Vision Transformers (ViTs) have achieved comparable or superior performance than Convolutional Neural Networks (CNNs) in computer vision. This empirical breakthrough is even more remarkable since, in contrast to CNNs, ViTs do not embed any visual inductive bias of spatial locality. Yet, recent works have shown that while minimizing their training loss, ViTs specifically learn spatially localized patterns. This raises a central question: how do ViTs learn these patterns by solely minimizing their training loss using gradient-based methods from random initialization? In this paper, we provide some theoretical justification of this phenomenon. We propose a spatially structured dataset and a simplified ViT model. In this model, the attention matrix solely depends on the positional encodings. We call this mechanism the positional attention mechanism. On the theoretical side, we consider a binary classification task and show that while the learning problem admits multiple solutions that generalize, our model implicitly learns the spatial structure of the dataset while generalizing: we call this phenomenon patch association. We prove that patch association helps to sample-efficiently transfer to downstream datasets that share the same structure as the pre-training one but differ in the features. Lastly, we empirically verify that a ViT with positional attention performs similarly to the original one on CIFAR-10/100, SVHN and ImageNet.

Current Explainable AI (ExAI) methods, especially in the NLP field, are conducted on various datasets by employing different metrics to evaluate several aspects. The lack of a common evaluation framework is hindering the progress tracking of such methods and their wider adoption. In this work, inspired by offline information retrieval, we propose different metrics and techniques to evaluate the explainability of SA models from two angles. First, we evaluate the strength of the extracted "rationales" in faithfully explaining the predicted outcome. Second, we measure the agreement between ExAI methods and human judgment on a homegrown dataset1 to reflect on the rationales plausibility. Our conducted experiments comprise four dimensions: (1) the underlying architectures of SA models, (2) the approach followed by the ExAI method, (3) the reasoning difficulty, and (4) the homogeneity of the ground-truth rationales. We empirically demonstrate that anchors explanations are more aligned with the human judgment and can be more confident in extracting supporting rationales. As can be foreseen, the reasoning complexity of sentiment is shown to thwart ExAI methods from extracting supporting evidence. Moreover, a remarkable discrepancy is discerned between the results of different explainability methods on the various architectures suggesting the need for consolidation to observe enhanced performance. Predominantly, transformers are shown to exhibit better explainability than convolutional and recurrent architectures. Our work paves the way towards designing more interpretable NLP models and enabling a common evaluation ground for their relative strengths and robustness.

Task transfer, transferring knowledge contained in related tasks, holds the promise of reducing the quantity of labeled data required to fine-tune language models. Dialogue understanding encompasses many diverse tasks, yet task transfer has not been thoroughly studied in conversational AI. This work explores conversational task transfer by introducing FETA: a benchmark for few-sample task transfer in open-domain dialogue. FETA contains two underlying sets of conversations upon which there are 10 and 7 tasks annotated, enabling the study of intra-dataset task transfer; task transfer without domain adaptation. We utilize three popular language models and three learning algorithms to analyze the transferability between 132 source-target task pairs and create a baseline for future work. We run experiments in the single- and multi-source settings and report valuable findings, e.g., most performance trends are model-specific, and span extraction and multiple-choice tasks benefit the most from task transfer. In addition to task transfer, FETA can be a valuable resource for future research into the efficiency and generalizability of pre-training datasets and model architectures, as well as for learning settings such as continual and multitask learning.

Despite tremendous progress in natural language processing using deep learning techniques in recent years, sign language production and comprehension has advanced very little. One critical barrier is the lack of largescale datasets available to the public due to the unbearable cost of labeled data generation. Efforts to provide public data for American Sign Language (ASL) comprehension have yielded two datasets, comprising more than thousand video clips. These datasets are large enough to enable a meaningful start to deep learning research on sign languages but are far too small to lead to any solution that can be practically deployed. So far, there is still no suitable dataset for ASL production. We proposed a system that can generate large scale ASL datasets for continuous ASL. It is suitable for general ASL processing and is particularly useful for ASL production. The continuous ASL dataset contains English labeled human articulations in condensed body pose data formats. To better serve the research community, we are releasing the first version of our ASL dataset, which contains 30k sentences, 416k words, a vocabulary of 18k words, in a total of 104 hours. This is the largest continuous sign language dataset published to date in terms of video duration. We also describe a system that can evolve and expand the dataset to incorporate better data processing techniques and more contents when available. It is our hope that the release of this ASL dataset and the sustainable dataset generation system to the public will propel better deep-learning research in ASL natural language processing.

Graph Convolutional Networks (GCNs) suffer from performance degradation when models go deeper. However, earlier works only attributed the performance degradation to over-smoothing. In this paper, we conduct theoretical and experimental analysis to explore the fundamental causes of performance degradation in deep GCNs: over-smoothing and gradient vanishing have a mutually reinforcing effect that causes the performance to deteriorate more quickly in deep GCNs. On the other hand, existing anti-over-smoothing methods all perform full convolutions up to the model depth. They could not well resist the exponential convergence of over-smoothing due to model depth increasing. In this work, we propose a simple yet effective plug-and-play module, SkipNode, to overcome the performance degradation of deep GCNs. It samples graph nodes in each convolutional layer to skip the convolution operation. In this way, both over-smoothing and gradient vanishing can be effectively suppressed since (1) not all nodes perform full convolutions up to the model depth and, (2) the gradient can be directly passed back through ``skipped'' nodes. We provide both theoretical analysis and empirical evaluation to demonstrate the efficacy of SkipNode and its superiority over SOTA baselines.

Code-switching automatic speech recognition becomes one of the most challenging and the most valuable scenarios of automatic speech recognition, due to the code-switching phenomenon between multilingual language and the frequent occurrence of code-switching phenomenon in daily life. The ISCSLP 2022 Chinese-English Code-Switching Automatic Speech Recognition (CSASR) Challenge aims to promote the development of code-switching automatic speech recognition. The ISCSLP 2022 CSASR challenge provided two training sets, TAL_CSASR corpus and MagicData-RAMC corpus, a development and a test set for participants, which are used for CSASR model training and evaluation. Along with the challenge, we also provide the baseline system performance for reference. As a result, more than 40 teams participated in this challenge, and the winner team achieved 16.70% Mixture Error Rate (MER) performance on the test set and has achieved 9.8% MER absolute improvement compared with the baseline system. In this paper, we will describe the datasets, the associated baselines system and the requirements, and summarize the CSASR challenge results and major techniques and tricks used in the submitted systems.

In many modern applications of deep neural networks, where the amount of labeled examples for training is limited, distillation with unlabeled examples [8, 20] has been enormously successful. In this two-stage training paradigm a larger and more sophisticated "teacher model" -- typically non-deployable for the purposes of the application -- is trained to learn from the limited amount of available training data in the first stage. In the second stage, the teacher model is used to generate labels on an unlabeled dataset, which is usually much larger in size than the original dataset the teacher itself was trained on. These labels are then utilized to train the "student model", namely the model which will actually be deployed. Notably, distillation with unlabeled examples is the most commonly used training-paradigm in applications where one finetunes and distills from very large-scale foundational models such as BERT [13] and GPT-3 [5] and, additionally, it can be used to significantly improve distillation on supervised examples only (see e.g.

Most research on natural language processing treats bias as an absolute concept: Based on a (probably complex) algorithmic analysis, a sentence, an article, or a text is classified as biased or not. Given the fact that for humans the question of whether a text is biased can be difficult to answer or is answered contradictory, we ask whether an "absolute bias classification" is a promising goal at all. We see the problem not in the complexity of interpreting language phenomena but in the diversity of sociocultural backgrounds of the readers, which cannot be handled uniformly: To decide whether a text has crossed the proverbial line between non-biased and biased is subjective. By asking "Is text X more [less, equally] biased than text Y?" we propose to analyze a simpler problem, which, by its construction, is rather independent of standpoints, views, or sociocultural aspects. In such a model, bias becomes a preference relation that induces a partial ordering from least biased to most biased texts without requiring a decision on where to draw the line. A prerequisite for this kind of bias model is the ability of humans to perceive relative bias differences in the first place. In our research, we selected a specific type of bias in argumentation, the stance bias, and designed a crowdsourcing study showing that differences in stance bias are perceptible when (light) support is provided through training or visual aid.

Despite substantial progress in 3D object detection, advanced 3D detectors often suffer from heavy computation overheads. To this end, we explore the potential of knowledge distillation (KD) for developing efficient 3D object detectors, focusing on popular pillar- and voxel-based detectors.In the absence of well-developed teacher-student pairs, we first study how to obtain student models with good trade offs between accuracy and efficiency from the perspectives of model compression and input resolution reduction. Then, we build a benchmark to assess existing KD methods developed in the 2D domain for 3D object detection upon six well-constructed teacher-student pairs. Further, we propose an improved KD pipeline incorporating an enhanced logit KD method that performs KD on only a few pivotal positions determined by teacher classification response, and a teacher-guided student model initialization to facilitate transferring teacher model's feature extraction ability to students through weight inheritance. Finally, we conduct extensive experiments on the Waymo dataset. Our best performing model achieves $65.75\%$ LEVEL 2 mAPH, surpassing its teacher model and requiring only $44\%$ of teacher flops. Our most efficient model runs 51 FPS on an NVIDIA A100, which is $2.2\times$ faster than PointPillar with even higher accuracy. Code is available at \url{https://github.com/CVMI-Lab/SparseKD}.

There is an emerging need for predictive models to be trained on-the-fly, since in numerous machine learning applications data are arriving in an online fashion. A critical challenge encountered is that of limited availability of ground truth information (e.g., labels in classification tasks) as new data are observed one-by-one online, while another significant challenge is that of class imbalance. This work introduces the novel Augmented Queues method, which addresses the dual-problem by combining in a synergistic manner online active learning, data augmentation, and a multi-queue memory to maintain separate and balanced queues for each class. We perform an extensive experimental study using image and time-series augmentations, in which we examine the roles of the active learning budget, memory size, imbalance level, and neural network type. We demonstrate two major advantages of Augmented Queues. First, it does not reserve additional memory space as the generation of synthetic data occurs only at training times. Second, learning models have access to more labelled data without the need to increase the active learning budget and / or the original memory size. Learning on-the-fly poses major challenges which, typically, hinder the deployment of learning models. Augmented Queues significantly improves the performance in terms of learning quality and speed. Our code is made publicly available.