AAAI 2021 论文列表

We propose two generic methods for improving semi-supervised learning (SSL). The first integrates weight perturbation (WP) into existing “consistency regularization” (CR) based methods. We implement WP by leveraging variational Bayesian inference (VBI). The second method proposes a novel consistency loss called “maximum uncertainty regularization” (MUR). While most consistency losses act on perturbations in the vicinity of each data point, MUR actively searches for “virtual” points situated beyond this region that cause the most uncertain class predictions. This allows MUR to impose smoothness on a wider area in the input-output manifold. Our experiments show clear improvements in classification errors of various CR based methods when they are combined with VBI or MUR or both.

We address the challenging problem of semi-supervised learning in the context of multiple visual interpretations of the world by finding consensus in a graph of neural networks. Each graph node is a scene interpretation layer, while each edge is a deep net that transforms one layer at one node into another from a different node. During the supervised phase edge networks are trained independently. During the next unsupervised stage edge nets are trained on the pseudo-ground truth provided by consensus among multiple paths that reach the nets' start and end nodes. These paths act as ensemble teachers for any given edge and strong consensus is used for high-confidence supervisory signal. The unsupervised learning process is repeated over several generations, in which each edge becomes a "student" and also part of different ensemble "teachers" for training other students. By optimizing such consensus between different paths, the graph reaches consistency and robustness over multiple interpretations and generations, in the face of unknown labels. We give theoretical justifications of the proposed idea and validate it on a large dataset. We show how prediction of different representations such as depth, semantic segmentation, surface normals and pose from RGB input could be effectively learned through self-supervised consensus in our graph. We also compare to state-of-the-art methods for multi-task and semi-supervised learning and show superior performance.

Sequence classification is essential for domains from medical diagnosis to online advertising. In these settings, data are typically proprietary, and annotations are expensive to acquire. Often times, so few annotations are available that training a robust model from scratch is impractical. Recently, knowledge amalgamation (KA) has emerged as a promising strategy for training models without this hard-to-come-by labeled training dataset. To achieve this, KA methods combine the knowledge of multiple pre-trained teacher models (trained on different classification tasks and proprietary datasets) into one student model that becomes an expert on the union of all teachers’ classes. However, we demonstrate that the state-of-the-art solutions fail in the presence of overconfident teachers, which make confident but incorrect predictions for instances from classes upon which they were not trained. Additionally, to-date no work has explored KA for sequence models. Therefore, we propose and then solve the open problem of semi-supervised KA for sequence classification (SKA). Our SKA approach first learns to estimate how trustworthy each teacher is for a given instance, then rescales the predicted probabilities from all teachers to supervise a student model. Our solution overcomes overconfident teachers through careful use of a very small amount of labeled instances. We demonstrate that this approach beats eight state-of-the-art alternatives on four real-world datasets by on average 15% in accuracy with as little as 2% of training data being annotated.

For natural language understanding tasks, either machine reading comprehension or natural language inference, both semantics-aware and inference are favorable features of the concerned modeling for better understanding performance. Thus we propose a Semantics-Aware Inferential Network (SAIN) to meet such a motivation. Taking explicit contextualized semantics as a complementary input, the inferential module of SAIN enables a series of reasoning steps over semantic clues through an attention mechanism. By stringing these steps, the inferential network effectively learns to perform iterative reasoning which incorporates both explicit semantics and contextualized representations. In terms of well pre-trained language models as front-end encoder, our model achieves significant improvement on 11 tasks including machine reading comprehension and natural language inference.

Advances in NLP have yielded impressive results for the task of machine reading comprehension (MRC), with approaches having been reported to achieve performance comparable to that of humans. In this paper, we investigate whether state-of-the-art MRC models are able to correctly process Semantics Altering Modifications (SAM): linguistically-motivated phenomena that alter the semantics of a sentence while preserving most of its lexical surface form. We present a method to automatically generate and align challenge sets featuring original and altered examples. We further propose a novel evaluation methodology to correctly assess the capability of MRC systems to process these examples independent of the data they were optimised on, by discounting for effects introduced by domain shift. In a large-scale empirical study, we apply the methodology in order to evaluate extractive MRC models with regard to their capability to correctly process SAM-enriched data. We comprehensively cover 12 different state-of-the-art neural architecture configurations and four training datasets and find that -- despite their well-known remarkable performance -- optimised models consistently struggle to correctly process semantically altered data.

Generalized zero-shot Learning (GZSL) aims to recognize images from either seen or unseen domain, mainly by learning a joint embedding space to associate image features with the corresponding category descriptions. Recent methods have proved that localizing important object regions can effectively bridge the semantic-visual gap. However, these are all based on one-off visual localizers, lacking of interpretability and flexibility. In this paper, we propose a novel Semantic-guided Reinforced Region Embedding (SR2E) network that can localize important objects in the long-term interests to construct semantic-visual embedding space. SR2E consists of Reinforced Region Module (R2M) and Semantic Alignment Module (SAM). First, without the annotated bounding box as supervision, R2M encodes the semantic category guidance into the reward and punishment criteria to teach the localizer serialized region searching. Besides, R2M explores different action spaces during the serialized searching path to avoid local optimal localization, which thereby generates discriminative visual features with less redundancy. Second, SAM preserves the semantic relationship into visual features via semantic-visual alignment and designs a domain detector to alleviate the domain confusion. Experiments on four public benchmarks demonstrate that the proposed SR2E is an effective GZSL method with reinforced embedding space, which obtains averaged 6.1% improvements.

We study the task of semantic mapping – specifically, an embodied agent (a robot or an egocentric AI assistant) is given a tour of a new environment and asked to build an allocentric top-down semantic map (‘what is where?’) from egocentric observations of an RGB-D camera with known pose (via localization sensors). Importantly, our goal is to build neural episodic memories and spatio-semantic representations of 3D spaces that enable the agent to easily learn subsequent tasks in the same space – navigating to objects seen during the tour (‘Find chair’) or answering questions about the space (‘How many chairs did you see in the house?’). Towards this goal, we present Semantic MapNet (SMNet), which consists of: (1) an Egocentric Visual Encoder that encodes each egocentric RGB-D frame, (2) a Feature Projector that projects egocentric features to appropriate locations on a floor-plan, (3) a Spatial Memory Tensor of size floor-plan length×width×feature-dims that learns to accumulate projected egocentric features, and (4) a Map Decoder that uses the memory tensor to produce semantic top-down maps. SMNet combines the strengths of (known) projective camera geometry and neural representation learning. On the task of semantic mapping in the Matterport3D dataset, SMNet significantly outperforms competitive baselines by 4.01−16.81% (absolute) on mean-IoU and 3.81−19.69% (absolute) on Boundary-F1 metrics. Moreover, we show how to use the spatio-semantic allocentric representations build by SMNet for the task of ObjectNav and Embodied Question Answering. Project page: https://vincentcartillier.github.io/smnet.html.

This paper considers a video caption generating network referred to as Semantic Grouping Network (SGN) that attempts (1) to group video frames with discriminating word phrases of partially decoded caption and then (2) to decode those semantically aligned groups in predicting the next word. As consecutive frames are not likely to provide unique information, prior methods have focused on discarding or merging repetitive information based only on the input video. The SGN learns an algorithm to capture the most discriminating word phrases of the partially decoded caption and a mapping that associates each phrase to the relevant video frames - establishing this mapping allows semantically related frames to be clustered, which reduces redundancy. In contrast to the prior methods, the continuous feedback from decoded words enables the SGN to dynamically update the video representation that adapts to the partially decoded caption. Furthermore, a contrastive attention loss is proposed to facilitate accurate alignment between a word phrase and video frames without manual annotations. The SGN achieves state-of-the-art performances by outperforming runner-up methods by a margin of 2.1%p and 2.4%p in a CIDEr-D score on MSVD and MSR-VTT datasets, respectively. Extensive experiments demonstrate the effectiveness and interpretability of the SGN.

Detecting 3D objects from point clouds is a significant yet challenging issue in many applications. While most existing approaches seek to leverage geometric information of point clouds, few studies accommodate the inherent semantic characteristics of each point and the consistency between the geometric and semantic cues. In this work, we propose a novel semantic consistency network (SCNet) driven by a natural principle: the class of a predicted 3D bounding box should be consistent with the classes of all the points inside this box. Specifically, our SCNet consists of a feature extraction structure, a detection decision structure, and a semantic segmentation structure. In inference, the feature extraction and the detection decision structures are used to detect 3D objects. In training, the semantic segmentation structure is jointly trained with the other two structures to produce more robust and applicative model parameters. A novel semantic consistency loss is proposed to regulate the output 3D object boxes and the segmented points to boost the performance. Our model is evaluated on two challenging datasets and achieves comparable results to the state-of-the-art methods.

Understanding real-world networks is a core research endeavor within the last two decades. Network Creation Games are a promising approach for this from a game-theoretic perspective. In these games, selfish agents corresponding to nodes in a network strategically decide which links to form to optimize their centrality. Many versions have been introduced and analyzed, but none of them fits to modeling the evolution of social networks. In real-world social networks connections are often established by recommendations from common acquaintances or by a chain of such recommendations. Thus establishing and maintaining a contact with a friend of a friend is easier than connecting to complete strangers. This explains the high clustering, i.e., the abundance of triangles, in real-world social networks. We propose and analyze a network creation model inspired by real-world social networks. In our model edges are formed via bilateral consent of both endpoints and the cost for establishing and maintaining an edge is proportional to the distance of the endpoints before establishing the connection. We provide results for generic cost functions which essentially only must be convex functions in the distance of the endpoints without the respective edge. For this broad class of cost functions we provide many structural properties of equilibrium networks and prove (almost) tight bounds on the diameter, the Price of Anarchy and the Price of Stability. Moreover, as a proof-of-concept we show via experiments that the created equilibrium networks of our model indeed closely mimic real-world social networks. We observe degree distributions that seem to follow a power-law, high clustering, and low diameters. This can be seen as a promising first step towards game-theoretic network creation models that predict networks featuring all core real-world properties.

Video Question Answering (VideoQA) requires fine-grained understanding of both video and language modalities to answer the given questions. In this paper, we propose novel training schemes for multiple-choice video question answering with a self-supervised pre-training stage and a supervised contrastive learning in the main stage as an auxiliary learning. In the self-supervised pre-training stage, we transform the original problem format of predicting the correct answer into the one that predicts the relevant question to provide a model with broader contextual inputs without any further dataset or annotation. For contrastive learning in the main stage, we add a masking noise to the input corresponding to the ground-truth answer, and consider the original input of the ground-truth answer as a positive sample, while treating the rest as negative samples. By mapping the positive sample closer to the masked input, we show that the model performance is improved. We further employ locally aligned attention to focus more effectively on the video frames that are particularly relevant to the given corresponding subtitle sentences. We evaluate our proposed model on highly competitive benchmark datasets related to multiple-choice video QA: TVQA, TVQA+, and DramaQA. Experimental results show that our model achieves state-of-the-art performance on all datasets. We also validate our approaches through further analyses.

Recent studies have witnessed that self-supervised methods based on view synthesis obtain clear progress on multi-view stereo (MVS). However, existing methods rely on the assumption that the corresponding points among different views share the same color, which may not always be true in practice. This may lead to unreliable self-supervised signal and harm the final reconstruction performance. To address the issue, we propose a framework integrated with more reliable supervision guided by semantic co-segmentation and data-augmentation. Specially, we excavate mutual semantic from multi-view images to guide the semantic consistency. And we devise effective data-augmentation mechanism which ensures the transformation robustness by treating the prediction of regular samples as pseudo ground truth to regularize the prediction of augmented samples. Experimental results on DTU dataset show that our proposed methods achieve the state-of-the-art performance among unsupervised methods, and even compete on par with supervised methods. Furthermore, extensive experiments on Tanks&Temples dataset demonstrate the effective generalization ability of the proposed method.

While various neural machine translation (NMT) methods have integrated mono-lingual syntax knowledge into the linguistic representation of sequence-to-sequence, no research is available on aligning the syntactic structures of target language with the corresponding source language syntactic structures. This work shows the first attempt of a source-target bilingual syntactic alignment approach SyntAligner by mutual information maximization-based self-supervised neural deep modeling. Building on the word alignment for NMT, our SyntAligner firstly aligns the syntactic structures of source and target sentences and then maximizes their mutual dependency by introducing a lower bound on their mutual information. In SyntAligner, the syntactic structure of span granularity is represented by transforming source or target word hidden state into a source or target syntactic span vector. A border-sensitive span attention mechanism then captures the correlation between the source and target syntactic span vectors, which also captures the self-attention between span border-words as alignment bias. Lastly, a self-supervised bilingual syntactic mutual information maximization-based learning objective dynamically samples the aligned syntactic spans to maximize their mutual dependency. Experiment results on three typical NMT tasks: WMT'14 English to German, IWSLT'14 German to English, and NC'11 English to French show the SyntAligner effectiveness and universality of syntactic alignment.

The Q-learning algorithm is known to be affected by the maximization bias, i.e. the systematic overestimation of action values, an important issue that has recently received renewed attention. Double Q-learning has been proposed as an efficient algorithm to mitigate this bias. However, this comes at the price of an underestimation of action values, in addition to increased memory requirements and a slower convergence. In this paper, we introduce a new way to address the maximization bias in the form of a "self-correcting algorithm" for approximating the maximum of an expected value. Our method balances the overestimation of the single estimator used in conventional Q-learning and the underestimation of the double estimator used in Double Q-learning. Applying this strategy to Q-learning results in Self-correcting Q-learning. We show theoretically that this new algorithm enjoys the same convergence guarantees as Q-learning while being more accurate. Empirically, it performs better than Double Q-learning in domains with rewards of high variance, and it even attains faster convergence than Q-learning in domains with rewards of zero or low variance. These advantages transfer to a Deep Q Network implementation that we call Self-correcting DQN and which outperforms regular DQN and Double DQN on several tasks in the Atari 2600 domain.

Imagining a colored realistic image from an arbitrary-drawn sketch is one of human capabilities that we eager machines to mimic. Unlike previous methods that either require the sketch-image pairs or utilize low-quantity detected edges as sketches, we study the exemplar-based sketch-to-image (s2i) synthesis task in a self-supervised learning manner, eliminating the necessity of the paired sketch data. To this end, we first propose an unsupervised method to efficiently synthesize line-sketches for general RGB-only datasets. With the synthetic paired-data, we then present a self-supervised Auto-Encoder (AE) to decouple the content/style features from sketches and RGB-images, and synthesize images both content-faithful to the sketches and style-consistent to the RGB-images. While prior works employ either the cycle-consistence loss or dedicated attentional modules to enforce the content/style fidelity, we show AE's superior performance with pure self-supervisions. To further improve the synthesis quality in high resolution, we also leverage an adversarial network to refine the details of synthetic images. Extensive experiments on $1024^2$ resolution demonstrate a new state-of-art-art performance of the proposed model on CelebA-HQ and Wiki-Art datasets. Moreover, with the proposed sketch generator, the model shows a promising performance on style mixing and style transfer, which the synthesized images are not only style-consistent but also semantically meaningful.

Labeling training examples at scale is a perennial challenge in machine learning. Self-supervision methods compensate for the lack of direct supervision by leveraging prior knowledge to automatically generate noisy labeled examples. Deep probabilistic logic (DPL) is a unifying framework for self-supervised learning that represents unknown labels as latent variables and incorporates diverse self-supervision using probabilistic logic to train a deep neural network end-to-end using variational EM. While DPL is successful at combining pre-specified self-supervision, manually crafting self-supervision to attain high accuracy may still be tedious and challenging. In this paper, we propose Self-Supervised Self-Supervision (S4), which adds to DPL the capability to learn new self-supervision automatically. Starting from an initial "seed," S4 iteratively uses the deep neural network to propose new self-supervision. These are either added directly (a form of structured self-training) or verified by a human expert (as in feature-based active learning). Experiments show that S4 is able to automatically propose accurate self-supervision and can often nearly match the accuracy of supervised methods with a tiny fraction of the human effort.

Event-based corporate profiling aims to assess the evolving operational status of the corresponding corporate from its event sequence. Existing studies on corporate profiling have partially addressed the problem via (i) case-by-case empirical analysis by leveraging traditional financial methods, or (ii) the automatic profile inference by reformulating the problem into a supervised learning task. However, both approaches heavily rely on domain knowledge and are labor-intensive. More importantly, the task-specific nature of both approaches prevents the obtained corporate profiles from being applied to diversified downstream applications. To this end, in this paper, we propose a Self-Supervised Prototype Representation Learning (SePaL) framework for dynamic corporate profiling. By exploiting the topological information of an event graph and exploring self-supervised learning techniques, SePaL can obtain unified corporate representations that are robust to event noises and can be easily fine-tuned to benefit various down-stream applications with only a few annotated data. Specifically, we first infer the initial cluster distribution of noise-resistant event prototypes based on latent representations of events. Then, we construct four permutation-invariant self-supervision signals to guide the representation learning of the event prototype. In terms of applications, we exploit the learned time-evolving corporate representations for both stock price spike prediction and corporate default risk evaluation. Experimental results on two real-world corporate event datasets demonstrate the effectiveness of SePaL for these two applications.

Session-based recommendation (SBR) focuses on next-item prediction at a certain time point. As user profiles are generally not available in this scenario, capturing the user intent lying in the item transitions plays a pivotal role. Recent graph neural networks (GNNs) based SBR methods regard the item transitions as pairwise relations, which neglect the complex high-order information among items. Hypergraph provides a natural way to capture beyond-pairwise relations, while its potential for SBR has remained unexplored. In this paper, we fill this gap by modeling session-based data as a hypergraph and then propose a dual channel hypergraph convolutional network -- DHCN to improve SBR. Moreover, to enhance hypergraph modeling, we innovatively integrate self-supervised learning into the training of our network by maximizing mutual information between the session representations learned via the two channels in DHCN, serving as an auxiliary task to improve the recommendation task. Extensive experiments on three benchmark datasets demonstrate the superiority of our model over the SOTA methods, and the ablation study validates the effectiveness and rationale of hypergraph modeling and self-supervised task. The implementation of our model is available via https://github.com/xiaxin1998/DHCN.

Visual saliency has emerged as a major visualization tool for interpreting deep reinforcement learning (RL) agents. However, much of the existing research uses it as an analyzing tool rather than an inductive bias for policy learning. In this work, we use visual attention as an inductive bias for RL agents. We propose a novel self-supervised attention learning approach which can 1. learn to select regions of interest without explicit annotations, and 2. act as a plug for existing deep RL methods to improve the learning performance. We empirically show that the self-supervised attention-aware deep RL methods outperform the baselines in the context of both the rate of convergence and performance. Furthermore, the proposed self-supervised attention is not tied with specific policies, nor restricted to a specific scene. We posit that the proposed approach is a general self-supervised attention module for multi-task learning and transfer learning, and empirically validate the generalization ability of the proposed method. Finally, we show that our method learns meaningful object keypoints highlighting improvements both qualitatively and quantitatively.