End-to-end speech translation relies on data that pair source-language speech inputs with corresponding translations into a target language. Such data are notoriously scarce, making synthetic data augmentation by back-translation or knowledge distillation a necessary ingredient of end-to-end training. In this paper, we present a novel approach to data augmentation that leverages audio alignments, linguistic properties, and translation. First, we augment a transcription by sampling from a suffix memory that stores text and audio data. Second, we translate the augmented transcript. Finally, we recombine concatenated audio segments and the generated translation. Besides training an MT-system, we only use basic off-the-shelf components without fine-tuning. While having similar resource demands as knowledge distillation, adding our method delivers consistent improvements of up to 0.9 and 1.1 BLEU points on five language pairs on CoVoST 2 and on two language pairs on Europarl-ST, respectively.
We aim to improve the performance of regressing hand keypoints and segmenting pixel-level hand masks under new imaging conditions (e.g., outdoors) when we only have labeled images taken under very different conditions (e.g., indoors). In the real world, it is important that the model trained for both tasks works under various imaging conditions. However, their variation covered by existing labeled hand datasets is limited. Thus, it is necessary to adapt the model trained on the labeled images (source) to unlabeled images (target) with unseen imaging conditions. While self-training domain adaptation methods (i.e., learning from the unlabeled target images in a self-supervised manner) have been developed for both tasks, their training may degrade performance when the predictions on the target images are noisy. To avoid this, it is crucial to assign a low importance (confidence) weight to the noisy predictions during self-training. In this paper, we propose to utilize the divergence of two predictions to estimate the confidence of the target image for both tasks. These predictions are given from two separate networks, and their divergence helps identify the noisy predictions. To integrate our proposed confidence estimation into self-training, we propose a teacher-student framework where the two networks (teachers) provide supervision to a network (student) for self-training, and the teachers are learned from the student by knowledge distillation. Our experiments show its superiority over state-of-the-art methods in adaptation settings with different lighting, grasping objects, backgrounds, and camera viewpoints. Our method improves by 4% the multi-task score on HO3D compared to the latest adversarial adaptation method. We also validate our method on Ego4D, egocentric videos with rapid changes in imaging conditions outdoors.
State-of-the-art distillation methods are mainly based on distilling deep features from intermediate layers, while the significance of logit distillation is greatly overlooked. To provide a novel viewpoint to study logit distillation, we reformulate the classical KD loss into two parts, i.e., target class knowledge distillation (TCKD) and non-target class knowledge distillation (NCKD). We empirically investigate and prove the effects of the two parts: TCKD transfers knowledge concerning the "difficulty" of training samples, while NCKD is the prominent reason why logit distillation works. More importantly, we reveal that the classical KD loss is a coupled formulation, which (1) suppresses the effectiveness of NCKD and (2) limits the flexibility to balance these two parts. To address these issues, we present Decoupled Knowledge Distillation (DKD), enabling TCKD and NCKD to play their roles more efficiently and flexibly. Compared with complex feature-based methods, our DKD achieves comparable or even better results and has better training efficiency on CIFAR-100, ImageNet, and MS-COCO datasets for image classification and object detection tasks. This paper proves the great potential of logit distillation, and we hope it will be helpful for future research. The code is available at https://github.com/megvii-research/mdistiller.
With the development of deep convolutional neural networks, medical image segmentation has achieved a series of breakthroughs in recent years. However, the high-performance convolutional neural networks always mean numerous parameters and high computation costs, which will hinder the applications in clinical scenarios. Meanwhile, the scarceness of large-scale annotated medical image datasets further impedes the application of high-performance networks. To tackle these problems, we propose Graph Flow, a comprehensive knowledge distillation framework, for both network-efficiency and annotation-efficiency medical image segmentation. Specifically, our core Graph Flow Distillation transfer the essence of cross-layer variations from a well-trained cumbersome teacher network to a non-trained compact student network. In addition, an unsupervised Paraphraser Module is integrated to purify the knowledge of the teacher network, which is also beneficial for the stabilization of training procedure. Furthermore, we build a unified distillation framework by integrating the adversarial distillation and the vanilla logits distillation, which can further refine the final predictions of the compact network. With different teacher networks (conventional convolutional architecture or prevalent transformer architecture) and student networks, we conduct extensive experiments on four medical image datasets with different modalities (Gastric Cancer, Synapse, BUSI, and CVC-ClinicDB).We demonstrate the prominent ability of our method which achieves competitive performance on these datasets. Moreover, we demonstrate the effectiveness of our Graph Flow through a novel semi-supervised paradigm for dual efficient medical image segmentation. Our code will be available at Graph Flow.
Vision transformers (ViTs) have gained popularity recently. Even without customized image operators such as convolutions, ViTs can yield competitive performance when properly trained on massive data. However, the computational overhead of ViTs remains prohibitive, due to stacking multi-head self-attention modules and else. Compared to the vast literature and prevailing success in compressing convolutional neural networks, the study of Vision Transformer compression has also just emerged, and existing works focused on one or two aspects of compression. This paper proposes a unified ViT compression framework that seamlessly assembles three effective techniques: pruning, layer skipping, and knowledge distillation. We formulate a budget-constrained, end-to-end optimization framework, targeting jointly learning model weights, layer-wise pruning ratios/masks, and skip configurations, under a distillation loss. The optimization problem is then solved using the primal-dual algorithm. Experiments are conducted with several ViT variants, e.g. DeiT and T2T-ViT backbones on the ImageNet dataset, and our approach consistently outperforms recent competitors. For example, DeiT-Tiny can be trimmed down to 50\% of the original FLOPs almost without losing accuracy. Codes are available online:~\url{https://github.com/VITA-Group/UVC}.
Continually learning to segment more and more types of image regions is a desired capability for many intelligent systems. However, such continual semantic segmentation suffers from the same catastrophic forgetting issue as in continual classification learning. While multiple knowledge distillation strategies originally for continual classification have been well adapted to continual semantic segmentation, they only consider transferring old knowledge based on the outputs from one or more layers of deep fully convolutional networks. Different from existing solutions, this study proposes to transfer a new type of information relevant to knowledge, i.e. the relationships between elements (Eg. pixels or small local regions) within each image which can capture both within-class and between-class knowledge. The relationship information can be effectively obtained from the self-attention maps in a Transformer-style segmentation model. Considering that pixels belonging to the same class in each image often share similar visual properties, a class-specific region pooling is applied to provide more efficient relationship information for knowledge transfer. Extensive evaluations on multiple public benchmarks support that the proposed self-attention transfer method can further effectively alleviate the catastrophic forgetting issue, and its flexible combination with one or more widely adopted strategies significantly outperforms state-of-the-art solutions.
Knowledge distillation is normally used to compress a big network, or teacher, onto a smaller one, the student, by training it to match its outputs. Recently, some works have shown that robustness against adversarial attacks can also be distilled effectively to achieve good rates of robustness on mobile-friendly models. In this work, however, we take a different point of view, and show that knowledge distillation can be used directly to boost the performance of state-of-the-art models in adversarial robustness. In this sense, we present a thorough analysis and provide general guidelines to distill knowledge from a robust teacher and boost the clean and adversarial performance of a student model even further. To that end, we present Adversarial Knowledge Distillation (AKD), a new framework to improve a model's robust performance, consisting on adversarially training a student on a mixture of the original labels and the teacher outputs. Through carefully controlled ablation studies, we show that using early-stopping, model ensembles and weak adversarial training are key techniques to maximize performance of the student, and show that these insights generalize across different robust distillation techniques. Finally, we provide insights on the effect of robust knowledge distillation on the dynamics of the student network, and show that AKD mostly improves the calibration of the network and modify its training dynamics on samples that the model finds difficult to learn, or even memorize.
Data augmentation has been proved effective in training deep models. Existing data augmentation methods tackle the fine-grained problem by blending image pairs and fusing corresponding labels according to the statistics of mixed pixels, which produces additional noise harmful to the performance of networks. Motivated by this, we present a simple yet effective cross ensemble knowledge distillation (CEKD) model for fine-grained feature learning. We innovatively propose a cross distillation module to provide additional supervision to alleviate the noise problem, and propose a collaborative ensemble module to overcome the target conflict problem. The proposed model can be trained in an end-to-end manner, and only requires image-level label supervision. Extensive experiments on widely used fine-grained benchmarks demonstrate the effectiveness of our proposed model. Specifically, with the backbone of ResNet-101, CEKD obtains the accuracy of 89.59%, 95.96% and 94.56% in three datasets respectively, outperforming state-of-the-art API-Net by 0.99%, 1.06% and 1.16%.
Dealing with missing modalities in Magnetic Resonance Imaging (MRI) and overcoming their negative repercussions is considered a hurdle in biomedical imaging. The combination of a specified set of modalities, which is selected depending on the scenario and anatomical part being scanned, will provide medical practitioners with full information about the region of interest in the human body, hence the missing MRI sequences should be reimbursed. The compensation of the adverse impact of losing useful information owing to the lack of one or more modalities is a well-known challenge in the field of computer vision, particularly for medical image processing tasks including tumour segmentation, tissue classification, and image generation. Various approaches have been developed over time to mitigate this problem's negative implications and this literature review goes through a significant number of the networks that seek to do so. The approaches reviewed in this work are reviewed in detail, including earlier techniques such as synthesis methods as well as later approaches that deploy deep learning, such as common latent space models, knowledge distillation networks, mutual information maximization, and generative adversarial networks (GANs). This work discusses the most important approaches that have been offered at the time of this writing, examining the novelty, strength, and weakness of each one. Furthermore, the most commonly used MRI datasets are highlighted and described. The main goal of this research is to offer a performance evaluation of missing modality compensating networks, as well as to outline future strategies for dealing with this issue.
Contrast-enhanced T1 (T1ce) is one of the most essential magnetic resonance imaging (MRI) modalities for diagnosing and analyzing brain tumors, especially gliomas. In clinical practice, common MRI modalities such as T1, T2, and fluid attenuation inversion recovery are relatively easy to access while T1ce is more challenging considering the additional cost and potential risk of allergies to the contrast agent. Therefore, it is of great clinical necessity to develop a method to synthesize T1ce from other common modalities. Current paired image translation methods typically have the issue of requiring a large amount of paired data and do not focus on specific regions of interest, e.g., the tumor region, in the synthesization process. To address these issues, we propose a Difficulty-perceived common-to-T1ce Semi-Supervised multimodal MRI Synthesis network (DS3-Net), involving both paired and unpaired data together with dual-level knowledge distillation. DS3-Net predicts a difficulty map to progressively promote the synthesis task. Specifically, a pixelwise constraint and a patchwise contrastive constraint are guided by the predicted difficulty map. Through extensive experiments on the publiclyavailable BraTS2020 dataset, DS3-Net outperforms its supervised counterpart in each respect. Furthermore, with only 5% paired data, the proposed DS3-Net achieves competitive performance with state-of-theart image translation methods utilizing 100% paired data, delivering an average SSIM of 0.8947 and an average PSNR of 23.60.
Audio classification is an active research area with a wide range of applications. Over the past decade, convolutional neural networks (CNNs) have been the de-facto standard building block for end-to-end audio classification models. Recently, neural networks based solely on self-attention mechanisms such as the Audio Spectrogram Transformer (AST) have been shown to outperform CNNs. In this paper, we find an intriguing interaction between the two very different models - CNN and AST models are good teachers for each other. When we use either of them as the teacher and train the other model as the student via knowledge distillation (KD), the performance of the student model noticeably improves, and in many cases, is better than the teacher model. In our experiments with this CNN/Transformer Cross-Model Knowledge Distillation (CMKD) method we achieve new state-of-the-art performance on FSD50K, AudioSet, and ESC-50.
The recent large-scale vision-language pre-training (VLP) of dual-stream architectures (e.g., CLIP) with a tremendous amount of image-text pair data, has shown its superiority on various multimodal alignment tasks. Despite its success, the resulting models are not capable of multimodal generative tasks due to the weak text encoder. To tackle this problem, we propose to augment the dual-stream VLP model with a textual pre-trained language model (PLM) via vision-language knowledge distillation (VLKD), enabling the capability for multimodal generation. VLKD is pretty data- and computation-efficient compared to the pre-training from scratch. Experimental results show that the resulting model has strong zero-shot performance on multimodal generation tasks, such as open-ended visual question answering and image captioning. For example, it achieves 44.5% zero-shot accuracy on the VQAv2 dataset, surpassing the previous state-of-the-art zero-shot model with $7\times$ fewer parameters. Furthermore, the original textual language understanding and generation ability of the PLM is maintained after VLKD, which makes our model versatile for both multimodal and unimodal tasks.
Remarkable achievements have been attained with Generative Adversarial Networks (GANs) in image-to-image translation. However, due to a tremendous amount of parameters, state-of-the-art GANs usually suffer from low efficiency and bulky memory usage. To tackle this challenge, firstly, this paper investigates GANs performance from a frequency perspective. The results show that GANs, especially small GANs lack the ability to generate high-quality high frequency information. To address this problem, we propose a novel knowledge distillation method referred to as wavelet knowledge distillation. Instead of directly distilling the generated images of teachers, wavelet knowledge distillation first decomposes the images into different frequency bands with discrete wavelet transformation and then only distills the high frequency bands. As a result, the student GAN can pay more attention to its learning on high frequency bands. Experiments demonstrate that our method leads to 7.08 times compression and 6.80 times acceleration on CycleGAN with almost no performance drop. Additionally, we have studied the relation between discriminators and generators which shows that the compression of discriminators can promote the performance of compressed generators.
In this paper, we focus on a new and challenging decentralized machine learning paradigm in which there are continuous inflows of data to be addressed and the data are stored in multiple repositories. We initiate the study of data decentralized class-incremental learning (DCIL) by making the following contributions. Firstly, we formulate the DCIL problem and develop the experimental protocol. Secondly, we introduce a paradigm to create a basic decentralized counterpart of typical (centralized) class-incremental learning approaches, and as a result, establish a benchmark for the DCIL study. Thirdly, we further propose a Decentralized Composite knowledge Incremental Distillation framework (DCID) to transfer knowledge from historical models and multiple local sites to the general model continually. DCID consists of three main components namely local class-incremental learning, collaborated knowledge distillation among local models, and aggregated knowledge distillation from local models to the general one. We comprehensively investigate our DCID framework by using different implementations of the three components. Extensive experimental results demonstrate the effectiveness of our DCID framework. The codes of the baseline methods and the proposed DCIL will be released at https://github.com/zxxxxh/DCIL.
Neural Machine Translation(NMT) models are usually trained via unidirectional decoder which corresponds to optimizing one-step-ahead prediction. However, this kind of unidirectional decoding framework may incline to focus on local structure rather than global coherence. To alleviate this problem, we propose a novel method, Self-Knowledge Distillation with Bidirectional Decoder for Neural Machine Translation(SBD-NMT). We deploy a backward decoder which can act as an effective regularization method to the forward decoder. By leveraging the backward decoder's information about the longer-term future, distilling knowledge learned in the backward decoder can encourage auto-regressive NMT models to plan ahead. Experiments show that our method is significantly better than the strong Transformer baselines on multiple machine translation data sets.
Temporal Graph Neural Networks (TGNNs) are powerful models to capture temporal, structural, and contextual information on temporal graphs. The generated temporal node embeddings outperform other methods in many downstream tasks. Real-world applications require high performance inference on real-time streaming dynamic graphs. However, these models usually rely on complex attention mechanisms to capture relationships between temporal neighbors. In addition, maintaining vertex memory suffers from intrinsic temporal data dependency that hinders task-level parallelism, making it inefficient on general-purpose processors. In this work, we present a novel model-architecture co-design for inference in memory-based TGNNs on FPGAs. The key modeling optimizations we propose include a light-weight method to compute attention scores and a related temporal neighbor pruning strategy to further reduce computation and memory accesses. These are holistically coupled with key hardware optimizations that leverage FPGA hardware. We replace the temporal sampler with an on-chip FIFO based hardware sampler and the time encoder with a look-up-table. We train our simplified models using knowledge distillation to ensure similar accuracy vis-\'a-vis the original model. Taking advantage of the model optimizations, we propose a principled hardware architecture using batching, pipelining, and prefetching techniques to further improve the performance. We also propose a hardware mechanism to ensure the chronological vertex updating without sacrificing the computation parallelism. We evaluate the performance of the proposed hardware accelerator on three real-world datasets.
Real-world object detection models should be cheap and accurate. Knowledge distillation (KD) can boost the accuracy of a small, cheap detection model by leveraging useful information from a larger teacher model. However, a key challenge is identifying the most informative features produced by the teacher for distillation. In this work, we show that only a very small fraction of features within a ground-truth bounding box are responsible for a teacher's high detection performance. Based on this, we propose Prediction-Guided Distillation (PGD), which focuses distillation on these key predictive regions of the teacher and yields considerable gains in performance over many existing KD baselines. In addition, we propose an adaptive weighting scheme over the key regions to smooth out their influence and achieve even better performance. Our proposed approach outperforms current state-of-the-art KD baselines on a variety of advanced one-stage detection architectures. Specifically, on the COCO dataset, our method achieves between +3.1% and +4.6% AP improvement using ResNet-101 and ResNet-50 as the teacher and student backbones, respectively. On the CrowdHuman dataset, we achieve +3.2% and +2.0% improvements in MR and AP, also using these backbones. Our code is available at https://github.com/ChenhongyiYang/PGD.
In this work, we study the continual semantic segmentation problem, where the deep neural networks are required to incorporate new classes continually without catastrophic forgetting. We propose to use a structural re-parameterization mechanism, named representation compensation (RC) module, to decouple the representation learning of both old and new knowledge. The RC module consists of two dynamically evolved branches with one frozen and one trainable. Besides, we design a pooled cube knowledge distillation strategy on both spatial and channel dimensions to further enhance the plasticity and stability of the model. We conduct experiments on two challenging continual semantic segmentation scenarios, continual class segmentation and continual domain segmentation. Without any extra computational overhead and parameters during inference, our method outperforms state-of-the-art performance. The code is available at \url{https://github.com/zhangchbin/RCIL}.
Image-to-image translation models are shown to be vulnerable to the Membership Inference Attack (MIA), in which the adversary's goal is to identify whether a sample is used to train the model or not. With daily increasing applications based on image-to-image translation models, it is crucial to protect the privacy of these models against MIAs. We propose adversarial knowledge distillation (AKD) as a defense method against MIAs for image-to-image translation models. The proposed method protects the privacy of the training samples by improving the generalizability of the model. We conduct experiments on the image-to-image translation models and show that AKD achieves the state-of-the-art utility-privacy tradeoff by reducing the attack performance up to 38.9% compared with the regular training model at the cost of a slight drop in the quality of the generated output images. The experimental results also indicate that the models trained by AKD generalize better than the regular training models. Furthermore, compared with existing defense methods, the results show that at the same privacy protection level, image translation models trained by AKD generate outputs with higher quality; while at the same quality of outputs, AKD enhances the privacy protection over 30%.
Large-scale pre-training has been proven to be crucial for various computer vision tasks. However, with the increase of pre-training data amount, model architecture amount, and the private/inaccessible data, it is not very efficient or possible to pre-train all the model architectures on large-scale datasets. In this work, we investigate an alternative strategy for pre-training, namely Knowledge Distillation as Efficient Pre-training (KDEP), aiming to efficiently transfer the learned feature representation from existing pre-trained models to new student models for future downstream tasks. We observe that existing Knowledge Distillation (KD) methods are unsuitable towards pre-training since they normally distill the logits that are going to be discarded when transferred to downstream tasks. To resolve this problem, we propose a feature-based KD method with non-parametric feature dimension aligning. Notably, our method performs comparably with supervised pre-training counterparts in 3 downstream tasks and 9 downstream datasets requiring 10x less data and 5x less pre-training time. Code is available at https://github.com/CVMI-Lab/KDEP.