PyTorchKR Papers

최신 논문

총 4,526편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2104.05279공식 코드 · Papers with Code

Class-Balanced Distillation for Long-Tailed Visual Recognition

Real-world imagery is often characterized by a significant imbalance of the number of images per class, leading to long-tailed distributions. An effective and simple approach to long-tailed visual recognition is to learn feature representations and a classifier separately, with instance and class-balanced sampling, respectively. In this work, we introduce a new framework, by making the key observation that a feature representation learned with instance sampling is far from optimal in a long-tailed setting. Our main contribution is a new training method, referred to as Class-Balanced Distillation (CBD), that leverages knowledge distillation to enhance feature representations. CBD allows the feature representation to evolve in the second training stage, guided by the teacher learned in the first stage. The second stage uses class-balanced sampling, in order to focus on under-represented classes. This framework can naturally accommodate the usage of multiple teachers, unlocking the information from an ensemble of models to enhance recognition capabilities. Our experiments show that the proposed technique consistently outperforms the state of the art on long-tailed recognition benchmarks such as ImageNet-LT, iNaturalist17 and iNaturalist18.

논문 ID: document/9265295

Head Network Distillation: Splitting Distilled Deep Neural Networks for Resource-Constrained Edge Computing Systems

As the complexity of Deep Neural Network (DNN) models increases, their deployment on mobile devices becomes increasingly challenging, especially in complex vision tasks such as image classification. Many of recent contributions aim either to produce compact models matching the limited computing capabilities of mobile devices or to offload the execution of such burdensome models to a compute-capable device at the network edge - the edge servers. In this paper, we propose to modify the structure and training process of DNN models for complex image classification tasks to achieve in-network compression in the early network layers. Our training process stems from knowledge distillation, a technique that has been traditionally used to build small - student - models mimicking the output of larger - teacher - models. Here, we adopt this idea to obtain aggressive compression while preserving accuracy. Our results demonstrate that our approach is effective for state-of-the-art models trained over complex datasets, and can extend the parameter region in which edge computing is a viable and advantageous option. Additionally, we demonstrate that in many settings of practical interest we reduce the inference time with respect to specialized models such as MobileNet v2 executed at the mobile device, while improving accuracy.

논문 ID: 9614.33560

Distilled Split Deep Neural Networks for Edge-Assisted Real-Time Systems

Offloading the execution of complex Deep Neural Networks (DNNs) models to compute-capable devices at the network edge, that is, edge servers, can significantly reduce capture-to-output delay. However, the communication link between the mobile devices and edge servers can become the bottleneck when channel conditions are poor. We propose a framework to split DNNs for image processing and minimize capture-to-output delay in a wide range of network conditions and computing parameters. The core idea is to split the DNN models into head and tail models, where the two sections are deployed at the mobile device and edge server, respectively. Different from prior literature presenting DNN splitting frameworks, we distill the architecture of the head DNN to reduce its computational complexity and introduce a bottleneck, thus minimizing processing load at the mobile device as well as the amount of wirelessly transferred data. Our results show 98% reduction in used bandwidth and 85% in computation load compared to straightforward splitting.

arXiv:2104.03736공식 코드 · Papers with Code

Towards Enabling Meta-Learning from Target Models

Meta-learning can extract an inductive bias from previous learning experience and assist the training of new tasks. It is often realized through optimizing a meta-model with the evaluation loss of task-specific solvers. Most existing algorithms sample non-overlapping $\mathit{support}$ sets and $\mathit{query}$ sets to train and evaluate the solvers respectively due to simplicity ($\mathcal{S}$/$\mathcal{Q}$ protocol). Different from $\mathcal{S}$/$\mathcal{Q}$ protocol, we can also evaluate a task-specific solver by comparing it to a target model $\mathcal{T}$, which is the optimal model for this task or a model that behaves well enough on this task ($\mathcal{S}$/$\mathcal{T}$ protocol). Although being short of research, $\mathcal{S}$/$\mathcal{T}$ protocol has unique advantages such as offering more informative supervision, but it is computationally expensive. This paper looks into this special evaluation method and takes a step towards putting it into practice. We find that with a small ratio of tasks armed with target models, classic meta-learning algorithms can be improved a lot without consuming many resources. We empirically verify the effectiveness of $\mathcal{S}$/$\mathcal{T}$ protocol in a typical application of meta-learning, $\mathit{i.e.}$, few-shot learning. In detail, after constructing target models by fine-tuning the pre-trained network on those hard tasks, we match the task-specific solvers and target models via knowledge distillation.

arXiv:2104.03597공식 코드 · Papers with Code

GKD: Semi-supervised Graph Knowledge Distillation for Graph-Independent Inference

The increased amount of multi-modal medical data has opened the opportunities to simultaneously process various modalities such as imaging and non-imaging data to gain a comprehensive insight into the disease prediction domain. Recent studies using Graph Convolutional Networks (GCNs) provide novel semi-supervised approaches for integrating heterogeneous modalities while investigating the patients' associations for disease prediction. However, when the meta-data used for graph construction is not available at inference time (e.g., coming from a distinct population), the conventional methods exhibit poor performance. To address this issue, we propose a novel semi-supervised approach named GKD based on knowledge distillation. We train a teacher component that employs the label-propagation algorithm besides a deep neural network to benefit from the graph and non-graph modalities only in the training phase. The teacher component embeds all the available information into the soft pseudo-labels. The soft pseudo-labels are then used to train a deep student network for disease prediction of unseen test data for which the graph modality is unavailable. We perform our experiments on two public datasets for diagnosing Autism spectrum disorder, and Alzheimer's disease, along with a thorough analysis on synthetic multi-modal datasets. According to these experiments, GKD outperforms the previous graph-based deep learning methods in terms of accuracy, AUC, and Macro F1.

arXiv:2104.03164공식 코드 · Papers with Code

Distilling and Transferring Knowledge via cGAN-generated Samples for Image Classification and Regression

Knowledge distillation (KD) has been actively studied for image classification tasks in deep learning, aiming to improve the performance of a student based on the knowledge from a teacher. However, applying KD in image regression with a scalar response variable has been rarely studied, and there exists no KD method applicable to both classification and regression tasks yet. Moreover, existing KD methods often require a practitioner to carefully select or adjust the teacher and student architectures, making these methods less flexible in practice. To address the above problems in a unified way, we propose a comprehensive KD framework based on cGANs, termed cGAN-KD. Fundamentally different from existing KD methods, cGAN-KD distills and transfers knowledge from a teacher model to a student model via cGAN-generated samples. This novel mechanism makes cGAN-KD suitable for both classification and regression tasks, compatible with other KD methods, and insensitive to the teacher and student architectures. An error bound for a student model trained in the cGAN-KD framework is derived in this work, providing a theory for why cGAN-KD is effective as well as guiding the practical implementation of cGAN-KD. Extensive experiments on CIFAR-100 and ImageNet-100 show that we can combine state of the art KD methods with the cGAN-KD framework to yield a new state of the art. Moreover, experiments on Steering Angle and UTKFace demonstrate the effectiveness of cGAN-KD in image regression tasks, where existing KD methods are inapplicable.

arXiv:2104.02244공식 코드 · Papers with Code

Content-Aware GAN Compression

Generative adversarial networks (GANs), e.g., StyleGAN2, play a vital role in various image generation and synthesis tasks, yet their notoriously high computational cost hinders their efficient deployment on edge devices. Directly applying generic compression approaches yields poor results on GANs, which motivates a number of recent GAN compression works. While prior works mainly accelerate conditional GANs, e.g., pix2pix and CycleGAN, compressing state-of-the-art unconditional GANs has rarely been explored and is more challenging. In this paper, we propose novel approaches for unconditional GAN compression. We first introduce effective channel pruning and knowledge distillation schemes specialized for unconditional GANs. We then propose a novel content-aware method to guide the processes of both pruning and distillation. With content-awareness, we can effectively prune channels that are unimportant to the contents of interest, e.g., human faces, and focus our distillation on these regions, which significantly enhances the distillation quality. On StyleGAN2 and SN-GAN, we achieve a substantial improvement over the state-of-the-art compression method. Notably, we reduce the FLOPs of StyleGAN2 by 11x with visually negligible image quality loss compared to the full-size model. More interestingly, when applied to various image manipulation tasks, our compressed model forms a smoother and better disentangled latent manifold, making it more effective for image editing.

arXiv:2104.02096

Compressing Visual-linguistic Model via Knowledge Distillation

Despite exciting progress in pre-training for visual-linguistic (VL) representations, very few aspire to a small VL model. In this paper, we study knowledge distillation (KD) to effectively compress a transformer-based large VL model into a small VL model. The major challenge arises from the inconsistent regional visual tokens extracted from different detectors of Teacher and Student, resulting in the misalignment of hidden representations and attention distributions. To address the problem, we retrain and adapt the Teacher by using the same region proposals from Student's detector while the features are from Teacher's own object detector. With aligned network inputs, the adapted Teacher is capable of transferring the knowledge through the intermediate representations. Specifically, we use the mean square error loss to mimic the attention distribution inside the transformer block and present a token-wise noise contrastive loss to align the hidden state by contrasting with negative representations stored in a sample queue. To this end, we show that our proposed distillation significantly improves the performance of small VL models on image captioning and visual question answering tasks. It reaches 120.8 in CIDEr score on COCO captioning, an improvement of 5.1 over its non-distilled counterpart; and an accuracy of 69.8 on VQA 2.0, a 0.8 gain from the baseline. Our extensive experiments and ablations confirm the effectiveness of VL distillation in both pre-training and fine-tuning stages.

arXiv:2104.00352

Decentralized and Model-Free Federated Learning: Consensus-Based Distillation in Function Space

This paper proposes a fully decentralized federated learning (FL) scheme for Internet of Everything (IoE) devices that are connected via multi-hop networks. Because FL algorithms hardly converge the parameters of machine learning (ML) models, this paper focuses on the convergence of ML models in function spaces. Considering that the representative loss functions of ML tasks e.g, mean squared error (MSE) and Kullback-Leibler (KL) divergence, are convex functionals, algorithms that directly update functions in function spaces could converge to the optimal solution. The key concept of this paper is to tailor a consensus-based optimization algorithm to work in the function space and achieve the global optimum in a distributed manner. This paper first analyzes the convergence of the proposed algorithm in a function space, which is referred to as a meta-algorithm, and shows that the spectral graph theory can be applied to the function space in a manner similar to that of numerical vectors. Then, consensus-based multi-hop federated distillation (CMFD) is developed for a neural network (NN) to implement the meta-algorithm. CMFD leverages knowledge distillation to realize function aggregation among adjacent devices without parameter averaging. An advantage of CMFD is that it works even with different NN models among the distributed learners. Although CMFD does not perfectly reflect the behavior of the meta-algorithm, the discussion of the meta-algorithm's convergence property promotes an intuitive understanding of CMFD, and simulation evaluations show that NN models converge using CMFD for several tasks. The simulation results also show that CMFD achieves higher accuracy than parameter aggregation for weakly connected networks, and CMFD is more stable than parameter aggregation methods.

arXiv:2104.00233공식 코드 · Papers with Code

Unsupervised Domain Expansion for Visual Categorization

Expanding visual categorization into a novel domain without the need of extra annotation has been a long-term interest for multimedia intelligence. Previously, this challenge has been approached by unsupervised domain adaptation (UDA). Given labeled data from a source domain and unlabeled data from a target domain, UDA seeks for a deep representation that is both discriminative and domain-invariant. While UDA focuses on the target domain, we argue that the performance on both source and target domains matters, as in practice which domain a test example comes from is unknown. In this paper we extend UDA by proposing a new task called unsupervised domain expansion (UDE), which aims to adapt a deep model for the target domain with its unlabeled data, meanwhile maintaining the model's performance on the source domain. We propose Knowledge Distillation Domain Expansion (KDDE) as a general method for the UDE task. Its domain-adaptation module can be instantiated with any existing model. We develop a knowledge distillation based learning mechanism, enabling KDDE to optimize a single objective wherein the source and target domains are equally treated. Extensive experiments on two major benchmarks, i.e., Office-Home and DomainNet, show that KDDE compares favorably against four competitive baselines, i.e., DDC, DANN, DAAN, and CDAN, for both UDA and UDE tasks. Our study also reveals that the current UDA models improve their performance on the target domain at the cost of noticeable performance loss on the source domain.

arXiv:2104.00676

Is Label Smoothing Truly Incompatible with Knowledge Distillation: An Empirical Study

This work aims to empirically clarify a recently discovered perspective that label smoothing is incompatible with knowledge distillation. We begin by introducing the motivation behind on how this incompatibility is raised, i.e., label smoothing erases relative information between teacher logits. We provide a novel connection on how label smoothing affects distributions of semantically similar and dissimilar classes. Then we propose a metric to quantitatively measure the degree of erased information in sample's representation. After that, we study its one-sidedness and imperfection of the incompatibility view through massive analyses, visualizations and comprehensive experiments on Image Classification, Binary Networks, and Neural Machine Translation. Finally, we broadly discuss several circumstances wherein label smoothing will indeed lose its effectiveness. Project page: http://zhiqiangshen.com/projects/LS_and_KD/index.html.

arXiv:2104.00299공식 코드 · Papers with Code

Students are the Best Teacher: Exit-Ensemble Distillation with Multi-Exits

This paper proposes a novel knowledge distillation-based learning method to improve the classification performance of convolutional neural networks (CNNs) without a pre-trained teacher network, called exit-ensemble distillation. Our method exploits the multi-exit architecture that adds auxiliary classifiers (called exits) in the middle of a conventional CNN, through which early inference results can be obtained. The idea of our method is to train the network using the ensemble of the exits as the distillation target, which greatly improves the classification performance of the overall network. Our method suggests a new paradigm of knowledge distillation; unlike the conventional notion of distillation where teachers only teach students, we show that students can also help other students and even the teacher to learn better. Experimental results demonstrate that our method achieves significant improvement of classification performance on various popular CNN architectures (VGG, ResNet, ResNeXt, WideResNet, etc.). Furthermore, the proposed method can expedite the convergence of learning with improved stability. Our code will be available on Github.

arXiv:2103.17012공식 코드 · Papers with Code

Knowledge Distillation By Sparse Representation Matching

Knowledge Distillation refers to a class of methods that transfers the knowledge from a teacher network to a student network. In this paper, we propose Sparse Representation Matching (SRM), a method to transfer intermediate knowledge obtained from one Convolutional Neural Network (CNN) to another by utilizing sparse representation learning. SRM first extracts sparse representations of the hidden features of the teacher CNN, which are then used to generate both pixel-level and image-level labels for training intermediate feature maps of the student network. We formulate SRM as a neural processing block, which can be efficiently optimized using stochastic gradient descent and integrated into any CNN in a plug-and-play manner. Our experiments demonstrate that SRM is robust to architectural differences between the teacher and student networks, and outperforms other KD techniques across several datasets.

arXiv:2103.16844

Fixing the Teacher-Student Knowledge Discrepancy in Distillation

Training a small student network with the guidance of a larger teacher network is an effective way to promote the performance of the student. Despite the different types, the guided knowledge used to distill is always kept unchanged for different teacher and student pairs in previous knowledge distillation methods. However, we find that teacher and student models with different networks or trained from different initialization could have distinct feature representations among different channels. (e.g. the high activated channel for different categories). We name this incongruous representation of channels as teacher-student knowledge discrepancy in the distillation process. Ignoring the knowledge discrepancy problem of teacher and student models will make the learning of student from teacher more difficult. To solve this problem, in this paper, we propose a novel student-dependent distillation method, knowledge consistent distillation, which makes teacher's knowledge more consistent with the student and provides the best suitable knowledge to different student networks for distillation. Extensive experiments on different datasets (CIFAR100, ImageNet, COCO) and tasks (image classification, object detection) reveal the widely existing knowledge discrepancy problem between teachers and students and demonstrate the effectiveness of our proposed method. Our method is very flexible that can be easily combined with other state-of-the-art approaches.

arXiv:2103.16617공식 코드 · Papers with Code

HAD-Net: A Hierarchical Adversarial Knowledge Distillation Network for Improved Enhanced Tumour Segmentation Without Post-Contrast Images

Segmentation of enhancing tumours or lesions from MRI is important for detecting new disease activity in many clinical contexts. However, accurate segmentation requires the inclusion of medical images (e.g., T1 post contrast MRI) acquired after injecting patients with a contrast agent (e.g., Gadolinium), a process no longer thought to be safe. Although a number of modality-agnostic segmentation networks have been developed over the past few years, they have been met with limited success in the context of enhancing pathology segmentation. In this work, we present HAD-Net, a novel offline adversarial knowledge distillation (KD) technique, whereby a pre-trained teacher segmentation network, with access to all MRI sequences, teaches a student network, via hierarchical adversarial training, to better overcome the large domain shift presented when crucial images are absent during inference. In particular, we apply HAD-Net to the challenging task of enhancing tumour segmentation when access to post-contrast imaging is not available. The proposed network is trained and tested on the BraTS 2019 brain tumour segmentation challenge dataset, where it achieves performance improvements in the ranges of 16% - 26% over (a) recent modality-agnostic segmentation methods (U-HeMIS, U-HVED), (b) KD-Net adapted to this problem, (c) the pre-trained student network and (d) a non-hierarchical version of the network (AD-Net), in terms of Dice scores for enhancing tumour (ET). The network also shows improvements in tumour core (TC) Dice scores. Finally, the network outperforms both the baseline student network and AD-Net in terms of uncertainty quantification for enhancing tumour segmentation based on the BraTs 2019 uncertainty challenge metrics. Our code is publicly available at: https://github.com/SaverioVad/HAD_Net

논문 ID: publication/3461729

Task Decoupled Knowledge Distillation For Lightweight Face Detectors

We propose a knowledge distillation method for the face detection task. This method decouples the distillation task of face detection into two subtasks, i.e., the classification distillation subtask and the regression distillation subtask. We add the task-specific convolutions in the teacher network and add the adaption convolutions on the feature maps of the student network to generate the task decoupled features. Then, each subtask uses different samples in distilling the features to be consistent with the corresponding detection subtask. Moreover, we propose an effective probability distillation method to joint boost the accuracy of the student network.

arXiv:2103.16367

Complementary Relation Contrastive Distillation

Knowledge distillation aims to transfer representation ability from a teacher model to a student model. Previous approaches focus on either individual representation distillation or inter-sample similarity preservation. While we argue that the inter-sample relation conveys abundant information and needs to be distilled in a more effective way. In this paper, we propose a novel knowledge distillation method, namely Complementary Relation Contrastive Distillation (CRCD), to transfer the structural knowledge from the teacher to the student. Specifically, we estimate the mutual relation in an anchor-based way and distill the anchor-student relation under the supervision of its corresponding anchor-teacher relation. To make it more robust, mutual relations are modeled by two complementary elements: the feature and its gradient. Furthermore, the low bound of mutual information between the anchor-teacher relation distribution and the anchor-student relation distribution is maximized via relation contrastive loss, which can distill both the sample representation and the inter-sample relations. Experiments on different benchmarks demonstrate the effectiveness of our proposed CRCD.

arXiv:2103.15871

Industry Scale Semi-Supervised Learning for Natural Language Understanding

This paper presents a production Semi-Supervised Learning (SSL) pipeline based on the student-teacher framework, which leverages millions of unlabeled examples to improve Natural Language Understanding (NLU) tasks. We investigate two questions related to the use of unlabeled data in production SSL context: 1) how to select samples from a huge unlabeled data pool that are beneficial for SSL training, and 2) how do the selected data affect the performance of different state-of-the-art SSL techniques. We compare four widely used SSL techniques, Pseudo-Label (PL), Knowledge Distillation (KD), Virtual Adversarial Training (VAT) and Cross-View Training (CVT) in conjunction with two data selection methods including committee-based selection and submodular optimization based selection. We further examine the benefits and drawbacks of these techniques when applied to intent classification (IC) and named entity recognition (NER) tasks, and provide guidelines specifying when each of these methods might be beneficial to improve large scale NLU systems.

arXiv:2103.15042

Distilling Virtual Examples for Long-tailed Recognition

We tackle the long-tailed visual recognition problem from the knowledge distillation perspective by proposing a Distill the Virtual Examples (DiVE) method. Specifically, by treating the predictions of a teacher model as virtual examples, we prove that distilling from these virtual examples is equivalent to label distribution learning under certain constraints. We show that when the virtual example distribution becomes flatter than the original input distribution, the under-represented tail classes will receive significant improvements, which is crucial in long-tailed recognition. The proposed DiVE method can explicitly tune the virtual example distribution to become flat. Extensive experiments on three benchmark datasets, including the large-scale iNaturalist ones, justify that the proposed DiVE method can significantly outperform state-of-the-art methods. Furthermore, additional analyses and experiments verify the virtual example interpretation, and demonstrate the effectiveness of tailored designs in DiVE for long-tailed problems.

arXiv:2103.14908

Embedding Transfer with Label Relaxation for Improved Metric Learning

This paper presents a novel method for embedding transfer, a task of transferring knowledge of a learned embedding model to another. Our method exploits pairwise similarities between samples in the source embedding space as the knowledge, and transfers them through a loss used for learning target embedding models. To this end, we design a new loss called relaxed contrastive loss, which employs the pairwise similarities as relaxed labels for inter-sample relations. Our loss provides a rich supervisory signal beyond class equivalence, enables more important pairs to contribute more to training, and imposes no restriction on manifolds of target embedding spaces. Experiments on metric learning benchmarks demonstrate that our method largely improves performance, or reduces sizes and output dimensions of target models effectively. We further show that it can be also used to enhance quality of self-supervised representation and performance of classification models. In all the experiments, our method clearly outperforms existing embedding transfer techniques.