AAAI 2021 论文列表

The problem of unsupervised domain adaptation in semantic segmentation is a major challenge for numerous computer vision tasks because acquiring pixel-level labels is time-consuming with expensive human labor. A large gap exists among data distributions in different domains, which will cause severe performance loss when a model trained with synthetic data is generalized to real data. Hence, we propose a novel domain adaptation approach, called Content Invariant Representation Network, to narrow the domain gap between the source (S) and target (T) domains. The previous works developed a network to directly transfer the knowledge from the S to T. On the contrary, the proposed method aims to progressively reduce the gap between S and T on the basis of a Content Invariant Representation (CIR). CIR is an intermediate domain (I) sharing invariant content with S and having similar data distribution to T. Then, an Ancillary Classifier Module (ACM) is designed to focus on pixel-level details and generate attention-aware results. ACM adaptively assigns different weights to pixels according to their domain offsets, thereby reducing local domain gaps. The global domain gap between CIR and T is also narrowed by enforcing local alignments. Last, we perform self-supervised training in the pseudo-labeled target domain to further fit the distribution of the real data. Comprehensive experiments on two domain adaptation tasks, that is, GTAV → Cityscapes and SYNTHIA → Cityscapes, clearly demonstrate the superiority of our method compared with state-of-the-art methods.

Federated learning (FL) is a promising approach for training decentralized data located on local client devices while improving efficiency and privacy. However, the distribution and quantity of the training data on the clients' side may lead to significant challenges such as class imbalance and non-IID (non-independent and identically distributed) data, which could greatly impact the performance of the common model. While much effort has been devoted to helping FL models converge when encountering non-IID data, the imbalance issue has not been sufficiently addressed. In particular, as FL training is executed by exchanging gradients in an encrypted form, the training data is not completely observable to either clients or server, and previous methods for class imbalance do not perform well for FL. Therefore, it is crucial to design new methods for detecting class imbalance in FL and mitigating its impact. In this work, we propose a monitoring scheme that can infer the composition of training data for each FL round, and design a new loss function -- Ratio Loss to mitigate the impact of the imbalance. Our experiments demonstrate the importance of acknowledging class imbalance and taking measures as early as possible in FL training, and the effectiveness of our method in mitigating the impact. Our method is shown to significantly outperform previous methods, while maintaining client privacy.

Deep reinforcement learning (DRL) algorithms have been demonstrated to be effective on a wide range of challenging decision making and control tasks. However, these methods typically suffer from severe action oscillations in particular in discrete action setting, which means that agents select different actions within consecutive steps even though states only slightly differ. This issue is often neglected since we usually evaluate the quality of a policy using cumulative rewards only. Action oscillation strongly affects the user experience and even causes serious potential security menace especially in real-world domains with the main concern of safety, such as autonomous driving. In this paper, we introduce Policy Inertia Controller (PIC) which serves as a generic plug-in framework to off-the-shelf DRL algorithms, to enable adaptive balance between the optimality and smoothness in a formal way. We propose Nested Policy Iteration as a general training algorithm for PIC-augmented policy which ensures monotonically non-decreasing updates.Further, we derive a practical DRL algorithm, namely Nested Soft Actor-Critic. Experiments on a collection of autonomous driving tasks and several Atari games suggest that our approach demonstrates substantial oscillation reduction than a range of commonly adopted baselines with almost no performance degradation.

Verifiable training has shown success in creating neural networks that are provably robust to a given amount of noise. However, despite only enforcing a single robustness criterion, its performance scales poorly with dataset complexity. On CIFAR10, a non-robust LeNet model has a 21.63% error rate, while a model created using verifiable training and a L-infinity robustness criterion of 8/255, has an error rate of 57.10%. Upon examination, we find that when labeling visually similar classes, the model's error rate is as high as 61.65%. Thus, we attribute the loss in performance to inter-class similarity. Classes that are similar (i.e., close in the feature space) increase the difficulty of learning a robust model. While it may be desirable to train a model to be robust for a large robustness region, pairwise class similarities limit the potential gains. Furthermore, consideration must be made regarding the relative cost of mistaking one class for another. In security or safety critical tasks, similar classes are likely to belong to the same group, and thus are equally sensitive. In this work, we propose a new approach that utilizes inter-class similarity to improve the performance of verifiable training and create robust models with respect to multiple adversarial criteria. First, we cluster similar classes using agglomerate clustering and assign robustness criteria based on the degree of similarity between clusters. Next, we propose two methods to apply our approach: (1) the Inter-Group Robustness Prioritization method, which uses a custom loss term to create a single model with multiple robustness guarantees and (2) the neural decision tree method, which trains multiple sub-classifiers with different robustness guarantees and combines them in a decision tree architecture. Our experiments on Fashion-MNIST and CIFAR10 demonstrate that by prioritizing the robustness between the most dissimilar groups, we improve clean performance by up to 9.63% and 30.89% respectively. Furthermore, on CIFAR100, our approach reduces the clean error rate by 26.32%.

Machine teaching is an algorithmic framework for teaching a target hypothesis via a sequence of examples or demonstrations. We investigate machine teaching for temporal logic formulas—a novel and expressive hypothesis class amenable to time-related task specifications. In the context of teaching temporal logic formulas, an exhaustive search even for a myopic solution takes exponential time (with respect to the time span of the task). We propose an efficient approach for teaching parametric linear temporal logic formulas. Concretely, we derive a necessary condition for the minimal time length of a demonstration to eliminate a set of hypotheses. Utilizing this condition, we propose an efficient myopic teaching algorithm by solving a sequence of integer programming problems. We further show that, under two notions of teaching complexity, the proposed algorithm has near-optimal performance. We evaluate our algorithm extensively under different classes of learners (i.e., learners with different preferences over hypotheses) and interaction protocols (e.g., non-adaptive and adaptive). Our results demonstrate the effectiveness of the proposed algorithm in teaching temporal logic formulas; in particular, we show that there are significant gains of teaching efficacy when the teacher adapts to feedback of the learner, or adapts to a (non-myopic) oracle.

Natural language generation (NLG) is an important task with various applications like neural machine translation (NMT) and image captioning. Since deep-learning-based methods have issues of exposure bias and loss inconsistency, reinforcement learning (RL) is widely adopted in NLG tasks recently. But most RL-based methods ignore the deviation ignorance issue, which means the model fails to understand the extent of token-level deviation well. It leads to semantic incorrectness and hampers the agent to perform well. To address the issue, we propose a technique called adaptive prior-dependent correction (APDC) to enhance RL. It leverages the distribution generated by computing the distances between the ground truth and all other words to correct the agent's stochastic policy. Additionally, some techniques on RL are explored to coordinate RL with APDC, which requires a reward estimation at every time step. We find that the RL-based NLG tasks are a special case in RL, where the state transition is deterministic and the afterstate value equals the Q-value at every time step. To utilize such prior knowledge, we estimate the advantage function with the difference of the Q-values which can be estimated by Monte Carlo rollouts. Experiments show that, on three tasks of NLG (NMT, image captioning, abstractive text summarization), our method consistently outperforms the state-of-the-art RL-based approaches on different frequently-used metrics.

Pedestrians with challenging patterns, e.g. small scale or heavy occlusion, appear frequently in practical applications like autonomous driving, which remains tremendous obstacle to higher robustness of detectors. Although plenty of previous works have been dedicated to these problems, properly matching patterns of pedestrian and parameters of detector, i.e., constructing a detector with proper parameter sizes for certain pedestrian patterns of different complexity, has been seldom investigated intensively. Pedestrian instances are usually handled equally with the same amount of parameters, which in our opinion is inadequate for those with more difficult patterns and leads to unsatisfactory performance. Thus, we propose in this paper a novel detection approach via adaptive pattern-parameter matching. The input pedestrian patterns, especially the complex ones, are first disentangled into simpler patterns for detection head by Pattern Disentangling Module (PDM) with various receptive fields. Then, Gating Feature Filtering Module (GFFM) dynamically decides the spatial positions where the patterns are still not simple enough and need further disentanglement by the next-level PDM. Cooperating with these two key components, our approach can adaptively select the best matched parameter size for the input patterns according to their complexity. Moreover, to further explore the relationship between parameter sizes and their performance on the corresponding patterns, two parameter selection policies are designed: 1) extending parameter size to maximum, aiming at more difficult patterns for different occlusion types; 2) specializing parameter size by group division, aiming at complex patterns for scale variations. Extensive experiments on two popular benchmarks, Caltech and CityPersons, show that our proposed method achieves superior performance compared with other state-of-the-art methods on subsets of different scales and occlusion types.

In many real-world applications, the amount of data available for training is often limited, and thus inductive bias and auxiliary knowledge are much needed for regularizing model training. One popular regularization method is to impose prior distribution assumptions on model parameters, and many recent works also attempt to regularize training by integrating external knowledge into specific neurons. However, existing regularization methods did not take account of the interaction between connected neuron pairs, which is invaluable internal knowledge for adaptive regularization for better representation learning as training progresses. In this paper, we explicitly take into account the interaction between connected neurons, and propose an adaptive internal knowledge driven regularization method, CORR-Reg. The key idea of CORR-Reg is to give a higher significance weight to connections of more correlated neuron pairs. The significance weights adaptively identify more important input neurons for each neuron. Instead of regularizing connection model parameters with a static strength such as weight decay, CORR-Reg imposes weaker regularization strength on more significant connections. As a consequence, neurons attend to more informative input features and thus learn more diversified and discriminative representation. We derive CORR-Reg with Bayesian inference framework and propose a novel optimization algorithm with Lagrange multiplier method and Stochastic Gradient Descent. Extensive evaluations on diverse benchmark datasets and neural network structures show that CORR-Reg achieves significant improvement over state-of-the-art regularization methods.

We provide new adaptive first-order methods for constrained convex optimization. Our main algorithms AdaACSA and AdaAGD+ are accelerated methods, which are universal in the sense that they achieve nearly-optimal convergence rates for both smooth and non-smooth functions, even when they only have access to stochastic gradients. In addition, they do not require any prior knowledge on how the objective function is parametrized, since they automatically adjust their per-coordinate learning rate. These can be seen as truly accelerated Adagrad methods for constrained optimization. We complement them with a simpler algorithm AdaGrad+ which enjoys the same features, and achieves the standard non-accelerated convergence rate. We also present a set of new results involving adaptive methods for unconstrained optimization and variational inequalities arising from monotone operators.

Keyphrase Generation compresses a document into some highly-summative phrases, which is an important task in natural language processing. Most state-of-the-art adopt greedy search or beam search decoding methods. These two decoding methods generate a large number of duplicated keyphrases and are time-consuming. Moreover, beam search only predicts a fixed number of keyphrases for different documents. In this paper, we propose an adaptive generation model-AdaGM, which is mainly inspired by the importance of the first words in keyphrase generation. In AdaGM, a novel reset state training mechanism is proposed to maximize the difference in the predicted first words. To ensure the discreteness and get an appropriate number of keyphrases according to the content of the document adaptively, we equip beam search with a highly effective filter mechanism. Experiments on five public datasets demonstrate the proposed model can generate marginally less duplicated and more accurate keyphrases. The codes of AdaGM are available at: https://github.com/huangxiaolist/adaGM.

We study the multi-armed bandit (MAB) problem with composite and anonymous feedback. In this model, the reward of pulling an arm spreads over a period of time (we call this period as reward interval) and the player receives partial rewards of the action, convoluted with rewards from pulling other arms, successively. Existing results on this model require prior knowledge about the reward interval size as an input to their algorithms. In this paper, we propose adaptive algorithms for both the stochastic and the adversarial cases, without requiring any prior information about the reward interval. For the stochastic case, we prove that our algorithm guarantees a regret that matches the lower bounds (in order). For the adversarial case, we propose the first algorithm to jointly handle non-oblivious adversary and unknown reward interval size. We also conduct simulations based on real-world dataset. The results show that our algorithms outperform existing benchmarks.

Panoptic segmentation that unifies instance segmentation and semantic segmentation has recently attracted increasing attention. While most existing methods focus on designing novel architectures, we steer toward a different perspective: performing automated multi-loss adaptation (named Ada-Segment) on the fly to flexibly adjust multiple training losses over the course of training using a controller trained to capture the learning dynamics. This offers a few advantages: it bypasses manual tuning of the sensitive loss combination, a decisive factor for panoptic segmentation; allows to explicitly model the learning dynamics, and reconcile the learning of multiple objectives (up to ten in our experiments); with an end-to-end architecture, it generalizes to different datasets without the need of re-tuning hyperparameters or re-adjusting the training process laboriously. Our Ada-Segment brings 2.7% panoptic quality (PQ) improvement on COCO val split from the vanilla baseline, achieving the state-of-the-art 48.5% PQ on COCO test-dev split and 32.9% PQ on ADE20K dataset. The extensive ablation studies reveal the ever-changing dynamics throughout the training process, necessitating the incorporation of an automated and adaptive learning strategy as presented in this paper.

With the rapid emergence of video data, image-to-video retrieval has attracted much attention. There are two types of image-to-video retrieval: instance-based and activity-based. The former task aims to retrieve videos containing the same main objects as the query image, while the latter focuses on finding the similar activity. Since dynamic information plays a significant role in the video, we pay attention to the latter task to explore the motion relation between images and videos. In this paper, we propose a Motion-assisted Activity Proposal-based Image-to-Video Retrieval (MAP-IVR) approach to disentangle the video features into motion features and appearance features and obtain appearance features from the images. Then, we perform image-to-video translation to improve the disentanglement quality. The retrieval is performed in both appearance and video feature spaces. Extensive experiments demonstrate that our MAP-IVR approach remarkably outperforms the state-of-the-art approaches on two benchmark activity-based video datasets.

We study active feature selection, a novel feature selection setting in which unlabeled data is available, but the budget for labels is limited, and the examples to label can be actively selected by the algorithm. We focus on feature selection using the classical mutual information criterion, which selects the k features with the largest mutual information with the label. In the active feature selection setting, the goal is to use significantly fewer labels than the data set size and still find k features whose mutual information with the label based on the entire data set is large. We explain and experimentally study the choices that we make in the algorithm, and show that they lead to a successful algorithm, compared to other more naive approaches. Our design draws on insights which relate the problem of active feature selection to the study of pure-exploration multi-armed bandits settings. While we focus here on mutual information, our general methodology can be adapted to other feature-quality measures as well. The extended version of this paper, reporting all experiment results, is available at Schnapp and Sabato (2020). The code is available at the following url: https://github.com/ShacharSchnapp/ActiveFeatureSelection

Recent advances in machine learning have led to increased deployment of black-box classifiers across a wide variety of applications. In many such situations there is a critical need to both reliably assess the performance of these pre-trained models and to perform this assessment in a label-efficient manner (given that labels may be scarce and costly to collect). In this paper, we introduce an active Bayesian approach for assessment of classifier performance to satisfy the desiderata of both reliability and label-efficiency. We begin by developing inference strategies to quantify uncertainty for common assessment metrics such as accuracy, misclassification cost, and calibration error. We then propose a general framework for active Bayesian assessment using inferred uncertainty to guide efficient selection of instances for labeling, enabling better performance assessment with fewer labels. We demonstrate significant gains from our proposed active Bayesian approach via a series of systematic empirical experiments assessing the performance of modern neural classifiers (e.g., ResNet and BERT) on several standard image and text classification datasets.

As mobile devices are becoming ubiquitous, regularly interacting with a variety of user interfaces (UIs) is a common aspect of daily life for many people. To improve the accessibility of these devices and to enable their usage in a variety of settings, building models that can assist users and accomplish tasks through the UI is vitally important. However, there are several challenges to achieve this. First, UI components of similar appearance can have different functionalities, making understanding their function more important than just analyzing their appearance. Second, domain-specific features like Document Object Model (DOM) in web pages and View Hierarchy (VH) in mobile applications provide important signals about the semantics of UI elements, but these features are not in a natural language format. Third, owing to a large diversity in UIs and absence of standard DOM or VH representations, building a UI understanding model with high coverage requires large amounts of training data. Inspired by the success of pre-training based approaches in NLP for tackling a variety of problems in a data-efficient way, we introduce a new pre-trained UI representation model called ActionBert. Our methodology is designed to leverage visual, linguistic and domain-specific features in user interaction traces to pre-train generic feature representations of UIs and their components. Our key intuition is that user actions, e.g., a sequence of clicks on different UI components, reveals important information about their functionality. We evaluate the proposed model on a wide variety of downstream tasks, ranging from icon classification to UI component retrieval based on its natural language description. Experiments show that the proposed ActionBert model outperforms multi-modal baselines across all downstream tasks by up to 15.5%.

Double Q-learning is a popular reinforcement learning algorithm in Markov decision process (MDP) problems. Clipped Double Q-learning, as an effective variant of Double Q-learning, employs the clipped double estimator to approximate the maximum expected action value. Due to the underestimation bias of the clipped double estimator, performance of clipped Double Q-learning may be degraded in some stochastic environments. In this paper, in order to reduce the underestimation bias, we propose an action candidate based clipped double estimator for Double Q-learning. Specifically, we first select a set of elite action candidates with the high action values from one set of estimators. Then, among these candidates, we choose the highest valued action from the other set of estimators. Finally, we use the maximum value in the second set of estimators to clip the action value of the chosen action in the first set of estimators and the clipped value is used for approximating the maximum expected action value. Theoretically, the underestimation bias in our clipped Double Q-learning decays monotonically as the number of the action candidates decreases. Moreover, the number of action candidates controls the trade-off between the overestimation and underestimation biases. In addition, we also extend our clipped Double Q-learning to continuous action tasks via approximating the elite continuous action candidates. We empirically verify that our algorithm can more accurately estimate the maximum expected action value on some toy environments and yield good performance on several benchmark problems.

We study the problem of fairly allocating indivisible goods and focus on the classic fairness notion of proportionality. The indivisibility of the goods is long known to pose highly non-trivial obstacles to achieving fairness, and a very vibrant line of research has aimed to circumvent them using appropriate notions of approximate fairness. Recent work has established that even approximate versions of proportionality (PROPx) may be impossible to achieve even for small instances, while the best known achievable approximations (PROP1) are much weaker. We introduce the notion of proportionality up to the maximin item (PROPm) and show how to reach an allocation satisfying this notion for any instance involving up to five agents with additive valuations. PROPm provides a well-motivated middle-ground between PROP1 and PROPx, while also capturing some elements of the well-studied maximin share (MMS) benchmark: another relaxation of proportionality that has attracted a lot of attention.

When allocating indivisible resources or tasks, an envy-free allocation or equitable allocation may not exist. We present a sufficient condition and an algorithm to achieve envy-freeness and equitability when monetary transfers are allowed. The approach works for any agent valuation functions (positive or negative) as long as they satisfy superadditivity. For the case of additive utilities, we present a characterization of allocations that can simultaneously be made equitable and envy-free via payments. Our study shows that superadditive valuations constitute the largest class of valuations for which an envy-free and equitable outcome exists for all instances. We then present a distributed algorithm to compute an approximately envy-free outcome for any class of valuations.