AAAI 2021 Papers

Large model size and high computational complexity prevent the neural machine translation (NMT) models from being deployed to low resource devices (e.g. mobile phones). Due to the large vocabulary, a large storage memory is required for the word embedding matrix in NMT models, in the meantime, high latency is introduced when constructing the word probability distribution. Based on reusing the word embedding matrix in the softmax layer, it is possible to handle the two problems brought by large vocabulary at the same time. In this paper, we propose Partial Vector Quantization (P-VQ) for NMT models, which can both compress the word embedding matrix and accelerate word probability prediction in the softmax layer. With P-VQ, the word embedding matrix is split into two low dimensional matrices, namely the shared part and the exclusive part. We compress the shared part by vector quantization and leave the exclusive part unchanged to maintain the uniqueness of each word. For acceleration, in the softmax layer, we replace most of the multiplication operations with the efficient looking-up operations based on our compression to reduce the computational complexity. Furthermore, we adopt curriculum learning and compact the word embedding matrix gradually to improve the compression quality. Experimental results on the Chinese-to-English translation task show that our method can reduce 74.35% of parameters of the word embedding and 74.42% of the FLOPs of the softmax layer. Meanwhile, the average BLEU score on the WMT test sets only drops 0.04.

In this paper, we propose an approach to effectively accelerating the computation of continuous normalizing flow (CNF), which has been proven to be a powerful tool for the tasks such as variational inference and density estimation. The training time cost of CNF can be extremely high because the required number of function evaluations (NFE) for solving corresponding ordinary differential equations (ODE) is very large. We think that the high NFE results from large truncation errors of solving ODEs. To address the problem, we propose to add a regularization. The regularization penalizes the difference between the trajectory of the ODE and its fitted polynomial regression. The trajectory of ODE will approximate a polynomial function, and thus the truncation error will be smaller. Furthermore, we provide two proofs and claim that the additional regularization does not harm training quality. Experimental results show that our proposed method can result in 42.3% to 71.3% reduction of NFE on the task of density estimation, and 19.3% to 32.1% reduction of NFE on variational auto-encoder, while the testing losses are not affected.

Outliers negatively affect the accuracy of data analysis. In this paper we are concerned with their influence on the accuracy of Principal Component Analysis (PCA). Algorithms that attempt to detect outliers and remove them from the data prior to applying PCA are sometimes called Robust PCA, or Robust Subspace Recovery algorithms. We propose a new algorithm for outlier detection that combines two ideas. The first is "chunk recursive elimination" that was used effectively to accelerate feature selection, and the second is combinatorial search, in a setting similar to A*. Our main result is showing how to combine these two ideas. One variant of our algorithm is guaranteed to compute an optimal solution according to some natural criteria, but its running time makes it impractical for large datasets. Other variants are much faster and come with provable bounds on sub-optimality. Experimental results show the effectiveness of the proposed approach.

Estimating the complete 3D point cloud from an incomplete one lies at the core of many vision and robotics applications. Existing methods typically predict the complete point cloud based on the global shape representation extracted from the incomplete input. Although they could predict the overall shape of 3D objects, they are incapable of generating structure details of objects. Moreover, the partial input point sets obtained from range scans are often sparse, noisy and non-uniform, which largely hinder shape completion. In this paper, we propose an adaptive sampling and hierarchical folding network (ASHF-Net) for robust 3D point cloud completion. Our main contributions are two-fold. First, we propose a denoising auto-encoder with an adaptive sampling module, aiming at learning robust local region features that are insensitive to noise. Second, we propose a hierarchical folding decoder with the gated skip-attention and multi-resolution completion goal to effectively exploit the local structure details of partial inputs. We also design a KL regularization term to evenly distribute the generated points. Extensive experiments demonstrate that our method outperforms existing state-of-the-art methods on multiple 3D point cloud completion benchmarks.

Knowledge distillation is considered as a training and compression strategy in which two neural networks, namely a teacher and a student, are coupled together during training. The teacher network is supposed to be a trustworthy predictor and the student tries to mimic its predictions. Usually, a student with a lighter architecture is selected so we can achieve compression and yet deliver high-quality results. In such a setting, distillation only happens for final predictions whereas the student could also benefit from teacher’s supervision for internal components. Motivated by this, we studied the problem of distillation for intermediate layers. Since there might not be a one-to-one alignment between student and teacher layers, existing techniques skip some teacher layers and only distill from a subset of them. This shortcoming directly impacts quality, so we instead propose a combinatorial technique which relies on attention. Our model fuses teacher-side information and takes each layer’s significance into consideration, then it performs distillation between combined teacher layers and those of the student. Using our technique, we distilled a 12-layer BERT (Devlin et al. 2019) into 6-, 4-, and 2-layer counterparts and evaluated them on GLUE tasks (Wang et al. 2018). Experimental results show that our combinatorial approach is able to outperform other existing techniques.

Although artificial intelligence has revolutionized data analysis, significantly less work has focused on using AI to improve scientific data collection. Past work in AI for data collection has typically assumed the objective function is well-defined by humans before starting an experiment; however, this is a poor fit for scientific domains where new discoveries and insights are made as data is being collected. In this paper we present a new framework to allow AI systems to work together with humans (e.g. scientists) to collect data more effectively in simple scientific domains. We present a novel algorithm, TESA, which seeks to achieve good performance by learning from past human behavior how to direct data to places that are likely to become scientifically interesting in the future. We analyze the problem theoretically, defining a novel notion of regret in this setting and showing that TESA is zero regret. Next, we show that TESA outperforms other related algorithms in simulations using real data drawn from three diverse domains (economics, mental health, and cognitive psychology). Finally, we run experiments with human subjects across these scientific domains to compare our iterative human-in-the-loop process to a (more standard) workflow in which information is communicated to the AI a priori.

Incorporating second-order curvature information into machine learning optimization algorithms can be subtle, and doing so naïvely can lead to high per-iteration costs associated with forming the Hessian and performing the associated linear system solve. To address this, we introduce ADAHESSIAN, a new stochastic optimization algorithm. ADAHESSIAN directly incorporates approximate curvature information from the loss function, and it includes several novel performance-improving features, including: (i) a fast Hutchinson based method to approximate the curvature matrix with low computational overhead; (ii) a spatial averaging to reduce the variance of the second derivative; and (iii) a root-mean-square exponential moving average to smooth out variations of the second-derivative across different iterations. We perform extensive tests on NLP, CV, and recommendation system tasks, and ADAHESSIAN achieves state-of-the-art results. In particular, we find that ADAHESSIAN: (i) outperforms AdamW for transformers by0.13/0.33 BLEU score on IWSLT14/WMT14, 2.7/1.0 PPLon PTB/Wikitext-103; (ii) outperforms AdamW for Squeeze-Bert by 0.41 points on GLUE; (iii) achieves 1.45%/5.55%higher accuracy on ResNet32/ResNet18 on Cifar10/ImageNetas compared to Adam; and (iv) achieves 0.032% better score than Adagrad for DLRM on the Criteo Ad Kaggle dataset. The cost per iteration of ADAHESSIANis comparable to first-order methods, and ADAHESSIAN exhibits improved robustness towards variations in hyperparameter values. The code for ADAHESSIAN is open-sourced and publicly-available [1].

Recently, Transformer has been demonstrating promising performance in many NLP tasks and showing a trend of replacing Recurrent Neural Network (RNN). Meanwhile, less attention is drawn to Convolutional Neural Network (CNN) due to its weak ability in capturing sequential and long-distance dependencies, although it has excellent local feature extraction capability. In this paper, we introduce an Attentive Convolutional Transformer (ACT) that takes the advantages of both Transformer and CNN for efficient text classification. Specifically, we propose a novel attentive convolution mechanism that utilizes the semantic meaning of convolutional filters attentively to transform text from complex word space to a more informative convolutional filter space where important n-grams are captured. ACT is able to capture both local and global dependencies effectively while preserving sequential information. Experiments on various text classification tasks and detailed analyses show that ACT is a lightweight, fast, and effective universal text classifier, outperforming CNNs, RNNs, and attentive models including Transformer.

The object of Weakly-supervised Temporal Action Localization (WS-TAL) is to localize all action instances in an untrimmed video with only video-level supervision. Due to the lack of frame-level annotations during training, current WS-TAL methods rely on attention mechanisms to localize the foreground snippets or frames that contribute to the video-level classification task. This strategy frequently confuse context with the actual action, in the localization result. Separating action and context is a core problem for precise WS-TAL, but it is very challenging and has been largely ignored in the literature. In this paper, we introduce an Action-Context Separation Network (ACSNet) that explicitly takes into account context for accurate action localization. It consists of two branches (i.e., the Foreground-Background branch and the Action-Context branch). The Foreground-Background branch first distinguishes foreground from background within the entire video while the Action-Context branch further separates the foreground as action and context. We associate video snippets with two latent components (i.e., a positive component and a negative component), and their different combinations can effectively characterize foreground, action and context. Furthermore, we introduce extended labels with auxiliary context categories to facilitate the learning of action-context separation. Experiments on THUMOS14 and ActivityNet v1.2/v1.3 datasets demonstrate the ACSNet outperforms existing state-of-the-art WS-TAL methods by a large margin.

Stochastic gradient descent (SGD) is a widely used method for its outstanding generalization ability and simplicity. Adaptive gradient methods have been proposed to further accelerate the optimization process. In this paper, we revisit existing adaptive gradient optimization methods with a new interpretation. Such new perspective leads to a refreshed understanding of the roles of second moments in stochastic optimization. Based on this, we propose Angle-Calibration Moment method (ACMo), a novel stochastic optimization method. It enjoys the benefits of second moments with only first moment updates. Theoretical analysis shows that ACMo is able to achieve the same convergence rate as mainstream adaptive methods. Experiments on a variety of CV and NLP tasks demonstrate that ACMo has a comparable convergence to state-of-the-art Adam-type optimizers, and even a better generalization performance in most cases. The code is available at https://github.com/Xunpeng746/ACMo.

Sequential recommender systems (SRS) have become a research hotspot in recent studies. Because of the requirement in capturing user's dynamic interests, sequential neural network based recommender models often need to be stacked with more hidden layers (e.g., up to 100 layers) compared with standard collaborative filtering methods. However, the high network latency has become the main obstacle when deploying very deep recommender models into a production environment. In this paper, we argue that the typical prediction framework that treats all users equally during the inference phase is inefficient in running time, as well as sub-optimal in accuracy. To resolve such an issue, we present SkipRec, an adaptive inference framework by learning to skip inactive hidden layers on a per-user basis. Specifically, we devise a policy network to automatically determine which layers should be retained and which layers are allowed to be skipped, so as to achieve user-specific decisions. To derive the optimal skipping policy, we propose using gumbel softmax and reinforcement learning to solve the non-differentiable problem during backpropagation. We perform extensive experiments on three real-world recommendation datasets, and demonstrate that SkipRec attains comparable or better accuracy with much less inference time.

Attribution methods have been developed to understand the decision making process of machine learning models, especially deep neural networks, by assigning importance scores to individual features. Existing attribution methods often built upon empirical intuitions and heuristics. There still lacks a general and theoretical framework that not only can unify these attribution methods, but also theoretically reveal their rationales, fidelity, and limitations. To bridge the gap, in this paper, we propose a Taylor attribution framework and reformulate seven mainstream attribution methods into the framework. Based on reformulations, we analyze the attribution methods in terms of rationale, fidelity, and limitation. Moreover, We establish three principles for a good attribution in the Taylor attribution framework, i.e., low approximation error, correct contribution assignment, and unbiased baseline selection. Finally, we empirically validate the Taylor reformulations, and reveal a positive correlation between the attribution performance and the number of principles followed by the attribution method via benchmarking on real-world datasets.

Pretrained language models have recently advanced a wide range of natural language processing tasks. Nowadays, the application of pretrained language models to IR tasks has also achieved impressive results. Typical methods either directly apply a pretrained model to improve the re-ranking stage, or use it to conduct passage expansion and term weighting for first-stage retrieval. We observe that the passage ranking and passage expansion tasks share certain inherent relations, and can benefit from each other. Therefore, in this paper, we propose a general pretraining framework to enhance both tasks with Unified Encoder-Decoder networks (UED). The overall ranking framework consists of two parts in a cascade manner: (1) passage expansion with a pretraining-based query generation method; (2) re-ranking of passage candidates from a traditional retrieval method with a pretrained transformer encoder. Both the two parts are based on the same pretrained UED model, where we jointly train the passage ranking and query generation tasks for further improving the full ranking pipeline. An extensive set of experiments have been conducted on two large-scale passage retrieval datasets to demonstrate the state-of-the-art results of the proposed framework in both the first-stage retrieval and the final re-ranking. In addition, we successfully deploy the framework to our online production system, which can stably serve industrial applications with a request volume of up to 100 QPS in less than 300ms.

Joint extraction of entities and relations focuses on detecting entity pairs and their relations simultaneously with a unified model. Based on the extraction order, previous works mainly solve this task through relation-last, relation-first and relation-middle manner. However, these methods still suffer from the template-dependency, non-entity detection and non-predefined relation prediction problem. To overcome these challenges, in this paper, we propose a unified multi-task learning framework to divide the task into three interacted sub-tasks. Specifically, we first introduce the type-attentional method for subject extraction to provide prior type information explicitly. Then, the subject-aware relation prediction is presented to select useful relations based on the combination of global and local semantics. Third, we propose a question generation based QA method for object extraction to obtain diverse queries automatically. Notably, our method detects subjects or objects without relying on NER models and thus it is capable of dealing with the non-entity scenario. Finally, three sub-tasks are integrated into a unified model through parameter sharing. Extensive experiments demonstrate that the proposed framework outperforms all the baseline methods on two benchmark datasets, and further achieve excellent performance for non-predefined relations.

Existing methods of adversarial attacks successfully generate adversarial examples to confuse Deep Neural Networks (DNNs) of image classification and object detection, resulting in wrong predictions. However, these methods are difficult to attack models of video object tracking, because the tracking algorithms could handle sequential information across video frames and the categories of targets tracked are normally unknown in advance. In this paper, we propose a Unified and Effective Network, named UEN, to attack visual object tracking models. There are several appealing characteristics of UEN: (1) UEN could produce various invisible adversarial perturbations according to different attack settings by using only one simple end-to-end network with three ingenious loss function; (2) UEN could generate general visible adversarial patch patterns to attack the advanced trackers in the real-world; (3) Extensive experiments show that UEN is able to attack many state-of-the-art trackers effectively (e.g. SiamRPN-based networks and DiMP) on popular tracking datasets including OTB100, UAV123, and GOT10K, making online real-time attacks possible. The attack results outperform the introduced baseline in terms of attacking ability and attacking efficiency.

Automated planning with neural network transition models is a two stage approach to solving planning problems with unknown transition models. The first stage of the approach learns the unknown transition model from data as a neural network model, and the second stage of the approach compiles the learned model to either a Mixed-Integer Linear Programming (MILP) model or a Recurrent Neural Network (RNN) model, and optimize it using an off-the-shelf solver. The previous studies have shown that both models have their advantages and disadvantages. Namely, the MILP model can be solved optimally using a branch-and-bound algorithm but has been experimentally shown not to scale well for neural networks with multiple hidden layers. In contrast, the RNN model can be solved effectively using a gradient descent algorithm but can only work under very restrictive assumptions. In this paper, we focus on improving the effectiveness of solving the second stage of the approach by introducing (i) a novel Lagrangian RNN architecture that can model the previously ignored components of the planning problem as Lagrangian functions, and (ii) a novel framework that unifies the MILP and the Lagrangian RNN models such that the weakness of one model is complemented by the strength of the other. Experimentally, we show that our unifying framework significantly outperforms the standalone MILP model by solving 80% more problem instances, and showcase the ability of our unifying framework to find high quality solutions to challenging automated planning problems with unknown transition models.

Second-order optimization methods have the ability to accelerate convergence by modifying the gradient through the curvature matrix. There have been many attempts to use second-order optimization methods for training deep neural networks. In this work, inspired by diagonal approximations and factored approximations such as Kronecker-factored Approximate Curvature (KFAC), we propose a new approximation to the Fisher information matrix (FIM) called Trace-restricted Kronecker-factored Approximate Curvature (TKFAC), which can hold the certain trace relationship between the exact and the approximate FIM. In TKFAC, we decompose each block of the approximate FIM as a Kronecker product of two smaller matrices and scaled by a coefficient related to trace. We theoretically analyze TKFAC's approximation error and give an upper bound of it. We also propose a new damping technique for TKFAC on convolutional neural networks to maintain the superiority of second-order optimization methods during training. Experiments show that our method has better performance compared with several state-of-the-art algorithms on some deep network architectures.

Generative models can be trained to emulate complex empirical data, but are they useful to make predictions in the context of previously unobserved environments? An intuitive idea to promote such extrapolation capabilities is to have the architecture of such model reflect a causal graph of the true data generating process, such that one can intervene on each node independently of the others. However, the nodes of this graph are usually unobserved, leading to overparameterization and lack of identifiability of the causal structure. We develop a theoretical framework to address this challenging situation by defining a weaker form of identifiability, based on the principle of independence of mechanisms. We demonstrate on toy examples that classical stochastic gradient descent can hinder the model's extrapolation capabilities, suggesting independence of mechanisms should be enforced explicitly during training. Experiments on deep generative models trained on real world data support these insights and illustrate how the extrapolation capabilities of such models can be leveraged.

Text generation tasks, including translation, summarization, language models, and etc. see rapid growth during recent years. Despite the remarkable achievements, the repetition problem has been observed in nearly all text generation models undermining the generation performance extensively. To solve the repetition problem, many methods have been proposed, but there is no existing theoretical analysis to show why this problem happens and how it is resolved. In this paper, we propose a new framework for theoretical analysis for the repetition problem. We first define the Average Repetition Probability (ARP) to characterize the repetition problem quantitatively. Then, we conduct an extensive analysis of the Markov generation model and derive several upper bounds of the average repetition probability with intuitive understanding. We show that most of the existing methods are essentially minimizing the upper bounds explicitly or implicitly. Grounded on our theory, we show that the repetition problem is, unfortunately, caused by the traits of our language itself. One major reason is attributed to the fact that there exist too many words predicting the same word as the subsequent word with high probability. Consequently, it is easy to go back to that word and form repetitions and we dub it as the high inflow problem. Furthermore, we extend our analysis to broader generation models by deriving a concentration bound of the average repetition probability for a general generation model. Finally, based on the theoretical upper bounds, we propose a novel rebalanced encoding approach to alleviate the high inflow problem and thus reducing the upper bound. The experimental results show that our theoretical framework is applicable in general generation models and our proposed rebalanced encoding approach alleviates the repetition problem significantly in both the translation task and the language modeling task. The source code of this paper can be obtained from https://github.com/fuzihaofzh/repetition-problem-nlg.