DearKD: Data-Efficient Early Knowledge Distillation for Vision Transformers

Transformers are successfully applied to computer vision due to their powerful modeling capacity with self-attention. However, the excellent performance of transformers heavily depends on enormous training images. Thus, a data-efficient transformer solution is urgently needed. In this work, we propose an early knowledge distillation framework, which is termed as DearKD, to improve the data efficiency required by transformers. Our DearKD is a two-stage framework that first distills the inductive biases from the early intermediate layers of a CNN and then gives the transformer full play by training without distillation. Further, our DearKD can be readily applied to the extreme data-free case where no real images are available. In this case, we propose a boundary-preserving intra-divergence loss based on DeepInversion to further close the performance gap against the full-data counterpart. Extensive experiments on ImageNet, partial ImageNet, data-free setting and other downstream tasks prove the superiority of DearKD over its baselines and state-of-the-art methods.

ViTPose: Simple Vision Transformer Baselines for Human Pose Estimation

Recently, customized vision transformers have been adapted for human pose estimation and have achieved superior performance with elaborate structures. However, it is still unclear whether plain vision transformers can facilitate pose estimation. In this paper, we take the first step toward answering the question by employing a plain and non-hierarchical vision transformer together with simple deconvolution decoders termed ViTPose for human pose estimation. We demonstrate that a plain vision transformer with MAE pretraining can obtain superior performance after finetuning on human pose estimation datasets. ViTPose has good scalability with respect to model size and flexibility regarding input resolution and token number. Moreover, it can be easily pretrained using the unlabeled pose data without the need for large-scale upstream ImageNet data. Our biggest ViTPose model based on the ViTAE-G backbone with 1 billion parameters obtains the best 80.9 mAP on the MS COCO test-dev set, while the ensemble models further set a new state-of-the-art for human pose estimation, i.e., 81.1 mAP. The source code and models will be released at https://github.com/ViTAE-Transformer/ViTPose.

Neural Maximum A Posteriori Estimation on Unpaired Data for Motion Deblurring

Real-world dynamic scene deblurring has long been a challenging task since paired blurry-sharp training data is unavailable. Conventional Maximum A Posteriori estimation and deep learning-based deblurring methods are restricted by handcrafted priors and synthetic blurry-sharp training pairs respectively, thereby failing to generalize to real dynamic blurriness. To this end, we propose a Neural Maximum A Posteriori (NeurMAP) estimation framework for training neural networks to recover blind motion information and sharp content from unpaired data. The proposed NeruMAP consists of a motion estimation network and a deblurring network which are trained jointly to model the (re)blurring process (i.e. likelihood function). Meanwhile, the motion estimation network is trained to explore the motion information in images by applying implicit dynamic motion prior, and in return enforces the deblurring network training (i.e. providing sharp image prior). The proposed NeurMAP is an orthogonal approach to existing deblurring neural networks, and is the first framework that enables training image deblurring networks on unpaired datasets. Experiments demonstrate our superiority on both quantitative metrics and visual quality over state-of-the-art methods. Codes are available on https://github.com/yjzhang96/NeurMAP-deblur.

BLISS: Robust Sequence-to-Sequence Learning via Self-Supervised Input Representation

Data augmentations (DA) are the cores to achieving robust sequence-to-sequence learning on various natural language processing (NLP) tasks. However, most of the DA approaches force the decoder to make predictions conditioned on the perturbed input representation, underutilizing supervised information provided by perturbed input. In this work, we propose a framework-level robust sequence-to-sequence learning approach, named BLISS, via self-supervised input representation, which has the great potential to complement the data-level augmentation approaches. The key idea is to supervise the sequence-to-sequence framework with both the \textit{supervised} ("input$\rightarrow$output") and \textit{self-supervised} ("perturbed input$\rightarrow$input") information. We conduct comprehensive experiments to validate the effectiveness of BLISS on various tasks, including machine translation, grammatical error correction, and text summarization. The results show that BLISS outperforms significantly the vanilla Transformer and consistently works well across tasks than the other five contrastive baselines. Extensive analyses reveal that BLISS learns robust representations and rich linguistic knowledge, confirming our claim. Source code will be released upon publication.

Source-Free Domain Adaptation via Distribution Estimation

Domain Adaptation aims to transfer the knowledge learned from a labeled source domain to an unlabeled target domain whose data distributions are different. However, the training data in source domain required by most of the existing methods is usually unavailable in real-world applications due to privacy preserving policies. Recently, Source-Free Domain Adaptation (SFDA) has drawn much attention, which tries to tackle domain adaptation problem without using source data. In this work, we propose a novel framework called SFDA-DE to address SFDA task via source Distribution Estimation. Firstly, we produce robust pseudo-labels for target data with spherical k-means clustering, whose initial class centers are the weight vectors (anchors) learned by the classifier of pretrained model. Furthermore, we propose to estimate the class-conditioned feature distribution of source domain by exploiting target data and corresponding anchors. Finally, we sample surrogate features from the estimated distribution, which are then utilized to align two domains by minimizing a contrastive adaptation loss function. Extensive experiments show that the proposed method achieves state-of-the-art performance on multiple DA benchmarks, and even outperforms traditional DA methods which require plenty of source data.

A Model-Agnostic Data Manipulation Method for Persona-based Dialogue Generation

Towards building intelligent dialogue agents, there has been a growing interest in introducing explicit personas in generation models. However, with limited persona-based dialogue data at hand, it may be difficult to train a dialogue generation model well. We point out that the data challenges of this generation task lie in two aspects: first, it is expensive to scale up current persona-based dialogue datasets; second, each data sample in this task is more complex to learn with than conventional dialogue data. To alleviate the above data issues, we propose a data manipulation method, which is model-agnostic to be packed with any persona-based dialogue generation model to improve its performance. The original training samples will first be distilled and thus expected to be fitted more easily. Next, we show various effective ways that can diversify such easier distilled data. A given base model will then be trained via the constructed data curricula, i.e. first on augmented distilled samples and then on original ones. Experiments illustrate the superiority of our method with two strong base dialogue models (Transformer encoder-decoder and GPT2).

Neural Collapse Inspired Attraction-Repulsion-Balanced Loss for Imbalanced Learning

Class imbalance distribution widely exists in real-world engineering. However, the mainstream optimization algorithms that seek to minimize error will trap the deep learning model in sub-optimums when facing extreme class imbalance. It seriously harms the classification precision, especially on the minor classes. The essential reason is that the gradients of the classifier weights are imbalanced among the components from different classes. In this paper, we propose Attraction-Repulsion-Balanced Loss (ARB-Loss) to balance the different components of the gradients. We perform experiments on the large-scale classification and segmentation datasets and our ARB-Loss can achieve state-of-the-art performance via only one-stage training instead of 2-stage learning like nowadays SOTA works.

VSA: Learning Varied-Size Window Attention in Vision Transformers

Attention within windows has been widely explored in vision transformers to balance the performance, computation complexity, and memory footprint. However, current models adopt a hand-crafted fixed-size window design, which restricts their capacity of modeling long-term dependencies and adapting to objects of different sizes. To address this drawback, we propose \textbf{V}aried-\textbf{S}ize Window \textbf{A}ttention (VSA) to learn adaptive window configurations from data. Specifically, based on the tokens within each default window, VSA employs a window regression module to predict the size and location of the target window, i.e., the attention area where the key and value tokens are sampled. By adopting VSA independently for each attention head, it can model long-term dependencies, capture rich context from diverse windows, and promote information exchange among overlapped windows. VSA is an easy-to-implement module that can replace the window attention in state-of-the-art representative models with minor modifications and negligible extra computational cost while improving their performance by a large margin, e.g., 1.1\% for Swin-T on ImageNet classification. In addition, the performance gain increases when using larger images for training and test. Experimental results on more downstream tasks, including object detection, instance segmentation, and semantic segmentation, further demonstrate the superiority of VSA over the vanilla window attention in dealing with objects of different sizes. The code will be released https://github.com/ViTAE-Transformer/ViTAE-VSA.

Bridging Cross-Lingual Gaps During Leveraging the Multilingual Sequence-to-Sequence Pretraining for Text Generation

For multilingual sequence-to-sequence pretrained language models (multilingual Seq2Seq PLMs), e.g. mBART, the self-supervised pretraining task is trained on a wide range of monolingual languages, e.g. 25 languages from commoncrawl, while the downstream cross-lingual tasks generally progress on a bilingual language subset, e.g. English-German, making there exists the cross-lingual data discrepancy, namely \textit{domain discrepancy}, and cross-lingual learning objective discrepancy, namely \textit{task discrepancy}, between the pretrain and finetune stages. To bridge the above cross-lingual domain and task gaps, we extend the vanilla pretrain-finetune pipeline with extra code-switching restore task. Specifically, the first stage employs the self-supervised code-switching restore task as a pretext task, allowing the multilingual Seq2Seq PLM to acquire some in-domain alignment information. And for the second stage, we continuously fine-tune the model on labeled data normally. Experiments on a variety of cross-lingual NLG tasks, including 12 bilingual translation tasks, 36 zero-shot translation tasks, and cross-lingual summarization tasks show our model outperforms the strong baseline mBART consistently. Comprehensive analyses indicate our approach could narrow the cross-lingual sentence representation distance and improve low-frequency word translation with trivial computational cost.