenges. LADA is based on a Transformer Encoder that utilizes lexicon to construct a directed graph and fuses word information through updating the optimal edge of the graph. Specially, we introduce the advanced data augmentation method to obtain the optimal representation for the NER task. Experimental results show that the augmentation done using LADA can considerably boost the performance of our NER system and achieve significantly better results than previous state-of-the-art methods and variant models in the literature on four publicly available NER datasets, namely Resume, MSRA, Weibo, and OntoNotes v4. We also observe better generalization and application to a real-world setting from LADA on multi-source complex entities.

Storytelling is an innate part of language-based communication. Today, current events are reported via Open Source Intelligence (OSINT) sources like news websites, blogs, and discussion forums. Scattered and fragmented sources such as these can be better understood when organized as chains of event plot points, or narratives, that have the ability to communicate end-end stories. Though search engines can retrieve aggregated event information, they lack the ability to sequence relevant events together to form narratives about different topics. I propose an AI system inspired by Gustav Freytag’s narrative theory called the Plot Element Pyramid and use knowledge graphs to represent, chain, and reason over narratives from disparately sourced event details to better comprehend convoluted, noisy information about critical events during intelligence analysis.

Open-ended Video question answering (open-ended VideoQA) aims to understand video content and question semantics to generate the correct answers. Most of the best performing models define the problem as a discriminative task of multi-label classification. In real-world scenarios, however, it is difficult to define a candidate set that includes all possible answers. In this paper, we propose a Knowledge-constrained Generative VideoQA Algorithm (KcGA) with an encoder-decoder pipeline, which enables out-of-domain answer generation through an adaptive external knowledge module and a multi-stream information control mechanism. We use ClipBERT to extract the video-question features, extract framewise object-level external knowledge from a commonsense knowledge base and compute the contextual-aware episode memory units via an attention based GRU to form the external knowledge features, and exploit multi-stream information control mechanism to fuse video-question and external knowledge features such that the semantic complementation and alignment are well achieved. We evaluate our model on two open-ended benchmark datasets to demonstrate that we can effectively and robustly generate high-quality answers without restrictions of training data.