AAAI 2021 论文列表

Teaching humans is an important topic under the umbrella of machine teaching, and its core problem is to design an algorithm for selecting teaching examples. Existing work typically regards humans as passive learners, where an ordered set of teaching examples are generated and fed to learners sequentially. However, such a mechanism is inconsistent with the behavior of human learners in practice. A real human learner can actively choose whether to review a historical example or to receive a new example depending on the belief of her learning states. In this work, we propose a model of active learners and design an efficient teaching algorithm accordingly. Experimental results with both simulated learners and real crowdsourcing workers demonstrate that our teaching algorithm has better teaching performance compared to existing methods.

Knowledge distillation is an effective tool to compress large pre-trained convolutional neural networks (CNNs) or their ensembles into models applicable to mobile and embedded devices. However, with expected flops or latency, existing methods are hand-crafted heuristics. They propose to pre-define the target student network for knowledge distillation, which may be sub-optimal because it requires much effort to explore a powerful student from the large design space. In this paper, we develop a novel teacher guided neural architecture search method to directly search for a student network with flexible channel and layer sizes. Specifically, we define the search space as the number of the channels/layers, which is sampled based on the probability distribution and is learned by minimizing the search objective of the student network. The maximum probability for the size in each distribution serves as the final searched width and depth of the target student network. Extensive experiments on a variety of face recognition benchmarks have demonstrated the superiority of our method over the state-of-the-art alternatives.

Automatically constructing taxonomy finds many applications in e-commerce and web search. One critical challenge is as data and business scope grow in real applications, new concepts are emerging and needed to be added to the existing taxonomy. Previous approaches focus on the taxonomy expansion, i.e. finding an appropriate hypernym concept from the taxonomy for a new query concept. In this paper, we formulate a new task, “taxonomy completion”, by discovering both the hypernym and hyponym concepts for a query. We propose Triplet Matching Network (TMN), to find the appropriate pairs for a given query concept. TMN consists of one primal scorer and multiple auxiliary scorers. These auxiliary scorers capture various fine-grained signals (e.g., query to hypernym or query to hyponym semantics), and the primal scorer makes a holistic prediction on triplet based on the internal feature representations of all auxiliary scorers. Also, an innovative channel-wise gating mechanism that retains task-specific information in concept representations is introduced to further boost model performance. Experiments on four real-world large-scale datasets show that TMN achieves the best performance on both taxonomy completion task and the previous taxonomy expansion task, outperforming existing methods.

Generalized Zero-Shot Learning (GZSL) targets recognizing new categories by learning transferable image representations. Existing methods find that, by aligning image representations with corresponding semantic labels, the semantic-aligned representations can be transferred to unseen categories. However, supervised by only seen category labels, the learned semantic knowledge is highly task-specific, which makes image representations biased towards seen categories. In this paper, we propose a novel Dual-Contrastive Embedding Network (DCEN) that simultaneously learns task-specific and task-independent knowledge via semantic alignment and instance discrimination. First, DCEN leverages task labels to cluster representations of the same semantic category by cross-modal contrastive learning and exploring semantic-visual complementarity. Besides task-specific knowledge, DCEN then introduces task-independent knowledge by attracting representations of different views of the same image and repelling representations of different images. Compared to high-level seen category supervision, this instance discrimination supervision encourages DCEN to capture low-level visual knowledge, which is less biased toward seen categories and alleviates the representation bias. Consequently, the task-specific and task-independent knowledge jointly make for transferable representations of DCEN, which obtains averaged 4.1% improvement on four public benchmarks.

In a reward-free environment, what is a suitable intrinsic objective for an agent to pursue so that it can learn an optimal task-agnostic exploration policy? In this paper, we argue that the entropy of the state distribution induced by finite-horizon trajectories is a sensible target. Especially, we present a novel and practical policy-search algorithm, Maximum Entropy POLicy optimization (MEPOL), to learn a policy that maximizes a non-parametric, $k$-nearest neighbors estimate of the state distribution entropy. In contrast to known methods, MEPOL is completely model-free as it requires neither to estimate the state distribution of any policy nor to model transition dynamics. Then, we empirically show that MEPOL allows learning a maximum-entropy exploration policy in high-dimensional, continuous-control domains, and how this policy facilitates learning meaningful reward-based tasks downstream.

Training a model with limited data is an essential task for machine learning and visual recognition. Few-shot learning approaches meta-learn a task-level inductive bias from SEEN class few-shot tasks, and the meta-model is expected to facilitate the few-shot learning with UNSEEN classes. Inspired by the idea that unlabeled data can be utilized to smooth the model space in traditional semi-supervised learning, we propose TAsk COoperation (TACO) which takes advantage of unsupervised tasks to smooth the meta-model space. Specifically, we couple the labeled support set in a few-shot task with easily-collected unlabeled instances, prediction agreement on which encodes the relationship between tasks. The learned smooth meta-model promotes the generalization ability on supervised UNSEEN few-shot tasks. The state-of-the-art few-shot classification results on MiniImageNet and TieredImageNet verify the superiority of TACO to leverage unlabeled data and task relationship in meta-learning.

Zero-shot learning (ZSL) refers to the problem of learning to classify instances from novel classes (unseen) that are absent in the training set (seen). Most ZSL methods infer the correlation between visual features and attributes to train the classifier for unseen classes. They may have a strong bias towards seen classes during training. Meta-learning has been introduced to mitigate the basis, but meta-ZSL methods are inapplicable when tasks used for training are sampled from diverse distributions. In this regard, we propose a novel Task-aligned Generative Meta-learning model for Zero-shot learning (TGMZ), aiming to mitigate the potentially biased training and to enable meta-ZSL to accommodate real-world datasets that contain diverse distributions. Specifically, TGMZ incorporates an attribute-conditioned task-wise distribution alignment network that projects tasks into a unified distribution to deliver an unbiased model. Our experiments show TGMZ achieves a relative improvement of 2.1%, 3.0%, 2.5%, and 7.6% over state-of-the-art algorithms on AWA1, AWA2, CUB, and aPY datasets, respectively. Overall, TGMZ outperforms competitors by 3.6% in the generalized zero-shot learning (GZSL) setting and 7.9% in our proposed fusion-ZSL setting.

Given the ubiquity of negative campaigning in recent political elections, we find it important to study its properties from a computational perspective. To this end, we present a model where elections can be manipulated by convincing voters to demote specific non-favored candidates, and study its properties in the classic setting of scoring rules. When the goal is constructive (making a preferred candidate win), we prove that finding such a demotion strategy is easy for Plurality and Veto, while generally hard for t-approval and Borda. We also provide a t-factor approximation for t-approval for every fixed t, and a 3-factor approximation algorithm for Borda. Interestingly enough - following recent trends in political science that show that the effectiveness of negative campaigning depends on the type of candidate and demographic - when assigning varying prices to different possible demotion operations, we are able to provide inapproximability results. When the goal is destructive (making the leading opponent lose), we show that the problem is easy for a broad class of scoring rules.

Few-Shot Learning (FSL) is essential for visual recognition. Many methods tackle this challenging problem via learning an embedding function from seen classes and transfer it to unseen classes with a few labeled instances. Researchers recently found it beneficial to incorporate task-specific feature adaptation into FSL models, which produces the most representative features for each task. However, these methods ignore the diversity of classes and apply a global transformation to the task. In this paper, we propose Global and Local Feature Adaptor (GLoFA), a unifying framework that tailors the instance representation to specific tasks by global and local feature adaptors. We claim that class-specific local transformation helps to improve the representation ability of feature adaptor. Global masks tend to capture sketchy patterns, while local masks focus on detailed characteristics. A strategy to measure the relationship between instances adaptively based on the characteristics of both tasks and classes endow GLoFA with the ability to handle mix-grained tasks. GLoFA outperforms other methods on a heterogeneous task distribution and achieves competitive results on benchmark datasets.

The drastic increase of data quantity often brings the severe decrease of data quality, such as incorrect label annotations. It poses a great challenge for robustly training Deep Neural Networks (DNNs). Existing learning methods with label noise either employ ad-hoc heuristics or restrict to specific noise assumptions. However, more general situations, such as instance-dependent label noise, have not been fully explored, as scarce studies focus on their label corruption process. By categorizing instances into confusing and unconfusing instances, this paper proposes a simple yet universal probabilistic model, which explicitly relates noisy labels to their instances. The resultant model can be realized by DNNs, where the training procedure is accomplished by employing a novel alternating optimization algorithm. Experiments on datasets with both synthetic and real-world label noise verify the proposed method yields significant improvements on robustness over state-of-the-art counterparts.

We propose a novel high-performance and interpretable canonical deep tabular data learning architecture, TabNet. TabNet uses sequential attention to choose which features to reason from at each decision step, enabling interpretability and more efficient learning as the learning capacity is used for the most salient features. We demonstrate that TabNet outperforms other variants on a wide range of non-performance-saturated tabular datasets and yields interpretable feature attributions plus insights into its global behavior. Finally, we demonstrate self-supervised learning for tabular data, significantly improving performance when unlabeled data is abundant.

This paper presents TaLNet, a model for voice reconstruction with ultrasound tongue and optical lip videos as inputs. TaLNet is based on an encoder-decoder architecture. Separate encoders are dedicated to processing the tongue and lip data streams respectively. The decoder predicts acoustic features conditioned on encoder outputs and speaker codes.To mitigate for having only relatively small amounts of dual articulatory-acoustic data available for training, and since our task here shares with text-to-speech (TTS) the common goal of speech generation, we propose a novel transfer learning strategy to exploit the much larger amounts of acoustic-only data available to train TTS models. For this, a Tacotron 2 TTS model is first trained, and then the parameters of its decoder are transferred to the TaLNet decoder. We have evaluated our approach on an unconstrained multi-speaker voice recovery task. Our results show the effectiveness of both the proposed model and the transfer learning strategy. Speech reconstructed using our proposed method significantly outperformed all baselines (DNN, BLSTM and without transfer learning) in terms of both naturalness and intelligibility. When using an ASR model decoding the recovery speech, the WER of our proposed method is relatively reduced over 30% compared to baselines.

Scenario-based question answering (SQA) has attracted an increasing research interest. Compared with the well-studied machine reading comprehension (MRC), SQA is a more challenging task: a scenario may contain not only a textual passage to read but also structured data like tables, i.e., tabular scenario based question answering (TSQA). AI applications of TSQA such as answering multiple-choice questions in high-school exams require synthesizing data in multiple cells and combining tables with texts and domain knowledge to infer answers. To support the study of this task, we construct GeoTSQA. This dataset contains 1k real questions contextualized by tabular scenarios in the geography domain. To solve the task, we extend state-of-the-art MRC methods with TTGen, a novel table-to-text generator. It generates sentences from variously synthesized tabular data and feeds the downstream MRC method with the most useful sentences. Its sentence ranking model fuses the information in the scenario, question, and domain knowledge. Our approach outperforms a variety of strong baseline methods on GeoTSQA.

Ternary neural networks (TNNs) are potential for network acceleration by reducing the full-precision weights in network to ternary ones, e.g., {-1,0,1}. However, existing TNNs are mostly calculated based on rule-of-thumb quantization methods by simply thresholding operations, which causes a significant accuracy loss. In this paper, we introduce a stem-residual framework which provides new insight into Ternary quantization, termed Residual Quantization (TRQ), to achieve more powerful TNNs. Rather than directly thresholding operations, TRQ recursively performs quantization on full-precision weights for a refined reconstruction by combining the binarized stem and residual parts. With such a unique quantization process, TRQ endows the quantizer with high flexibility and precision. Our TRQ is generic, which can be easily extended to multiple bits through recursively encoded residual for a better recognition accuracy. Extensive experimental results demonstrate that the proposed method yields great recognition accuracy while being accelerated.

Focusing on text-to-image (T2I) generation, we propose Text and Image Mutual-Translation Adversarial Networks (TIME), a lightweight but effective model that jointly learns a T2I generator G and an image captioning discriminator D under the Generative Adversarial Network framework. While previous methods tackle the T2I problem as a uni-directional task and use pre-trained language models to enforce the image--text consistency, TIME requires neither extra modules nor pre-training. We show that the performance of G can be boosted substantially by training it jointly with D as a language model. Specifically, we adopt Transformers to model the cross-modal connections between the image features and word embeddings, and design an annealing conditional hinge loss that dynamically balances the adversarial learning. In our experiments, TIME achieves state-of-the-art (SOTA) performance on the CUB dataset (Inception Score of 4.91 and Fréchet Inception Distance of 14.3 on CUB), and shows promising performance on MS-COCO dataset on image captioning and downstream vision-language tasks.

It is well-known that second-order optimizer can accelerate the training of deep neural networks, however, the huge computation cost of second-order optimization makes it impractical to apply in real practice. In order to reduce the cost, many methods have been proposed to approximate a second-order matrix. Inspired by KFAC, we propose a novel Trace-based Hardware-driven layer-ORiented Natural Gradient Descent Computation method, called THOR, to make the second-order optimization applicable in the real application models. Specifically, we gradually increase the update interval and use the matrix trace to determine which blocks of Fisher Information Matrix (FIM) need to be updated. Moreover, by resorting the power of hardware, we have designed a Hardware-driven approximation method for computing FIM to achieve better performance. To demonstrate the effectiveness of THOR, we have conducted extensive experiments. The results show that training ResNet-50 on ImageNet with THOR only takes 66.7 minutes to achieve a top-1 accuracy of 75.9 % under an 8 Ascend 910 environment with MindSpore, a new deep learning computing framework. Moreover, with more computational resources, THOR can only takes 2.7 minutes to 75.9 % with 256 Ascend 910.

Human attention mechanisms often work in a top-down manner, yet it is not well explored in vision research. Here, we propose the Top-Down Attention Framework (TDAF) to capture top-down attentions, which can be easily adopted in most existing models. The designed Recursive Dual-Directional Nested Structure in it forms two sets of orthogonal paths, recursive and structural ones, where bottom-up spatial features and top-down attention features are extracted respectively. Such spatial and attention features are nested deeply, therefore, the proposed framework works in a mixed top-down and bottom-up manner. Empirical evidence shows that our TDAF can capture effective stratified attention information and boost performance. ResNet with TDAF achieves 2.0% improvements on ImageNet. For object detection, the performance is improved by 2.7% AP over FCOS. For pose estimation, TDAF improves the baseline by 1.6%. And for action recognition, the 3D-ResNet adopting TDAF achieves improvements of 1.7% accuracy.

Automated temporal planning is the problem of synthesizing, starting from a model of a system, a course of actions to achieve a desired goal when temporal constraints, such as deadlines, are present in the problem. Despite considerable successes in the literature, scalability is still a severe limitation for existing planners, especially when confronted with real-world, industrial scenarios. In this paper, we aim at exploiting recent advances in reinforcement learning, for the synthesis of heuristics for temporal planning. Starting from a set of problems of interest for a specific domain, we use a customized reinforcement learning algorithm to construct a value function that is able to estimate the expected reward for as many problems as possible. We use a reward schema that captures the semantics of the temporal planning problem and we show how the value function can be transformed in a planning heuristic for a semi-symbolic heuristic search exploration of the planning model. We show on two case-studies how this method can widen the reach of current temporal planners with encouraging results.

Time-series is ubiquitous across applications, such as transportation, finance and healthcare. Time-series is often influenced by external factors, especially in the form of asynchronous events, making forecasting difficult. However, existing models are mainly designated for either synchronous time-series or asynchronous event sequence, and can hardly provide a synthetic way to capture the relation between them. We propose Variational Synergetic Multi-Horizon Network (VSMHN), a novel deep conditional generative model. To learn complex correlations across heterogeneous sequences, a tailored encoder is devised to combine the advances in deep point processes models and variational recurrent neural networks. In addition, an aligned time coding and an auxiliary transition scheme are carefully devised for batched training on unaligned sequences. Our model can be trained effectively using stochastic variational inference and generates probabilistic predictions with Monte-Carlo simulation. Furthermore, our model produces accurate, sharp and more realistic probabilistic forecasts. We also show that modeling asynchronous event sequences is crucial for multi-horizon time-series forecasting.