AAAI 2023 Papers

3D object detection received increasing attention in autonomous driving recently. Objects in 3D scenes are distributed with diverse orientations. Ordinary detectors do not explicitly model the variations of rotation and reflection transformations. Consequently, large networks and extensive data augmentation are required for robust detection. Recent equivariant networks explicitly model the transformation variations by applying shared networks on multiple transformed point clouds, showing great potential in object geometry modeling. However, it is difficult to apply such networks to 3D object detection in autonomous driving due to its large computation cost and slow reasoning speed. In this work, we present TED, an efficient Transformation-Equivariant 3D Detector to overcome the computation cost and speed issues. TED first applies a sparse convolution backbone to extract multi-channel transformation-equivariant voxel features; and then aligns and aggregates these equivariant features into lightweight and compact representations for high-performance 3D object detection. On the highly competitive KITTI 3D car detection leaderboard, TED ranked 1st among all submissions with competitive efficiency. Code is available at https://github.com/hailanyi/TED.

Recent work has demonstrated that pretrained transformers are overconfident in text classification tasks, which can be calibrated by the famous post-hoc calibration method temperature scaling (TS). Character or word spelling mistakes are frequently encountered in real applications and greatly threaten transformer model safety. Research on calibration under noisy settings is rare, and we focus on this direction. Based on a toy experiment, we discover that TS performs poorly when the datasets are perturbed by slight noise, such as swapping the characters, which results in distribution shift. We further utilize two metrics, predictive uncertainty and maximum mean discrepancy (MMD), to measure the distribution shift between clean and noisy datasets, based on which we propose a simple yet effective transferable TS method for calibrating models dynamically. To evaluate the performance of the proposed methods under noisy settings, we construct a benchmark consisting of four noise types and five shift intensities based on the QNLI, AG-News, and Emotion tasks. Experimental results on the noisy benchmark show that (1) the metrics are effective in measuring distribution shift and (2) transferable TS can significantly decrease the expected calibration error (ECE) compared with the competitive baseline ensemble TS by approximately 46.09%.

Deep operator network (DeepONet) has demonstrated great success in various learning tasks, including learning solution operators of partial differential equations. In particular, it provides an efficient approach to predicting the evolution equations in a finite time horizon. Nevertheless, the vanilla DeepONet suffers from the issue of stability degradation in the long- time prediction. This paper proposes a transfer-learning aided DeepONet to enhance the stability. Our idea is to use transfer learning to sequentially update the DeepONets as the surro- gates for propagators learned in different time frames. The evolving DeepONets can better track the varying complexities of the evolution equations, while only need to be updated by efficient training of a tiny fraction of the operator networks. Through systematic experiments, we show that the proposed method not only improves the long-time accuracy of Deep- ONet while maintaining similar computational cost but also substantially reduces the sample size of the training set.

Video copy localization aims to precisely localize all the copied segments within a pair of untrimmed videos in video retrieval applications. Previous methods typically start from frame-to-frame similarity matrix generated by cosine similarity between frame-level features of the input video pair, and then detect and refine the boundaries of copied segments on similarity matrix under temporal constraints. In this paper, we propose TransVCL: an attention-enhanced video copy localization network, which is optimized directly from initial frame-level features and trained end-to-end with three main components: a customized Transformer for feature enhancement, a correlation and softmax layer for similarity matrix generation, and a temporal alignment module for copied segments localization. In contrast to previous methods demanding the handcrafted similarity matrix, TransVCL incorporates long-range temporal information between feature sequence pair using self- and cross- attention layers. With the joint design and optimization of three components, the similarity matrix can be learned to present more discriminative copied patterns, leading to significant improvements over previous methods on segment-level labeled datasets (VCSL and VCDB). Besides the state-of-the-art performance in fully supervised setting, the attention architecture facilitates TransVCL to further exploit unlabeled or simply video-level labeled data. Additional experiments of supplementing video-level labeled datasets including SVD and FIVR reveal the high flexibility of TransVCL from full supervision to semi-supervision (with or without video-level annotation). Code is publicly available at https://github.com/transvcl/TransVCL.

Heuristic search algorithms, e.g. A*, are the commonly used tools for pathfinding on grids, i.e. graphs of regular structure that are widely employed to represent environments in robotics, video games, etc. Instance-independent heuristics for grid graphs, e.g. Manhattan distance, do not take the obstacles into account, and thus the search led by such heuristics performs poorly in obstacle-rich environments. To this end, we suggest learning the instance-dependent heuristic proxies that are supposed to notably increase the efficiency of the search. The first heuristic proxy we suggest to learn is the correction factor, i.e. the ratio between the instance-independent cost-to-go estimate and the perfect one (computed offline at the training phase). Unlike learning the absolute values of the cost-to-go heuristic function, which was known before, learning the correction factor utilizes the knowledge of the instance-independent heuristic. The second heuristic proxy is the path probability, which indicates how likely the grid cell is lying on the shortest path. This heuristic can be employed in the Focal Search framework as the secondary heuristic, allowing us to preserve the guarantees on the bounded sub-optimality of the solution. We learn both suggested heuristics in a supervised fashion with the state-of-the-art neural networks containing attention blocks (transformers). We conduct a thorough empirical evaluation on a comprehensive dataset of planning tasks, showing that the suggested techniques i) reduce the computational effort of the A* up to a factor of 4x while producing the solutions, whose costs exceed those of the optimal solutions by less than 0.3% on average; ii) outperform the competitors, which include the conventional techniques from the heuristic search, i.e. weighted A*, as well as the state-of-the-art learnable planners. The project web-page is: https://airi-institute.github.io/TransPath/.

Recently, transformer architecture has gained great success in the computer vision community, such as image classification, object detection, etc. Nonetheless, its application for 3D vision remains to be explored, given that point cloud is inherently sparse, irregular, and unordered. Furthermore, existing point transformer frameworks usually feed raw point cloud of N×3 dimension into transformers, which limits the point processing scale because of their quadratic computational costs to the input size N. In this paper, we rethink the structure of point transformer. Instead of directly applying transformer to points, our network (TransLO) can process tens of thousands of points simultaneously by projecting points onto a 2D surface and then feeding them into a local transformer with linear complexity. Specifically, it is mainly composed of two components: Window-based Masked transformer with Self Attention (WMSA) to capture long-range dependencies; Masked Cross-Frame Attention (MCFA) to associate two frames and predict pose estimation. To deal with the sparsity issue of point cloud, we propose a binary mask to remove invalid and dynamic points. To our knowledge, this is the first transformer-based LiDAR odometry network. The experiment results on the KITTI odometry dataset show that our average rotation and translation RMSE achieves 0.500°/100m and 0.993% respectively. The performance of our network surpasses all recent learning-based methods and even outperforms LOAM on most evaluation sequences.Codes will be released on https://github.com/IRMVLab/TransLO.

Nearest neighbor-based methods are commonly used for classification tasks and as subroutines of other data-analysis methods. An attacker with the capability of inserting their own data points into the training set can manipulate the inferred nearest neighbor structure. We distill this goal to the task of performing a training-set data insertion attack against k-Nearest Neighbor classification (kNN). We prove that computing an optimal training-time (a.k.a. poisoning) attack against kNN classification is NP-Hard, even when k = 1 and the attacker can insert only a single data point. We provide an anytime algorithm to perform such an attack, and a greedy algorithm for general k and attacker budget. We provide theoretical bounds and empirically demonstrate the effectiveness and practicality of our methods on synthetic and real-world datasets. Empirically, we find that kNN is vulnerable in practice and that dimensionality reduction is an effective defense. We conclude with a discussion of open problems illuminated by our analysis.

The problem of adversarial attacks to a black-box model when no queries are allowed has posed a great challenge to the community and has been extensively investigated. In this setting, one simple yet effective method is to transfer the obtained adversarial examples from attacking surrogate models to fool the target model. Previous works have studied what kind of attacks to the surrogate model can generate more transferable adversarial examples, but their performances are still limited due to the mismatches between surrogate models and the target model. In this paper, we tackle this problem from a novel angle---instead of using the original surrogate models, can we obtain a Meta-Surrogate Model (MSM) such that attacks to this model can be easily transferred to other models? We show that this goal can be mathematically formulated as a bi-level optimization problem and design a differentiable attacker to make training feasible. Given one or a set of surrogate models, our method can thus obtain an MSM such that adversarial examples generated on MSM enjoy eximious transferability. Comprehensive experiments on Cifar-10 and ImageNet demonstrate that by attacking the MSM, we can obtain stronger transferable adversarial examples to deceive black-box models including adversarially trained ones, with much higher success rates than existing methods.

Traffic prediction is an important component of the intelligent transportation system. Existing deep learning methods encode temporal information and spatial information separately or iteratively. However, the spatial and temporal information is highly correlated in a traffic network, so existing methods may not learn the complex spatial-temporal dependencies hidden in the traffic network due to the decomposed model design. To overcome this limitation, we propose a new model named Trafformer, which unifies spatial and temporal information in one transformer-style model. Trafformer enables every node at every timestamp interact with every other node in every other timestamp in just one step in the spatial-temporal correlation matrix. This design enables Trafformer to catch complex spatial-temporal dependencies. Following the same design principle, we use the generative style decoder to predict multiple timestamps in only one forward operation instead of the iterative style decoder in Transformer. Furthermore, to reduce the complexity brought about by the huge spatial-temporal self-attention matrix, we also propose two variants of Trafformer to further improve the training and inference speed without losing much effectivity. Extensive experiments on two traffic datasets demonstrate that Trafformer outperforms existing methods and provides a promising future direction for the spatial-temporal traffic prediction problem.

In this work, we tackle the challenging task of jointly tracking hand object poses and reconstructing their shapes from depth point cloud sequences in the wild, given the initial poses at frame 0. We for the first time propose a point cloud-based hand joint tracking network, HandTrackNet, to estimate the inter-frame hand joint motion. Our HandTrackNet proposes a novel hand pose canonicalization module to ease the tracking task, yielding accurate and robust hand joint tracking. Our pipeline then reconstructs the full hand via converting the predicted hand joints into a MANO hand. For object tracking, we devise a simple yet effective module that estimates the object SDF from the first frame and performs optimization-based tracking. Finally, a joint optimization step is adopted to perform joint hand and object reasoning, which alleviates the occlusion-induced ambiguity and further refines the hand pose. During training, the whole pipeline only sees purely synthetic data, which are synthesized with sufficient variations and by depth simulation for the ease of generalization. The whole pipeline is pertinent to the generalization gaps and thus directly transferable to real in-the-wild data. We evaluate our method on two real hand object interaction datasets, e.g. HO3D and DexYCB, without any fine-tuning. Our experiments demonstrate that the proposed method significantly outperforms the previous state-of-the-art depth-based hand and object pose estimation and tracking methods, running at a frame rate of 9 FPS. We have released our code on https://github.com/PKU-EPIC/HOTrack.

Text recognition is a long-standing research problem for document digitalization. Existing approaches are usually built based on CNN for image understanding and RNN for char-level text generation. In addition, another language model is usually needed to improve the overall accuracy as a post-processing step. In this paper, we propose an end-to-end text recognition approach with pre-trained image Transformer and text Transformer models, namely TrOCR, which leverages the Transformer architecture for both image understanding and wordpiece-level text generation. The TrOCR model is simple but effective, and can be pre-trained with large-scale synthetic data and fine-tuned with human-labeled datasets. Experiments show that the TrOCR model outperforms the current state-of-the-art models on the printed, handwritten and scene text recognition tasks. The TrOCR models and code are publicly available at https://aka.ms/trocr.

With rapid development in hardware (sensors and processors) and AI algorithms, automated driving techniques have entered the public’s daily life and achieved great success in supporting human driving performance. However, due to the high contextual variations and temporal dynamics in pedestrian behaviors, the interaction between autonomous-driving cars and pedestrians remains challenging, impeding the development of fully autonomous driving systems. This paper focuses on predicting pedestrian intention with a novel transformer-based evidential prediction (TrEP) algorithm. We develop a transformer module towards the temporal correlations among the input features within pedestrian video sequences and a deep evidential learning model to capture the AI uncertainty under scene complexities. Experimental results on three popular pedestrian intent benchmarks have verified the effectiveness of our proposed model over the state-of-the-art. The algorithm performance can be further boosted by controlling the uncertainty level. We systematically compare human disagreements with AI uncertainty to further evaluate AI performance in confusing scenes. The code is released at https://github.com/zzmonlyyou/TrEP.git.

Understanding mathematical questions effectively is a crucial task, which can benefit many applications, such as difficulty estimation. Researchers have drawn much attention to designing pre-training models for question representations due to the scarcity of human annotations (e.g., labeling difficulty). However, unlike general free-format texts (e.g., user comments), mathematical questions are generally designed with explicit purposes and mathematical logic, and usually consist of more complex content, such as formulas, and related mathematical knowledge (e.g., Function). Therefore, the problem of holistically representing mathematical questions remains underexplored. To this end, in this paper, we propose a novel contrastive pre-training approach for mathematical question representations, namely QuesCo, which attempts to bring questions with more similar purposes closer. Specifically, we first design two-level question augmentations, including content-level and structure-level, which generate literally diverse question pairs with similar purposes. Then, to fully exploit hierarchical information of knowledge concepts, we propose a knowledge hierarchy-aware rank strategy (KHAR), which ranks the similarities between questions in a fine-grained manner. Next, we adopt a ranking contrastive learning task to optimize our model based on the augmented and ranked questions. We conduct extensive experiments on two real-world mathematical datasets. The experimental results demonstrate the effectiveness of our model.

Translating imagined speech from human brain activity into voice is a challenging and absorbing research issue that can provide new means of human communication via brain signals. Efforts to reconstruct speech from brain activity have shown their potential using invasive measures of spoken speech data, but have faced challenges in reconstructing imagined speech. In this paper, we propose NeuroTalk, which converts non-invasive brain signals of imagined speech into the user's own voice. Our model was trained with spoken speech EEG which was generalized to adapt to the domain of imagined speech, thus allowing natural correspondence between the imagined speech and the voice as a ground truth. In our framework, an automatic speech recognition decoder contributed to decomposing the phonemes of the generated speech, demonstrating the potential of voice reconstruction from unseen words. Our results imply the potential of speech synthesis from human EEG signals, not only from spoken speech but also from the brain signals of imagined speech.

Deep neural networks (DNNs) are known to be vulnerable to adversarial geometric transformation. This paper aims to verify the robustness of large-scale DNNs against the combination of multiple geometric transformations with a provable guarantee. Given a set of transformations (e.g., rotation, scaling, etc.), we develop GeoRobust, a black-box robustness analyser built upon a novel global optimisation strategy, for locating the worst-case combination of transformations that affect and even alter a network's output. GeoRobust can provide provable guarantees on finding the worst-case combination based on recent advances in Lipschitzian theory. Due to its black-box nature, GeoRobust can be deployed on large-scale DNNs regardless of their architectures, activation functions, and the number of neurons. In practice, GeoRobust can locate the worst-case geometric transformation with high precision for the ResNet50 model on ImageNet in a few seconds on average. We examined 18 ImageNet classifiers, including the ResNet family and vision transformers, and found a positive correlation between the geometric robustness of the networks and the parameter numbers. We also observe that increasing the depth of DNN is more beneficial than increasing its width in terms of improving its geometric robustness. Our tool GeoRobust is available at https://github.com/TrustAI/GeoRobust.

Humans perceive surrounding scenes through multiple senses with multisensory integration. For example, hearing helps capture the spatial location of a racing car behind us; seeing peoples' talking faces can strengthen our perception of their speech. However, today's state-of-the-art scene understanding systems are usually designed to rely on a single audio or visual modality. Ignoring multisensory cooperation has become one of the key bottlenecks in creating intelligent systems with human-level perception capability, which impedes the real-world applications of existing scene understanding models. To address this limitation, my research has pioneered marrying computer vision with computer audition to create multimodal systems that can learn to understand audio and visual data. In particular, my current research focuses on asking and solving fundamental problems in a fresh research area: audio-visual scene understanding and strives to develop unified, explainable, and robust multisensory perception machines. The three themes are distinct yet interconnected, and all of them are essential for designing powerful and trustworthy perception systems. In my talk, I will give a brief overview about this new research area and then introduce my works in the three research thrusts.

Artificial intelligence (AI) and Machine Learning (ML) have shown great success in many areas such as computer vision, natural language processing, and knowledge discovery. However, AI research to deliver social benefits and impacts is less explored while imminent needed. Guided by the United Nations’ Sustainable Development Goals, my research involves the development of advanced AI techniques, in particular Deep Graph Learning (DGL), to address the grand societal challenges and further apply them to various social good applications for improving our society and people’s daily life, namely DGL for Social Good (DGL4SG). Achieving the goal is not easy since challenges come from the increasing complexity of many factors including problems, data, and techniques, which require long-term and concentrated effort. DGL presents a good opportunity to build better solutions and tools due to its strong capability in learning and inferring graph data which is ideal for modeling many real-world social good systems. Fortunately, I have been working on DGL with continued contributions and impacts since my graduate study. The special research experience lifts me up to a unique position for conducting research that intersects AI, DGL, and social good, and pushing the field of DGL4SG forward.

In the near future, autonomous systems such as multi-robot systems are envisioned to increasingly co-exist with hu- mans in our daily lives, from household service to large- scale warehouse logistics, agriculture environment sampling, and smart city. In these applications, robots and humans as networked heterogeneous components will frequently inter- act with each other in a variety of scenarios under uncer- tain, rapidly-changing, and possibly hostile environment. On one hand, harmonious interactions among robots, as well as between robots and humans, would require safe integration (e.g. collision-free close-proximity interactions) of heteroge- neous robots, human, and human-robot autonomy. On the other hand, reliable interactions among autonomous multi- robot systems often call for resilient system integrity (e.g. communication capability with potential robot failures) to re- tain its capability of accomplishing complex tasks through coordinated behaviors. In the proposed talk, I will discuss our recent works towards safe autonomy and resilient autonomy that aim to facilitate correct-by-design robotic behaviors in a variety of applications.

Deep reinforcement learning (DRL) has proven effective in training agents to achieve goals in complex environments. However, a trained RL agent may exhibit, during deployment, unexpected behavior when faced with a situation where its state transitions differ even slightly from the training environment. Such a situation can arise for a variety of reasons. Rapid and accurate detection of anomalous behavior appears to be a prerequisite for using DRL in safety-critical systems, such as autonomous driving. We propose a novel OOD detection algorithm based on modeling the transition function of the training environment. Our method captures the bias of model behavior when encountering subtle changes of dynamics while maintaining a low false positive rate. Preliminary evaluations on the realistic simulator CARLA corroborate the relevance of our proposed method.