AAAI 2021 论文列表

In this paper, we devise a novel semantic completion network, called point cloud semantic scene completion network (PCSSC-Net), for indoor scenes solely based on point clouds. Existing point cloud completion networks still suffer from their inability of fully recovering complex structures and contents from global geometric descriptions neglecting semantic hints. To extract and infer comprehensive information from partial input, we design a patch-based contextual encoder to hierarchically learn point-level, patch-level, and scene-level geometric and contextual semantic information with a divide-and-conquer strategy. Consider that the scene semantics afford a high-level clue of constituting geometry for an indoor scene environment, we articulate a semantics-guided completion decoder where semantics could help cluster isolated points in the latent space and infer complicated scene geometry. Given the fact that real-world scans tend to be incomplete as ground truth, we choose to synthesize scene dataset with RGB-D images and annotate complete point clouds as ground truth for the supervised training purpose. Extensive experiments validate that our new method achieves the state-of-the-art performance, in contrast with the current methods applied to our dataset.

We expand the literature on the price of anarchy (PoA) of simultaneous item auctions by considering settings with correlated values; we do this via the fundamental economic model of interdependent values (IDV). It is well-known that in multi-item settings with private values, correlated values can lead to bad PoA, which can be polynomially large in the number of agents~n. In the more general model of IDV, we show that the PoA can be polynomially large even in single-item settings. On the positive side, we identify a natural condition on information dispersion in the market, which enables good PoA guarantees. Under this condition, we show that for single-item settings, the PoA of standard mechanisms degrades gracefully. For settings with multiple items we show a separation between two domains: If there are more buyers, we devise a new simultaneous item auction with good PoA, under limited information asymmetry. To the best of our knowledge, this is the first positive PoA result for correlated values in multi-item settings. The main technical difficulty in establishing this result is that the standard tool for establishing PoA results --- the smoothness framework --- is unsuitable for IDV settings, and so we must introduce new techniques to address the unique challenges imposed by such settings. In the domain of more items, we establish impossibility results even for surprisingly simple scenarios.

Human emotion decoding in affective brain-computer interfaces suffers a major setback due to the inter-subject variability of electroencephalography (EEG) signals. Existing approaches usually require amassing extensive EEG data of each new subject, which is prohibitively time-consuming along with poor user experience. To tackle this issue, we divide EEG representations into private components specific to each subject and shared emotional components that are universal to all subjects. According to this representation partition, we propose a plug-and-play domain adaptation method for dealing with the inter-subject variability. In the training phase, subject-invariant emotional representations and private components of source subjects are separately captured by a shared encoder and private encoders. Furthermore, we build one emotion classifier on the shared partition and subjects' individual classifiers on the combination of these two partitions. In the calibration phase, the model only requires few unlabeled EEG data from incoming target subjects to model their private components. Therefore, besides the shared emotion classifier, we have another pipeline to use the knowledge of source subjects through the similarity of private components. In the test phase, we integrate predictions of the shared emotion classifier with those of individual classifiers ensemble after modulation by similarity weights. Experimental results on the SEED dataset show that our model greatly shortens the calibration time within a minute while maintaining the recognition accuracy, all of which make emotion decoding more generalizable and practicable.

Real-world planning problems often involve hundreds or even thousands of objects, straining the limits of modern planners. In this work, we address this challenge by learning to predict a small set of objects that, taken together, would be sufficient for finding a plan. We propose a graph neural network architecture for predicting object importance in a single inference pass, thus incurring little overhead while greatly reducing the number of objects that must be considered by the planner. Our approach treats the planner and transition model as black boxes, and can be used with any off-the-shelf planner. Empirically, across classical planning, probabilistic planning, and robotic task and motion planning, we find that our method results in planning that is significantly faster than several baselines, including other partial grounding strategies and lifted planners. We conclude that learning to predict a sufficient set of objects for a planning problem is a simple, powerful, and general mechanism for planning in large instances. Video: https://youtu.be/FWsVJc2fvCE Code: https://git.io/JIsqX

Width-based planning methods have been shown to yield state-of-the-art performance in the Atari 2600 domain using pixel input. One successful approach, RolloutIW, represents states with the B-PROST boolean feature set. An augmented version of RolloutIW, pi-IW, shows that learned features can be competitive with handcrafted ones for width-based search. In this paper, we leverage variational autoencoders (VAEs) to learn features directly from pixels in a principled manner, and without supervision. The inference model of the trained VAEs extracts boolean features from pixels, and RolloutIW plans with these features. The resulting combination outperforms the original RolloutIW and human professional play on Atari 2600 and drastically reduces the size of the feature set.

Automatic Feature Engineering (AFE) aims to extract useful knowledge for interpretable predictions given data for the machine learning tasks. Here, we develop AFE to extract dependency relationships that can be interpreted with functional formulas to discover physics meaning or new hypotheses for the problems of interest. We focus on materials science applications, where interpretable predictive modeling may provide principled understanding of materials systems and guide new materials discovery. It is often computationally prohibitive to exhaust all the potential relationships to construct and search the whole feature space to identify interpretable and predictive features. We develop and evaluate new AFE strategies by exploring a feature generation tree (FGT) with deep Q-network (DQN) for scalable and efficient exploration policies. The developed DQN-based AFE strategies are benchmarked with the existing AFE methods on several materials science datasets.

Traffic state estimation (TSE) reconstructs the traffic variables (e.g., density or average velocity) on road segments using partially observed data, which is important for traffic managements. Traditional TSE approaches mainly bifurcate into two categories: model-driven and data-driven, and each of them has shortcomings. To mitigate these limitations, hybrid TSE methods, which combine both model-driven and data-driven, are becoming a promising solution. This paper introduces a hybrid framework, physics-informed deep learning (PIDL), to combine second-order traffic flow models and neural networks to solve the TSE problem. PIDL can encode traffic flow models into deep neural networks to regularize the learning process to achieve improved data efficiency and estimation accuracy. We focus on highway TSE with observed data from loop detectors and probe vehicles, using both density and average velocity as the traffic variables. With numerical examples, we show the use of PIDL to solve a popular second-order traffic flow model, i.e., a Greenshields-based Aw-Rascle-Zhang (ARZ) model, and discover the model parameters. We then evaluate the PIDL-based TSE method using the Next Generation SIMulation (NGSIM) dataset. Experimental results demonstrate the proposed PIDL-based approach to outperform advanced baseline methods in terms of data efficiency and estimation accuracy.

Consider a learning algorithm, which involves an internal call to an optimization routine such as a generalized eigenvalue problem, a cone programming problem or even sorting. Integrating such a method as layers within a trainable deep network in a numerically stable way is not simple – for instance, only recently, strategies have emerged for eigendecomposition and differentiable sorting. We propose an efficient and differentiable solver for general linear programming problems which can be used in a plug and play manner within deep neural networks as a layer. Our development is inspired by a fascinating but not widely used link between dynamics of slime mold (physarum) and mathematical optimization schemes such as steepest descent. We describe our development and demonstrate the use of our solver in a video object segmentation task and meta-learning for few-shot learning. We review the relevant known results and provide a technical analysis describing its applicability for our use cases. Our solver performs comparably with a customized projected gradient descent method on the first task and outperforms the very recently proposed differentiable CVXPY solver on the second task. Experiments show that our solver converges quickly without the need for a feasible initial point. Interestingly, our scheme is easy to implement and can easily serve as layers whenever a learning procedure needs a fast approximate solution to a LP, within a larger network.

We focus on the scenario in which an agent can exploit his information advantage to manipulate the outcome of an election. In particular, we study district-based elections with two candidates, in which the winner of the election is the candidate that wins in the majority of the districts. District-based elections are adopted worldwide (e.g., UK and USA) and are a natural extension of widely studied voting mechanisms (e.g., k-voting and plurality voting). We resort to the Bayesian persuasion framework, where the manipulator (sender) strategically discloses information to the voters (receivers) that update their beliefs rationally. We study both private signaling in which the sender can use a private communication channel per receiver and public signaling in which the sender can use a single communication channel for all the receivers. Furthermore, for the first time, we introduce semi-public signaling in which the sender can use a single communication channel per district. We show that there is a sharp distinction between private and (semi-)public signaling. In particular, optimal private signaling schemes can provide an arbitrarily better probability of victory than (semi-)public ones and can be computed efficiently, while optimal (semi-)public signaling schemes cannot be approximated to within any factor in polynomial time unless P=NP. However, we show that reasonable relaxations allow the design of multi-criteria PTASs for optimal (semi-)public signaling schemes. In doing so, we introduce a novel property, namely comparative stability, and we design a bi-criteria PTAS for public signaling in general Bayesian persuasion problems beyond elections when the sender's utility function is state-dependent.

Non-IID data present a tough challenge for federated learning. In this paper, we explore a novel idea of facilitating pairwise collaborations between clients with similar data. We propose FedAMP, a new method employing federated attentive message passing to facilitate similar clients to collaborate more. We establish the convergence of FedAMP for both convex and non-convex models, and propose a heuristic method to further improve the performance of FedAMP when clients adopt deep neural networks as personalized models. Our extensive experiments on benchmark data sets demonstrate the superior performance of the proposed methods.

A common challenge in personalized user preference prediction is the cold-start problem. Due to the lack of user-item interactions, directly learning from the new users' log data causes serious over-fitting problem. Recently, many existing studies regard the cold-start personalized preference prediction as a few-shot learning problem, where each user is the task and recommended items are the classes, and the gradient-based meta learning method (MAML) is leveraged to address this challenge. However, in real-world application, the users are not uniformly distributed (i.e., different users may have different browsing history, recommended items, and user profiles. We define the major users as the users in the groups with large numbers of users sharing similar user information, and other users are the minor users), existing MAML approaches tend to fit the major users and ignore the minor users. To address the task-overfitting problem, we propose a novel personalized adaptive meta learning approach to consider both the major and the minor users with three key contributions: 1) We are the first to present a personalized adaptive learning rate meta-learning approach to improve the performance of MAML by focusing on both the major and minor users. 2) To provide better personalized learning rates for each user, we introduce a similarity-based method to find similar users as a reference and a tree-based method to store users' features for fast search. 3) To reduce the memory usage, we design a memory agnostic regularizer to further reduce the space complexity to constant while maintain the performance. Experiments on MovieLens, BookCrossing, and real-world production datasets reveal that our method outperforms the state-of-the-art methods dramatically for both the minor and major users.

Vaccination is the primary intervention for controlling the spread of infectious diseases. A certain level of vaccination rate (referred to as "herd immunity'') is needed for this intervention to be effective. However, there are concerns that herd immunity might not be achieved due to an increasing level of hesitancy and opposition to vaccines. One of the primary reasons for this is the cost of non-conformance with one's peers. We use the framework of network coordination games to study the persistence of anti-vaccine sentiment in a population. We extend it to incorporate the opposing forces of the pressure of conforming to peers, herd-immunity and vaccination benefits. We study the structure of the equilibria in such games, and the characteristics of unvaccinated nodes. We also study Stackelberg strategies to reduce the number of nodes with anti-vaccine sentiment. Finally, we evaluate our results on different kinds of real world social networks.

Automatic evaluation for open-ended natural language generation tasks remains a challenge. We propose a learned evaluation metric: Perception Score. It utilizes a pre-trained model and considers context information for conditional generation. Perception Score assigns a holistic score along with the uncertainty measurement. We conduct experiments on three open-ended conditional generation tasks and two open-ended unconditional generation tasks. Perception Score achieves state-of-the-art results on all the tasks consistently in terms of correlation with human evaluation scores.

When deploying machine learning models in the real-world, system designers may wish that models exhibit certain shape behavior, i.e., model outputs follow a particular shape with respect to input features. Trends such as monotonicity, convexity, diminishing or accelerating returns are some of the desired shapes. Presence of these shapes makes the model more interpretable for the system designers, and adequately fair for the customers. We notice that many such common shapes are related to derivatives, and propose a new approach, PenDer (Penalizing Derivatives), which incorporates these shape constraints by penalizing the derivatives. We further present an Augmented Lagrangian Method (ALM) to solve this constrained optimization problem. Experiments on three real-world datasets illustrate that even though both PenDer and state-of-the-art Lattice models achieve similar conformance to shape, PenDer captures better sensitivity of prediction with respect to intended features. We also demonstrate that PenDer achieves better test performance than Lattice while enforcing more desirable shape behavior.

Traditional knowledge distillation uses a two-stage training strategy to transfer knowledge from a high-capacity teacher model to a compact student model, which relies heavily on the pre-trained teacher. Recent online knowledge distillation alleviates this limitation by collaborative learning, mutual learning and online ensembling, following a one-stage end-to-end training fashion. However, collaborative learning and mutual learning fail to construct an online high-capacity teacher, whilst online ensembling ignores the collaboration among branches and its logit summation impedes the further optimisation of the ensemble teacher. In this work, we propose a novel Peer Collaborative Learning method for online knowledge distillation, which integrates online ensembling and network collaboration into a unified framework. Specifically, given a target network, we construct a multi-branch network for training, in which each branch is called a peer. We perform random augmentation multiple times on the inputs to peers and assemble feature representations outputted from peers with an additional classifier as the peer ensemble teacher. This helps to transfer knowledge from a high-capacity teacher to peers, and in turn further optimises the ensemble teacher. Meanwhile, we employ the temporal mean model of each peer as the peer mean teacher to collaboratively transfer knowledge among peers, which helps each peer to learn richer knowledge and facilitates to optimise a more stable model with better generalisation. Extensive experiments on CIFAR-10, CIFAR-100 and ImageNet show that the proposed method significantly improves the generalisation of various backbone networks and outperforms the state-of-the-art methods.

In computer vision, monocular depth estimation is the problem of obtaining a high-quality depth map from a two-dimensional image. This map provides information on three-dimensional scene geometry, which is necessary for various applications in academia and industry, such as robotics and autonomous driving. Recent studies based on convolutional neural networks achieved impressive results for this task. However, most previous studies did not consider the relationships between the neighboring pixels in a local area of the scene. To overcome the drawbacks of existing methods, we propose a patch-wise attention method for focusing on each local area. After extracting patches from an input feature map, our module generates attention maps for each local patch, using two attention modules for each patch along the channel and spatial dimensions. Subsequently, the attention maps return to their initial positions and merge into one attention feature. Our method is straightforward but effective. The experimental results on two challenging datasets, KITTI and NYU Depth V2, demonstrate that the proposed method achieves significant performance. Furthermore, our method outperforms other state-of-the-art methods on the KITTI depth estimation benchmark.

Current state-of-the-art image captioning systems usually generated descriptions autoregressively, i.e., every forward step conditions on the given image and previously produced words. The sequential attribution causes a unavoidable decoding latency. Non-autoregressive image captioning, on the other hand, predicts the entire sentence simultaneously and accelerates the inference process significantly. However, it removes the dependence in a caption and commonly suffers from repetition or missing issues. To make a better trade-off between speed and quality, we introduce a partially non-autoregressive model, named PNAIC, which considers a caption as a series of concatenated word groups. The groups are generated parallelly in global while each word in group is predicted from left to right, and thus the captioner can create multiple discontinuous words concurrently at each time step. More importantly, by incorporating curriculum learning-based training tasks of group length prediction and invalid group deletion, our model is capable of generating accurate captions as well as preventing common incoherent errors. Extensive experiments on MS COCO benchmark demonstrate that our proposed method achieves more than 3.5× speedup while maintaining competitive performance.

In this paper, we technically propose an enriched prior guided framework, called Dual-constrained Deep Semi-Supervised Coupled Factorization Network (DS2CF-Net), for discovering hierarchical coupled data representation. To extract hidden deep features, DS2CF-Net is formulated as a partial-label and geometrical structure-constrained framework. Specifically, DS2CF-Net designs a deep factorization architecture using multilayers of linear transformations, which can coupled update both the basis vectors and new representations in each layer. To enable learned deep representations and coefficients to be discriminative, we also consider enriching the supervised prior by joint deep coefficients-based label prediction and then incorporate the enriched prior information as additional label and structure constraints. The label constraint can enable the intra-class samples to have same coordinate in feature space, and the structure constraint forces the coefficients in each layer to be block-diagonal so that the enriched prior using the self-expressive label propagation are more accurate. Our network also integrates the adaptive dual-graph learning to retain the local structures of both data and feature manifolds in each layer. Extensive experiments on image datasets demonstrate the effectiveness of DS2CF-Net for representation learning and clustering.

The goal of few-shot learning is to learn a classifier that can recognize unseen classes from limited support data with labels. A common practice for this task is to train a model on the base set first and then transfer to novel classes through fine-tuning or meta-learning. However, as the base classes have no overlap to the novel set, simply transferring whole knowledge from base data is not an optimal solution since some knowledge in the base model may be biased or even harmful to the novel class. In this paper, we propose to transfer partial knowledge by freezing or fine-tuning particular layer(s) in the base model. Specifically, layers will be imposed different learning rates if they are chosen to be fine-tuned, to control the extent of preserved transferability. To determine which layers to be recast and what values of learning rates for them, we introduce an evolutionary search based method that is efficient to simultaneously locate the target layers and determine their individual learning rates. We conduct extensive experiments on CUB and mini-ImageNet to demonstrate the effectiveness of our proposed method. It achieves the state-of-the-art performance on both meta-learning and non-meta based frameworks. Furthermore, we extend our method to the conventional pre-training + fine-tuning paradigm and obtain consistent improvement.