AAAI 2022 Papers

Self-supervised learning (SSL), especially contrastive methods, has raised attraction recently as it learns effective transferable representations without semantic annotations. A common practice for self-supervised pre-training is to use as much data as possible. For a specific downstream task, however, involving irrelevant data in pre-training may degenerate the downstream performance, observed from our extensive experiments. On the other hand, for existing SSL methods, it is burdensome and infeasible to use different downstream-task-customized datasets in pre-training for different tasks. To address this issue, we propose a novel SSL paradigm called Scalable Dynamic Routing (SDR), which can be trained once and deployed efficiently to different downstream tasks with task-customized pre-trained models. Specifically, we construct the SDRnet with various sub-nets and train each sub-net with only one subset of the data by data-aware progressive training. When a downstream task arrives, we route among all the pre-trained sub-nets to get the best along with its corresponding weights. Experiment results show that our SDR can train 256 sub-nets on ImageNet simultaneously, which provides better transfer performance than a unified model trained on the full ImageNet, achieving state-of-the-art (SOTA) averaged accuracy over 11 downstream classification tasks and AP on PASCAL VOC detection task.

Multi-modal Multi-label Emotion Recognition (MMER) aims to identify various human emotions from heterogeneous visual, audio and text modalities. Previous methods mainly focus on projecting multiple modalities into a common latent space and learning an identical representation for all labels, which neglects the diversity of each modality and fails to capture richer semantic information for each label from different perspectives. Besides, associated relationships of modalities and labels have not been fully exploited. In this paper, we propose versaTile multi-modAl learning for multI-labeL emOtion Recognition (TAILOR), aiming to refine multi-modal representations and enhance discriminative capacity of each label. Specifically, we design an adversarial multi-modal refinement module to sufficiently explore the commonality among different modalities and strengthen the diversity of each modality. To further exploit label-modal dependence, we devise a BERT-like cross-modal encoder to gradually fuse private and common modality representations in a granularity descent way, as well as a label-guided decoder to adaptively generate a tailored representation for each label with the guidance of label semantics. In addition, we conduct experiments on the benchmark MMER dataset CMU-MOSEI in both aligned and unaligned settings, which demonstrate the superiority of TAILOR over the state-of-the-arts.

In this paper, we study the zero-shot sketch-based image retrieval (ZS-SBIR) task, which retrieves natural images related to sketch queries from unseen categories. In the literature, convolutional neural networks (CNNs) have become the de-facto standard and they are either trained end-to-end or used to extract pre-trained features for images and sketches. However, CNNs are limited in modeling the global structural information of objects due to the intrinsic locality of convolution operations. To this end, we propose a Transformer-based approach called Three-Way Vision Transformer (TVT) to leverage the ability of Vision Transformer (ViT) to model global contexts due to the global self-attention mechanism. Going beyond simply applying ViT to this task, we propose a token-based strategy of adding fusion and distillation tokens and making them complementary to each other. Specifically, we integrate three ViTs, which are pre-trained on data of each modality, into a three-way pipeline through the processes of distillation and multi-modal hypersphere learning. The distillation process is proposed to supervise fusion ViT (ViT with an extra fusion token) with soft targets from modality-specific ViTs, which prevents fusion ViT from catastrophic forgetting. Furthermore, our method learns a multi-modal hypersphere by performing inter- and intra-modal alignment without loss of uniformity, which aims to bridge the modal gap between modalities of sketch and image and avoid the collapse in dimensions. Extensive experiments on three benchmark datasets, i.e., Sketchy, TU-Berlin, and QuickDraw, demonstrate the superiority of our TVT method over the state-of-the-art ZS-SBIR methods.

This paper presents TS2Vec, a universal framework for learning representations of time series in an arbitrary semantic level. Unlike existing methods, TS2Vec performs contrastive learning in a hierarchical way over augmented context views, which enables a robust contextual representation for each timestamp. Furthermore, to obtain the representation of an arbitrary sub-sequence in the time series, we can apply a simple aggregation over the representations of corresponding timestamps. We conduct extensive experiments on time series classification tasks to evaluate the quality of time series representations. As a result, TS2Vec achieves significant improvement over existing SOTAs of unsupervised time series representation on 125 UCR datasets and 29 UEA datasets. The learned timestamp-level representations also achieve superior results in time series forecasting and anomaly detection tasks. A linear regression trained on top of the learned representations outperforms previous SOTAs of time series forecasting. Furthermore, we present a simple way to apply the learned representations for unsupervised anomaly detection, which establishes SOTA results in the literature. The source code is publicly available at https://github.com/yuezhihan/ts2vec.

Random Fourier features (RFF) are a popular set of tools for constructing low-dimensional approximations of translation-invariant kernels, allowing kernel methods to be scaled to big data. Apart from their computational advantages, by working in the spectral domain random Fourier features expose the translation invariant kernel as a density function that may, in principle, be manipulated directly to tune the kernel. In this paper we propose selecting the density function from a reproducing kernel Hilbert space to allow us to search the space of all translation-invariant kernels. Our approach, which we call tuned random features (TRF), achieves this by approximating the density function as the RKHS-norm regularised least-squares best fit to an unknown ``true'' optimal density function, resulting in a RFF formulation where kernel selection is reduced to regularised risk minimisation with a novel regulariser. We derive bounds on the Rademacher complexity for our method showing that our random features approximation method converges to optimal kernel selection in the large N,D limit. Finally, we prove experimental results for a variety of real-world learning problems, demonstrating the performance of our approach compared to comparable methods.

Conventional static knowledge graphs model entities in relational data as nodes, connected by edges of specific relation types. However, information and knowledge evolve continuously, and temporal dynamics emerge, which are expected to influence future situations. In temporal knowledge graphs, time information is integrated into the graph by equipping each edge with a timestamp or a time range. Embedding-based methods have been introduced for link prediction on temporal knowledge graphs, but they mostly lack explainability and comprehensible reasoning chains. Particularly, they are usually not designed to deal with link forecasting -- event prediction involving future timestamps. We address the task of link forecasting on temporal knowledge graphs and introduce TLogic, an explainable framework that is based on temporal logical rules extracted via temporal random walks. We compare TLogic with state-of-the-art baselines on three benchmark datasets and show better overall performance while our method also provides explanations that preserve time consistency. Furthermore, in contrast to most state-of-the-art embedding-based methods, TLogic works well in the inductive setting where already learned rules are transferred to related datasets with a common vocabulary.

There has been a recent surge in learning generative models for graphs. While impressive progress has been made on static graphs, work on generative modeling of temporal graphs is at a nascent stage with significant scope for improvement. First, existing generative models do not scale with either the time horizon or the number of nodes. Second, existing techniques are transductive in nature and thus do not facilitate knowledge transfer. Finally, due to relying on one-to-one node mapping from source to the generated graph, existing models leak node identity information and do not allow up-scaling/down-scaling the source graph size. In this paper, we bridge these gaps with a novel generative model called TIGGER. TIGGER derives its power through a combination of temporal point processes with auto-regressive modeling enabling both transductive and inductive variants. Through extensive experiments on real datasets, we establish TIGGER generates graphs of superior fidelity, while also being up to 3 orders of magnitude faster than the state-of-the-art.

Robots operating in human spaces must be able to engage in natural language interaction, both understanding and executing instructions, and using conversation to resolve ambiguity and correct mistakes. To study this, we introduce TEACh, a dataset of over 3,000 human-human, interactive dialogues to complete household tasks in simulation. A Commander with access to oracle information about a task communicates in natural language with a Follower. The Follower navigates through and interacts with the environment to complete tasks varying in complexity from "Make Coffee" to "Prepare Breakfast", asking questions and getting additional information from the Commander. We propose three benchmarks using TEACh to study embodied intelligence challenges, and we evaluate initial models' abilities in dialogue understanding, language grounding, and task execution.

In recent years, tree decoders become more popular than LaTeX string decoders in the field of handwritten mathematical expression recognition (HMER) as they can capture the hierarchical tree structure of mathematical expressions. However previous tree decoders converted the tree structure labels into a fixed and ordered sequence, which could not make full use of the diversified expression of tree labels. In this study, we propose a novel tree decoder (TDv2) to fully utilize the tree structure labels. Compared with previous tree decoders, this new model does not require a fixed priority for different branches of a node during training and inference, which can effectively improve the model generalization capability. The input and output of the model make full use of the tree structure label, so that there is no need to find the parent node in the decoding process, which simplifies the decoding process and adds a prior information to help predict the node. We verified the effectiveness of each part of the model through comprehensive ablation experiments and attention visualization analysis. On the authoritative CROHME 14/16/19 datasets, our method achieves the state-of-the-art results.

Event logs are often one of the main sources of information to understand the behavior of a system. While numerous approaches have extracted partial information from event logs, in this work, we aim at inferring a global model of a system from its event logs. We consider real-time systems, which can be modeled with Timed Automata: our approach is thus a Timed Automata learner. There is a handful of related work, however, they might require a lot of parameters or produce Timed Automata that either are undeterministic or lack precision. In contrast, our proposed approach, called TAG, requires only one parameter and learns a deterministic Timed Automaton having a good tradeoff between accuracy and complexity of the automata. This allows getting an interpretable and accurate global model of the real-time system considered. Our experiments compare our approach to the related work and demonstrate its merits.

Few-shot action recognition aims to recognize novel action classes (query) using just a few samples (support). The majority of current approaches follow the metric learning paradigm, which learns to compare the similarity between videos. Recently, it has been observed that directly measuring this similarity is not ideal since different action instances may show distinctive temporal distribution, resulting in severe misalignment issues across query and support videos. In this paper, we arrest this problem from two distinct aspects -- action duration misalignment and action evolution misalignment. We address them sequentially through a Two-stage Action Alignment Network (TA2N). The first stage locates the action by learning a temporal affine transform, which warps each video feature to its action duration while dismissing the action-irrelevant feature (e.g. background). Next, the second stage coordinates query feature to match the spatial-temporal action evolution of support by performing temporally rearrange and spatially offset prediction. Extensive experiments on benchmark datasets show the potential of the proposed method in achieving state-of-the-art performance for few-shot action recognition.

Automated fact-checking is a needed technology to curtail the spread of online misinformation. One current framework for such solutions proposes to verify claims by retrieving supporting or refuting evidence from related textual sources. However, the realistic use cases for fact-checkers will require verifying claims against evidence sources that could be affected by the same misinformation. Furthermore, the development of modern NLP tools that can produce coherent, fabricated content would allow malicious actors to systematically generate adversarial disinformation for fact-checkers. In this work, we explore the sensitivity of automated fact-checkers to synthetic adversarial evidence in two simulated settings: ADVERSARIAL ADDITION, where we fabricate documents and add them to the evidence repository available to the fact-checking system, and ADVERSARIAL MODIFICATION, where existing evidence source documents in the repository are automatically altered. Our study across multiple models on three benchmarks demonstrates that these systems suffer significant performance drops against these attacks. Finally, we discuss the growing threat of modern NLG systems as generators of disinformation in the context of the challenges they pose to automated fact-checkers.

Reactive synthesis from high-level specifications that combine hard constraints expressed in Linear Temporal Logic (LTL) with soft constraints expressed by discounted sum (DS) rewards has applications in planning and reinforcement learning. An existing approach combines techniques from LTL synthesis with optimization for the DS rewards but has failed to yield a sound algorithm. An alternative approach combining LTL synthesis with satisficing DS rewards (rewards that achieve a threshold) is sound and complete for integer discount factors, but, in practice, a fractional discount factor is desired. This work extends the existing satisficing approach, presenting the first sound algorithm for synthesis from LTL and DS rewards with fractional discount factors. The utility of our algorithm is demonstrated on robotic planning domains.

The challenge of talking face generation from speech lies in aligning two different modal information, audio and video, such that the mouth region corresponds to input audio. Previous methods either exploit audio-visual representation learning or leverage intermediate structural information such as landmarks and 3D models. However, they struggle to synthesize fine details of the lips varying at the phoneme level as they do not sufficiently provide visual information of the lips at the video synthesis step. To overcome this limitation, our work proposes Audio-Lip Memory that brings in visual information of the mouth region corresponding to input audio and enforces fine-grained audio-visual coherence. It stores lip motion features from sequential ground truth images in the value memory and aligns them with corresponding audio features so that they can be retrieved using audio input at inference time. Therefore, using the retrieved lip motion features as visual hints, it can easily correlate audio with visual dynamics in the synthesis step. By analyzing the memory, we demonstrate that unique lip features are stored in each memory slot at the phoneme level, capturing subtle lip motion based on memory addressing. In addition, we introduce visual-visual synchronization loss which can enhance lip-syncing performance when used along with audio-visual synchronization loss in our model. Extensive experiments are performed to verify that our method generates high-quality video with mouth shapes that best align with the input audio, outperforming previous state-of-the-art methods.

Neural sequence models trained with maximum likelihood estimation have led to breakthroughs in many tasks, where success is defined by the gap between training and test performance. However, their ability to achieve stronger forms of generalization remains unclear. We consider the problem of symbolic mathematical integration, as it requires generalizing systematically beyond the training set. We develop a methodology for evaluating generalization that takes advantage of the problem domain's structure and access to a verifier. Despite promising in-distribution performance of sequence-to-sequence models in this domain, we demonstrate challenges in achieving robustness, compositionality, and out-of-distribution generalization, through both carefully constructed manual test suites and a genetic algorithm that automatically finds large collections of failures in a controllable manner. Our investigation highlights the difficulty of generalizing well with the predominant modeling and learning approach, and the importance of evaluating beyond the test set, across different aspects of generalization.

Despite the great progress in video understanding made by deep convolutional neural networks, feature representation learned by existing methods may be biased to static visual cues. To address this issue, we propose a novel method to suppress static visual cues (SSVC) based on probabilistic analysis for self-supervised video representation learning. In our method, video frames are first encoded to obtain latent variables under standard normal distribution via normalizing flows. By modelling static factors in a video as a random variable, the conditional distribution of each latent variable becomes shifted and scaled normal. Then, the less-varying latent variables along time are selected as static cues and suppressed to generate motion-preserved videos. Finally, positive pairs are constructed by motion-preserved videos for contrastive learning to alleviate the problem of representation bias to static cues. The less-biased video representation can be better generalized to various downstream tasks. Extensive experiments on publicly available benchmarks demonstrate that the proposed method outperforms the state of the art when only single RGB modality is used for pre-training.

Natural Language Inference (NLI) models are known to learn from biases and artefacts within their training data, impacting how well they generalise to other unseen datasets. Existing de-biasing approaches focus on preventing the models from learning these biases, which can result in restrictive models and lower performance. We instead investigate teaching the model how a human would approach the NLI task, in order to learn features that will generalise better to previously unseen examples. Using natural language explanations, we supervise the model’s attention weights to encourage more attention to be paid to the words present in the explanations, significantly improving model performance. Our experiments show that the in-distribution improvements of this method are also accompanied by out-of-distribution improvements, with the supervised models learning from features that generalise better to other NLI datasets. Analysis of the model indicates that human explanations encourage increased attention on the important words, with more attention paid to words in the premise and less attention paid to punctuation and stopwords.

Recent work has unveiled a theory for reasoning about the decisions made by binary classifiers: a classifier describes a Boolean function, and the reasons behind an instance being classified as positive are the prime-implicants of the function that are satisfied by the instance. One drawback of these works is that they do not explicitly treat scenarios where the underlying data is known to be constrained, e.g., certain combinations of features may not exist, may not be observable, or may be required to be disregarded. We propose a more general theory, also based on prime-implicants, tailored to taking constraints into account. The main idea is to view classifiers as describing partial Boolean functions that are undefined on instances that do not satisfy the constraints. We prove that this simple idea results in more parsimonious reasons. That is, not taking constraints into account (e.g., ignoring, or taking them as negative instances) results in reasons that are subsumed by reasons that do take constraints into account. We illustrate this improved succinctness on synthetic classifiers and classifiers learnt from real data.

Many data applications have certain invariant constraints due to practical needs. Data curators who employ differential privacy need to respect such constraints on the sanitized data product as a primary utility requirement. Invariants challenge the formulation, implementation, and interpretation of privacy guarantees. We propose subspace differential privacy, to honestly characterize the dependence of the sanitized output on confidential aspects of the data. We discuss two design frameworks that convert well-known differentially private mechanisms, such as the Gaussian and the Laplace mechanisms, to subspace differentially private ones that respect the invariants specified by the curator. For linear queries, we discuss the design of near-optimal mechanisms that minimize the mean squared error. Subspace differentially private mechanisms rid the need for post-processing due to invariants, preserve transparency and statistical intelligibility of the output, and can be suitable for distributed implementation. We showcase the proposed mechanisms on the 2020 Census Disclosure Avoidance demonstration data, and a spatio-temporal dataset of mobile access point connections on a large university campus.