AAAI 2023 Papers

With the advent of deep learning, a huge number of text-to-speech (TTS) models which produce human-like speech have emerged. Recently, by introducing syntactic and semantic information w.r.t the input text, various approaches have been proposed to enrich the naturalness and expressiveness of TTS models. Although these strategies showed impressive results, they still have some limitations in utilizing language information. First, most approaches only use graph networks to utilize syntactic and semantic information without considering linguistic features. Second, most previous works do not explicitly consider adjacent words when encoding syntactic and semantic information, even though it is obvious that adjacent words are usually meaningful when encoding the current word. To address these issues, we propose Relation-aware Word Encoding Network (RWEN), which effectively allows syntactic and semantic information based on two modules (i.e., Semantic-level Relation Encoding and Adjacent Word Relation Encoding). Experimental results show substantial improvements compared to previous works.

Active Object Tracking (AOT) aims to maintain a specific relation between the tracker and object(s) by autonomously controlling the motion system of a tracker given observations. It is widely used in various applications such as mobile robots and autonomous driving. However, Building a generalizable active tracker that works robustly across various scenarios remains a challenge, particularly in unstructured environments with cluttered obstacles and diverse layouts. To realize this, we argue that the key is to construct a state representation that can model the geometry structure of the surroundings and the dynamics of the target. To this end, we propose a framework called RSPT to form a structure-aware motion representation by Reconstructing Surroundings and Predicting the target Trajectory. Moreover, we further enhance the generalization of the policy network by training in the asymmetric dueling mechanism. Empirical results show that RSPT outperforms existing methods in unseen environments, especially those with cluttered obstacles and diverse layouts. We also demonstrate good sim-to-real transfer when deploying RSPT in real-world scenarios.

Multi-hop questions are associated with a series of justifications, and one needs to obtain the answers by following the reasoning path (RP) that orders the justifications adequately. So reasoning path retrieval becomes a critical preliminary stage for multi-hop Question Answering (QA). Within the RP, two fundamental challenges emerge for better performance: (i) what the order of the justifications in the RP should be, and (ii) what if the wrong justification has been in the path. In this paper, we propose Reasoning Path Augmentation (RPA), which uses reasoning path reordering and augmentation to handle the above two challenges, respectively. Reasoning path reordering restructures the reasoning by targeting the easier justification first but difficult one later, in which the difficulty is determined by the overlap between query and justifications since the higher overlap means more lexical relevance and easier searchable. Reasoning path augmentation automatically generates artificial RPs, in which the distracted justifications are inserted to aid the model recover from the wrong justification. We build RPA with a naive pre-trained model and evaluate RPA on the QASC and MultiRC datasets. The evaluation results demonstrate that RPA outperforms previously published reasoning path retrieval methods, showing the effectiveness of the proposed methods. Moreover, we present detailed experiments on how the orders of justifications and the percent of augmented paths affect the question- answering performance, revealing the importance of polishing RPs and the necessity of augmentation.

The exploration of mutual-benefit cross-domains has shown great potential toward accurate self-supervised depth estimation. In this work, we revisit feature fusion between depth and semantic information and propose an efficient local adaptive attention method for geometric aware representation enhancement. Instead of building global connections or deforming attention across the feature space without restraint, we bound the spatial interaction within a learnable region of interest. In particular, we leverage geometric cues from semantic information to learn local adaptive bounding boxes to guide unsupervised feature aggregation. The local areas preclude most irrelevant reference points from attention space, yielding more selective feature learning and faster convergence. We naturally extend the paradigm into a multi-head and hierarchic way to enable the information distillation in different semantic levels and improve the feature discriminative ability for fine-grained depth estimation. Extensive experiments on the KITTI dataset show that our proposed method establishes a new state-of-the-art in self-supervised monocular depth estimation task, demonstrating the effectiveness of our approach over former Transformer variants.

Whole-slide images (WSI) in computational pathology have high resolution with gigapixel size, but are generally with sparse regions of interest, which leads to weak diagnostic relevance and data inefficiency for each area in the slide. Most of the existing methods rely on a multiple instance learning framework that requires densely sampling local patches at high magnification. The limitation is evident in the application stage as the heavy computation for extracting patch-level features is inevitable. In this paper, we develop RLogist, a benchmarking deep reinforcement learning (DRL) method for fast observation strategy on WSIs. Imitating the diagnostic logic of human pathologists, our RL agent learns how to find regions of observation value and obtain representative features across multiple resolution levels, without having to analyze each part of the WSI at the high magnification. We benchmark our method on two whole-slide level classification tasks, including detection of metastases in WSIs of lymph node sections, and subtyping of lung cancer. Experimental results demonstrate that RLogist achieves competitive classification performance compared to typical multiple instance learning algorithms, while having a significantly short observation path. In addition, the observation path given by RLogist provides good decision-making interpretability, and its ability of reading path navigation can potentially be used by pathologists for educational/assistive purposes. Our code is available at: https://github.com/tencent-ailab/RLogist.

It is imperative to accelerate the training of neural network force field such as Deep Potential, which usually requires thousands of images based on first-principles calculation and a couple of days to generate an accurate potential energy surface. To this end, we propose a novel optimizer named reorganized layer extended Kalman filtering (RLEKF), an optimized version of global extended Kalman filtering (GEKF) with a strategy of splitting big and gathering small layers to overcome the O(N^2) computational cost of GEKF. This strategy provides an approximation of the dense weights error covariance matrix with a sparse diagonal block matrix for GEKF. We implement both RLEKF and the baseline Adam in our alphaDynamics package and numerical experiments are performed on 13 unbiased datasets. Overall, RLEKF converges faster with slightly better accuracy. For example, a test on a typical system, bulk copper, shows that RLEKF converges faster by both the number of training epochs (x11.67) and wall-clock time (x1.19). Besides, we theoretically prove that the updates of weights converge and thus are against the gradient exploding problem. Experimental results verify that RLEKF is not sensitive to the initialization of weights. The RLEKF sheds light on other AI-for-science applications where training a large neural network (with tons of thousands parameters) is a bottleneck.

Most approaches used in open-domain question answering on hybrid data that comprises both tabular-and-textual contents are based on a Retrieval-Reader pipeline in which the retrieval module finds relevant 
“heterogenous” evidence for a given question and the reader module generates an answer from the retrieved evidence. In this paper, we present a Retriever-Reranker-Reader framework by newly proposing a Reader-INherited evidence reranKer (RINK) where a reranker module is designed by finetuning the reader’s neural architecture based on a simple prompting method. Our underlying assumption of reusing the reader’s module for the reranker is that the reader’s ability to generating an answer from evidence contains the knowledge required for the reranking, because the reranker needs to “read” in-depth a question and evidences more carefully and elaborately than a baseline retriever. Furthermore, we present a simple and effective pretraining method by extensively deploying the commonly used data augmentation methods of cell corruption and cell reordering based on the pretraining tasks - tabular-and-textual entailment and cross-modal masked language modeling. Experimental results on OTT-QA, a large-scale table-and-text open-domain question answering dataset, show that the proposed RINK armed with our pretraining procedure makes improvements over the baseline reranking method and leads to state-of-the-art performance.

RGB-D object tracking has attracted considerable attention recently, achieving promising performance thanks to the symbiosis between visual and depth channels. However, given a limited amount of annotated RGB-D tracking data, most state-of-the-art RGB-D trackers are simple extensions of high-performance RGB-only trackers, without fully exploiting the underlying potential of the depth channel in the offline training stage. To address the dataset deficiency issue, a new RGB-D dataset named RGBD1K is released in this paper. The RGBD1K contains 1,050 sequences with about 2.5M frames in total. To demonstrate the benefits of training on a larger RGB-D data set in general, and RGBD1K in particular, we develop a transformer-based RGB-D tracker, named SPT, as a baseline for future visual object tracking studies using the new dataset. The results, of extensive experiments using the SPT tracker demonstrate the potential of the RGBD1K dataset to improve the performance of RGB-D tracking, inspiring future developments of effective tracker designs. The dataset and codes will be available on the project homepage: https://github.com/xuefeng-zhu5/RGBD1K.

Learning High-Resolution representations is essential for semantic segmentation. Convolutional neural network (CNN) architectures with downstream and upstream propagation flow are popular for segmentation in medical diagnosis. However, due to performing spatial downsampling and upsampling in multiple stages, information loss is inexorable. On the contrary, connecting layers densely on high spatial resolution is computationally expensive. In this work, we devise a Loose Dense Connection Strategy to connect neurons in subsequent layers with reduced parameters. On top of that, using a m-way Tree structure for feature propagation we propose Receptive Field Chain Network (RFC-Net) that learns high-resolution global features on a compressed computational space. Our experiments demonstrates that RFC Net achieves state-of-the-art performance on Kvasir and CVC-ClinicDB benchmarks for Polyp segmentation. Our code is publicly available at github.com/sourajitcs/RFC-NetAAAI23.

One of the recent best attempts at Text-to-SQL is the pre-trained language model. Due to the structural property of the SQL queries, the seq2seq model takes the responsibility of parsing both the schema items (i.e., tables and columns) and the skeleton (i.e., SQL keywords). Such coupled targets increase the difficulty of parsing the correct SQL queries especially when they involve many schema items and logic operators. This paper proposes a ranking-enhanced encoding and skeleton-aware decoding framework to decouple the schema linking and the skeleton parsing. Specifically, for a seq2seq encoder-decode model, its encoder is injected by the most relevant schema items instead of the whole unordered ones, which could alleviate the schema linking effort during SQL parsing, and its decoder first generates the skeleton and then the actual SQL query, which could implicitly constrain the SQL parsing. We evaluate our proposed framework on Spider and its three robustness variants: Spider-DK, Spider-Syn, and Spider-Realistic. The experimental results show that our framework delivers promising performance and robustness. Our code is available at https://github.com/RUCKBReasoning/RESDSQL.

Cold-start problem is one of the most challenging problems for recommender systems. One promising solution to this problem is cross-domain recommendation (CDR) which leverages rich information from an auxiliary source domain to improve the performance of recommender system in the target domain. In particular, the family of embedding and mapping methods for CDR is very effective, which explicitly learn a mapping function from source embeddings to target embeddings to transfer user’s preferences. Recent works usually transfer an overall source embedding by modeling a common or personalized preference bridge for all users. However, a unified user embedding cannot reflect the user’s multiple interests in auxiliary source domain. In this paper, we propose a novel framework called reinforced multi-interest transfer for CDR (REMIT). Specifically, we first construct a heterogeneous information network and employ different meta-path based aggregations to get user’s multiple interests in source domain, then transform different interest embeddings with different meta-generated personalized bridge functions for each user. To better coordinate the transformed user interest embeddings and the item embedding in target domain, we systematically develop a reinforced method to dynamically assign weights to transformed interests for different training instances and optimize the performance of target model. In addition, the REMIT is a general framework that can be applied upon various base models in target domain. Our extensive experimental results on large real-world datasets demonstrate the superior performance and compatibility of REMIT.

With the development of advanced driver assistance systems~(ADAS) and autonomous vehicles, conducting experiments in various scenarios becomes an urgent need. Although having been capable of synthesizing photo-realistic street scenes, conventional image-to-image translation methods cannot produce coherent scenes due to the lack of 3D information. In this paper, a large-scale neural rendering method is proposed to synthesize the autonomous driving scene~(READ), which makes it possible to generate large-scale driving scenes in real time on a PC through a variety of sampling schemes. In order to effectively represent driving scenarios, we propose an ω-net rendering network to learn neural descriptors from sparse point clouds. Our model can not only synthesize photo-realistic driving scenes but also stitch and edit them. The promising experimental results show that our model performs well in large-scale driving scenarios.

Detecting out-of-distribution (OOD) samples is crucial to the safe deployment of a classifier in the real world. However, deep neural networks are known to be overconfident for abnormal data. Existing works directly design score function by mining the inconsistency from classifier for in-distribution (ID) and OOD. In this paper, we further complement this inconsistency with reconstruction error, based on the assumption that an autoencoder trained on ID data cannot reconstruct OOD as well as ID. We propose a novel method, READ (Reconstruction Error Aggregated Detector), to unify inconsistencies from classifier and autoencoder. Specifically, the reconstruction error of raw pixels is transformed to latent space of classifier. We show that the transformed reconstruction error bridges the semantic gap and inherits detection performance from the original. Moreover, we propose an adjustment strategy to alleviate the overconfidence problem of autoencoder according to a fine-grained characterization of OOD data. Under two scenarios of pre-training and retraining, we respectively present two variants of our method, namely READ-MD (Mahalanobis Distance) only based on pre-trained classifier and READ-ED (Euclidean Distance) which retrains the classifier. Our methods do not require access to test time OOD data for fine-tuning hyperparameters. Finally, we demonstrate the effectiveness of the proposed methods through extensive comparisons with state-of-the-art OOD detection algorithms. On a CIFAR-10 pre-trained WideResNet, our method reduces the average FPR@95TPR by up to 9.8% compared with previous state-of-the-art.

We propose a robust and accurate non-parametric method for single-view 3D face reconstruction (SVFR). While tremendous efforts have been devoted to parametric SVFR, a visible gap still lies between the result 3D shape and the ground truth. We believe there are two major obstacles: 1) the representation of the parametric model is limited to a certain face database; 2) 2D images and 3D shapes in the fitted datasets are distinctly misaligned. To resolve these issues, a large-scale pseudo 2D&3D dataset is created by first rendering the detailed 3D faces, then swapping the face in the wild images with the rendered face. These pseudo 2D&3D pairs are created from publicly available datasets which eliminate the gaps between 2D and 3D data while covering diverse appearances, poses, scenes, and illumination. We further propose a non-parametric scheme to learn a well-generalized SVFR model from the created dataset, and the proposed hierarchical signed distance function turns out to be effective in predicting middle-scale and small-scale 3D facial geometry. Our model outperforms previous methods on FaceScape-wild/lab and MICC benchmarks and is well generalized to various appearances, poses, expressions, and in-the-wild environments. The code is released at https://github.com/zhuhao-nju/rafare.

As a direct depth sensor, radar holds promise as a tool to improve monocular 3D object detection, which suffers from depth errors, due in part to the depth-scale ambiguity. On the other hand, leveraging radar depths is hampered by difficulties in precisely associating radar returns with 3D estimates from monocular methods, effectively erasing its benefits. This paper proposes a fusion network that addresses this radar-camera association challenge. We train our network to predict the 3D offsets between radar returns and object centers, enabling radar depths to enhance the accuracy of 3D monocular detection. By using parallel radar and camera backbones, our network fuses information at both the feature level and detection level, while at the same time leveraging a state-of-the-art monocular detection technique without retraining it. Experimental results show significant improvement in mean average precision and translation error on the nuScenes dataset over monocular counterparts. Our source code is available at https://github.com/longyunf/radiant.

Knowledge base question answering (KBQA) has attracted a lot of interest in recent years, especially for complex questions which require multiple facts to answer. Question decomposition is a promising way to answer complex questions. Existing decomposition methods split the question into sub-questions according to a single compositionality type, which is not sufficient for questions involving multiple compositionality types. In this paper, we propose Question Decomposition Tree (QDT) to represent the structure of complex questions. Inspired by recent advances in natural language generation (NLG), we present a two-staged method called Clue-Decipher to generate QDT. It can leverage the strong ability of NLG model and simultaneously preserve the original questions. To verify that QDT can enhance KBQA task, we design a decomposition-based KBQA system called QDTQA. Extensive experiments show that QDTQA outperforms previous state-of-the-art methods on ComplexWebQuestions dataset. Besides, our decomposition method improves an existing KBQA system by 12% and sets a new state-of-the-art on LC-QuAD 1.0.

There is a longstanding interest in capturing the error behaviour of object detectors by finding images where their performance is likely to be unsatisfactory. In real-world applications such as autonomous driving, it is also crucial to characterise potential failures beyond simple requirements of detection performance. For example, a missed detection of a pedestrian close to an ego vehicle will generally require closer inspection than a missed detection of a car in the distance. The problem of predicting such potential failures at test time has largely been overlooked in the literature and conventional approaches based on detection uncertainty fall short in that they are agnostic to such fine-grained characterisation of errors. In this work, we propose to reformulate the problem of finding "hard" images as a query-based hard image retrieval task, where queries are specific definitions of "hardness", and offer a simple and intuitive method that can solve this task for a large family of queries. Our method is entirely post-hoc, does not require ground-truth annotations, is independent of the choice of a detector, and relies on an efficient Monte Carlo estimation that uses a simple stochastic model in place of the ground-truth. We show experimentally that it can be applied successfully to a wide variety of queries for which it can reliably identify hard images for a given detector without any labelled data. We provide results on ranking and classification tasks using the widely used RetinaNet, Faster-RCNN, Mask-RCNN, and Cascade Mask-RCNN object detectors. The code for this project is available at https://github.com/fiveai/hardest.

Maximum Inner Product Search (MIPS) plays an essential role in many applications ranging from information retrieval, recommender systems to natural language processing. However, exhaustive MIPS is often expensive and impractical when there are a large number of candidate items. The state-of-the-art quantization method of approximated MIPS is product quantization with a score-aware loss, developed by assuming that queries are uniformly distributed in the unit sphere. However, in real-world datasets, the above assumption about queries does not necessarily hold. To this end, we propose a quantization method based on the distribution of queries combined with sampled softmax. Further, we introduce a general framework encompassing the proposed method and multiple quantization methods, and we develop an effective optimization for the proposed general framework. The proposed method is evaluated on three real-world datasets. The experimental results show that it outperforms the state-of-the-art baselines.