AAAI 2021 Papers

In this paper, we propose a Meta Cooperative Learning (MCL) framework for task-oriented dialog systems (TDSs). Our model consists of an auxiliary KB reasoning task for learning meta KB knowledge, an auxiliary dialogue reasoning task for learning dialogue patterns, and a TDS task (primary task) that aims at not only retrieving accurate entities from KB but also generating natural responses, which are coordinated to achieve collective success in both retrieving accurate KB entities and generating human-like responses via meta learning. Concretely, the dialog generation model amalgamates complementary meta KB and dialog knowledge from two novel auxiliary reasoning tasks that together provide integrated guidance to build a high-quality TDS by adding regularization terms to force primary network to produce similar results to auxiliary networks. While MCL automatically learns appropriate labels for the two auxiliary reasoning tasks from the primary task, without requiring access to any further data. The key idea behind MCL is to use the performance of the primary task, which is trained alongside the auxiliary tasks in one iteration, to improve the auxiliary labels for the next iteration with meta learning. Experimental results on three benchmark datasets show that MCL can generate higher quality responses compared to several strong baselines in terms of both automatic and human evaluations. Code to reproduce the results in this paper is available at: https://github.com/siat-nlp/MCL.

In conventional supervised learning, a training dataset is given with ground-truth labels from a known label set, and the learned model will classify unseen instances to known labels. In real situations, when the learned models do not work well, users generally attribute the failure to the inadequate selection of learning algorithms or the lack of enough labeled training samples. In this paper, we point out that there is an important category of failure, which owes to the fact that there are unknown classes in the training data misperceived as other labels, and thus their existence was unknown from the given supervision. Such problems of unknown unknown classes can hardly be addressed by common re-selection of algorithms or accumulation of training samples. For this purpose, we propose the exploratory machine learning, where in this paradigm once user encounters unsatisfactory learning performance, she can examine the possibility and, if unknown unknowns really exist, deploy the optimal strategy of feature space augmentation to make the unknown classes observable and be enabled for learning. Theoretical analysis and empirical study on both synthetic and real datasets validate the efficacy of our proposal.

Exploration-exploitation is a powerful and practical tool in multi-agent learning (MAL), however, its effects are far from understood. To make progress in this direction, we study a smooth analogue of Q-learning. We start by showing that our learning model has strong theoretical justification as an optimal model for studying exploration-exploitation. Specifically, we prove that smooth Q-learning has bounded regret in arbitrary games for a cost model that explicitly captures the balance between game and exploration costs and that it always converges to the set of quantal-response equilibria (QRE), the standard solution concept for games under bounded rationality, in weighted potential games with heterogeneous learning agents. In our main task, we then turn to measure the effect of exploration in collective system performance. We characterize the geometry of the QRE surface in low-dimensional MAL systems and link our findings with catastrophe (bifurcation) theory. In particular, as the exploration hyperparameter evolves over-time, the system undergoes phase transitions where the number and stability of equilibria can change radically given an infinitesimal change to the exploration parameter. Based on this, we provide a formal theoretical treatment of how tuning the exploration parameter can provably lead to equilibrium selection with both positive as well as negative (and potentially unbounded) effects to system performance.

This paper studies the challenging problem of reinforcement learning (RL) in hard exploration tasks with sparse rewards. It focuses on the exploration stage before the agent gets the first positive reward, in which case, traditional RL algorithms with simple exploration strategies often work poorly. Unlike previous methods using some attribute of a single state as the intrinsic reward to encourage exploration, this work leverages the social influence between different states to permit more efficient exploration. It introduces a general intrinsic reward construction method to evaluate the social influence of states dynamically. Three kinds of social influence are introduced for a state: conformity, power, and authority. By measuring the state’s social influence, agents quickly find the focus state during the exploration process. The proposed RL framework with state social influence evaluation works well in hard exploration task. Extensive experimental analyses and comparisons in Grid Maze and many hard exploration Atari 2600 games demonstrate its high exploration efficiency.

Exploration is essential for reinforcement learning (RL). To face the challenges of exploration, we consider a reward-free RL framework that completely separates exploration from exploitation and brings new challenges for exploration algorithms. In the exploration phase, the agent learns an exploratory policy by interacting with a reward-free environment and collects a dataset of transitions by executing the policy. In the planning phase, the agent computes a good policy for any reward function based on the dataset without further interacting with the environment. This framework is suitable for the meta RL setting where there are many reward functions of interest. In the exploration phase, we propose to maximize the Renyi entropy over the state-action space and justify this objective theoretically. The success of using Renyi entropy as the objective results from its encouragement to explore the hard-to-reach state-actions. We further deduce a policy gradient formulation for this objective and design a practical exploration algorithm that can deal with complex environments. In the planning phase, we solve for good policies given arbitrary reward functions using a batch RL algorithm. Empirically, we show that our exploration algorithm is effective and sample efficient, and results in superior policies for arbitrary reward functions in the planning phase.

In semi-supervised learning, one key strategy in exploiting unlabeled data is trying to estimate its pseudo-label based on current predictive model, where the unlabeled data assigned with pseudo-label is further utilized to enlarge labeled data set for model update. Nonetheless, the supervision information conveyed by pseudo-label is prone to error especially when the performance of initial predictive model is mediocre due to limited amount of labeled data. In this paper, an intermediate unlabeled data exploitation strategy is investigated via partial label assignment, i.e. a set of candidate labels other than a single pseudo-label are assigned to the unlabeled data. We only assume that the ground-truth label of unlabeled data resides in the assigned candidate label set, which is less error-prone than trying to identify the single ground-truth label via pseudo-labeling. Specifically, a multi-class classifier is induced from the partial label examples with candidate labels to facilitate model induction with labeled examples. An iterative procedure is designed to enable labeling information communication between the classifiers induced from partial label examples and labeled examples, whose classification outputs are integrated to yield the final prediction. Comparative studies against state-of-the-art approaches clearly show the effectiveness of the proposed unlabeled data exploitation strategy for multi-class semi-supervised learning.

Many unsupervised domain adaptive (UDA) person ReID approaches combine clustering-based pseudo-label prediction with feature fine-tuning. However, because of domain gap, the pseudo-labels are not always reliable and there are noisy/incorrect labels. This would mislead the feature representation learning and deteriorate the performance. In this paper, we propose to estimate and exploit the credibility of the assigned pseudo-label of each sample to alleviate the influence of noisy labels, by suppressing the contribution of noisy samples. We build our baseline framework using the mean teacher method together with an additional contrastive loss. We have observed that a sample with a wrong pseudo-label through clustering in general has a weaker consistency between the output of the mean teacher model and the student model. Based on this finding, we propose to exploit the uncertainty (measured by consistency levels) to evaluate the reliability of the pseudo-label of a sample and incorporate the uncertainty to re-weight its contribution within various ReID losses, including the ID classification loss per sample, the triplet loss, and the contrastive loss. Our uncertainty-guided optimization brings significant improvement and achieves the state-of-the-art performance on benchmark datasets.

The significant progress on Generative Adversarial Networks (GANs) has facilitated realistic single-object image generation based on language input. However, complex-scene generation (with various interactions among multiple objects) still suffers from messy layouts and object distortions, due to diverse configurations in layouts and appearances. Prior methods are mostly object-driven and ignore their inter-relations that play a significant role in complex-scene images. This work explores relationship-aware complex-scene image generation, where multiple objects are inter-related as a scene graph. With the help of relationships, we propose three major updates in the generation framework. First, reasonable spatial layouts are inferred by jointly considering the semantics and relationships among objects. Compared to standard location regression, we show relative scales and distances serve a more reliable target. Second, since the relations between objects have significantly influenced an object's appearance, we design a relation-guided generator to generate objects reflecting their relationships. Third, a novel scene graph discriminator is proposed to guarantee the consistency between the generated image and the input scene graph. Our method tends to synthesize plausible layouts and objects, respecting the interplay of multiple objects in an image. Experimental results on Visual Genome and HICO-DET datasets show that our proposed method significantly outperforms prior arts in terms of IS and FID metrics. Based on our user study and visual inspection, our method is more effective in generating logical layout and appearance for complex-scenes.

In this paper, we propose to estimate 3D hand pose by recovering the 3D coordinates of joints in a group-wise manner, where less-related joints are automatically categorized into different groups and exhibit different features. This is different from the previous methods where all the joints are considered holistically and share the same feature. The benefits of our method are illustrated by the principle of multi-task learning (MTL), i.e., by separating less-related joints into different groups (as different tasks), our method learns different features for each of them, therefore efficiently avoids the negative transfer (among less related tasks/groups of joints). The key of our method is a novel binary selector that automatically selects related joints into the same group. We implement such a selector with binary values stochastically sampled from a Concretedistribution, which is constructed using Gumbel softmax on trainable parameters. This enables us to preserve the differentiable property of the whole network. We further exploit features from those less-related groups by carrying out an additional feature fusing scheme among them, to learn more discriminative features. This is realized by implementing multiple 1x1 convolutions on the concatenated features, where each joint group contains a unique 1x1convolution for feature fusion. The detailed ablation analysis and the extensive experiments on several benchmark datasets demonstrate the promising performance of the proposed method over the state-of-the-art (SOTA) methods. Besides, our method achieves top-1 among all the methods that do not exploit the dense 3D shape labels on the most recently released FreiHAND competition at the submission date. The source code and models are available at https://github.com/moranli-aca/LearnableGroups-Hand.

Adapting semantic segmentation models to new domains is an important but challenging problem. Recently enlightening progress has been made, but the performance of existing methods is unsatisfactory on real datasets where the new target domain comprises of heterogeneous sub-domains (e.g. diverse weather characteristics). We point out that carefully reasoning about the multiple modalities in the target domain can improve the robustness of adaptation models. To this end, we propose a condition-guided adaptation framework that is empowered by a special attentive progressive adversarial training (APAT) mechanism and a novel self-training policy. The APAT strategy progressively performs condition-specific alignment and attentive global feature matching. The new self-training scheme exploits the adversarial ambivalences of easy and hard adaptation regions and the correlations among target sub-domains effectively. We evaluate our method (DCAA) on various adaptation scenarios where the target images vary in weather conditions. The comparisons against baselines and the state-of-the-art approaches demonstrate the superiority of DCAA over the competitors.

User modeling is critical for developing personalized services in industry. A common way for user modeling is to learn user representations that can be distinguished by their interests or preferences. In this work, we focus on developing universal user representation model. The obtained universal representations are expected to contain rich information, and be applicable to various downstream applications without further modifications (e.g., user preference prediction and user profiling). Accordingly, we can be free from the heavy work of training task-specific models for every downstream task as in previous works. In specific, we propose Self-supervised User Modeling Network (SUMN) to encode behavior data into the universal representation. It includes two key components. The first one is a new learning objective, which guides the model to fully identify and preserve valuable user information under a self-supervised learning framework. The other one is a multi-hop aggregation layer, which benefits the model capacity in aggregating diverse behaviors. Extensive experiments on benchmark datasets show that our approach can outperform state-of-the-art unsupervised representation methods, and even compete with supervised ones.

Human perceives rich auditory experience with distinct sound heard by ears. Videos recorded with binaural audio particular simulate how human receives ambient sound. However, a large number of videos are with monaural audio only, which would degrade the user experience due to the lack of ambient information. To address this issue, we propose an audio spatialization framework to convert a monaural video into a binaural one exploiting the relationship across audio and visual components. By preserving the left-right consistency in both audio and visual modalities, our learning strategy can be viewed as a self-supervised learning technique, and alleviates the dependency on a large amount of video data with ground truth binaural audio data during training. Experiments on benchmark datasets confirm the effectiveness of our proposed framework in both semi-supervised and fully supervised scenarios, with ablation studies and visualization further support the use of our model for audio spatialization.

Self-attention networks have shown remarkable progress in computer vision tasks such as image classification. The main benefit of the self-attention mechanism is the ability to capture long-range feature interactions in attention-maps. However, the computation of attention-maps requires a learnable key, query, and positional encoding, whose usage is often not intuitive and computationally expensive. To mitigate this problem, we propose a novel self-attention module with explicitly modeled attention-maps using only a single learnable parameter for low computational overhead. The design of explicitly modeled attention-maps using geometric prior is based on the observation that the spatial context for a given pixel within an image is mostly dominated by its neighbors, while more distant pixels have a minor contribution. Concretely, the attention-maps are parametrized via simple functions (e.g., Gaussian kernel) with a learnable radius, which is modeled independently of the input content. Our evaluation shows that our method achieves an accuracy improvement of up to 2.2% over the ResNet-baselines in ImageNet ILSVRC and outperforms other self-attention methods such as AA-ResNet152 in accuracy by 0.9% with 6.4% fewer parameters and 6.7% fewer GFLOPs. This result empirically indicates the value of incorporating geometric prior into self-attention mechanism when applied in image classification.

Unlabeled data exploitation and interpretability are usually both required in reality. They, however, are conducted independently, and very few works try to connect the two. For unlabeled data exploitation, state-of-the-art semi-supervised learning (SSL) results have been achieved via encouraging the consistency of model output on data perturbation, that is, consistency assumption. However, it remains hard for users to understand how particular decisions are made by state-of-the-art SSL models. To this end, in this paper we first disclose that the consistency assumption is closely related to causality invariance, where causality invariance lies in the main reason why the consistency assumption is valid. We then propose ECT (Explanation Consistency Training) which encourages a consistent reason of model decision under data perturbation. ECT employs model explanation as a surrogate of the causality of model output, which is able to bridge state-of-the-art interpretability to SSL models and alleviate the high complexity of causality. We realize ECT-SM for vision and ECT-ATT for NLP tasks. Experimental results on real-world data sets validate the highly competitive performance and better explanation of the proposed algorithms.

Explaining the predictions of neural black-box models is an important problem, especially when such models are used in applications where user trust is crucial. Estimating the influence of training examples on a learned neural model's behavior allows us to identify training examples most responsible for a given prediction and, therefore, to faithfully explain the output of a black-box model. The most generally applicable existing method is based on influence functions, which scale poorly for larger sample sizes and models. We propose gradient rollback, a general approach for influence estimation, applicable to neural models where each parameter update step during gradient descent touches a smaller number of parameters, even if the overall number of parameters is large. Neural matrix factorization models trained with gradient descent are part of this model class. These models are popular and have found a wide range of applications in industry. Especially knowledge graph embedding methods, which belong to this class, are used extensively. We show that gradient rollback is highly efficient at both training and test time. Moreover, we show theoretically that the difference between gradient rollback's influence approximation and the true influence on a model's behavior is smaller than known bounds on the stability of stochastic gradient descent. This establishes that gradient rollback is robustly estimating example influence. We also conduct experiments which show that gradient rollback provides faithful explanations for knowledge base completion and recommender datasets. An implementation and an appendix are available.

As an emerging field in Machine Learning, Explainable AI (XAI) has been offering remarkable performance in interpreting the decisions made by Convolutional Neural Networks (CNNs). To achieve visual explanations for CNNs, methods based on class activation mapping and randomized input sampling have gained great popularity. However, the attribution methods based on these techniques provide lower-resolution and blurry explanation maps that limit their explanation power. To circumvent this issue, visualization based on various layers is sought. In this work, we collect visualization maps from multiple layers of the model based on an attribution-based input sampling technique and aggregate them to reach a fine-grained and complete explanation. We also propose a layer selection strategy that applies to the whole family of CNN-based models, based on which our extraction framework is applied to visualize the last layers of each convolutional block of the model. Moreover, we perform an empirical analysis of the efficacy of derived lower-level information to enhance the represented attributions. Comprehensive experiments conducted on shallow and deep models trained on natural and industrial datasets, using both ground-truth and model-truth based evaluation metrics validate our proposed algorithm by meeting or outperforming the state-of-the-art methods in terms of explanation ability and visual quality, demonstrating that our method shows stability regardless of the size of objects or instances to be explained.

Interpretable machine learning has gained much attention recently. Briefness and comprehensiveness are necessary in order to provide a large amount of information concisely when explaining a black-box decision system. However, existing interpretable machine learning methods fail to consider briefness and comprehensiveness simultaneously, leading to redundant explanations. We propose the variational information bottleneck for interpretation, VIBI, a system-agnostic interpretable method that provides a brief but comprehensive explanation. VIBI adopts an information theoretic principle, information bottleneck principle, as a criterion for finding such explanations. For each instance, VIBI selects key features that are maximally compressed about an input (briefness), and informative about a decision made by a black-box system on that input (comprehensive). We evaluate VIBI on three datasets and compare with state-of-the-art interpretable machine learning methods in terms of both interpretability and fidelity evaluated by human and quantitative metrics.

Given the widespread deployment of black box deep neural networks in computer vision applications, the interpretability aspect of these black box systems has recently gained traction. Various methods have been proposed to explain the results of such deep neural networks. However, some recent works have shown that such explanation methods are biased and do not produce consistent interpretations. Hence, rather than introducing a novel explanation method, we learn models that are encouraged to be interpretable given an explanation method. We use Grad-CAM as the explanation algorithm and encourage the network to learn consistent interpretations along with maximizing the log-likelihood of the correct class. We show that our method outperforms the baseline on the pointing game evaluation on ImageNet and MS-COCO datasets respectively. We also introduce new evaluation metrics that penalize the saliency map if it lies outside the ground truth bounding box or segmentation mask, and show that our method outperforms the baseline on these metrics as well. Moreover, our model trained with interpretation consistency generalizes to other explanation algorithms on all the evaluation metrics. The code and models are publicly available.

A desirable goal for autonomous agents is to be able to coordinate on the fly with previously unknown teammates. Known as “ad hoc teamwork”, enabling such a capability has been receiving increasing attention in the research community. One of the central challenges in ad hoc teamwork is quickly recognizing the current plans of other agents and planning accordingly. In this paper, we focus on the scenario in which teammates can communicate with one another, but only at a cost. Thus, they must carefully balance plan recognition based on observations vs. that based on communication. This paper proposes a new metric for evaluating how similar are two policies that a teammate may be following - the Expected Divergence Point (EDP). We then present a novel planning algorithm for ad hoc teamwork, determining which query to ask and planning accordingly. We demonstrate the effectiveness of this algorithm in a range of increasingly general communication in ad hoc teamwork problems.