PyTorchKR Papers

최신 논문

총 4,526편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2105.10633

Revisiting Knowledge Distillation for Object Detection

The existing solutions for object detection distillation rely on the availability of both a teacher model and ground-truth labels. We propose a new perspective to relax this constraint. In our framework, a student is first trained with pseudo labels generated by the teacher, and then fine-tuned using labeled data, if any available. Extensive experiments demonstrate improvements over existing object detection distillation algorithms. In addition, decoupling the teacher and ground-truth distillation in this framework provides interesting properties such: as 1) using unlabeled data to further improve the student's performance, 2) combining multiple teacher models of different architectures, even with different object categories, and 3) reducing the need for labeled data (with only 20% of COCO labels, this method achieves the same performance as the model trained on the entire set of labels). Furthermore, a by-product of this approach is the potential usage for domain adaptation. We verify these properties through extensive experiments.

arXiv:2105.11166

AirNet: Neural Network Transmission over the Air

State-of-the-art performance for many edge applications is achieved by deep neural networks (DNNs). Often, these DNNs are location- and time-sensitive, and must be delivered over a wireless channel rapidly and efficiently. In this paper, we introduce AirNet, a family of novel training and transmission methods that allow DNNs to be efficiently delivered over wireless channels under stringent transmit power and latency constraints. This corresponds to a new class of joint source-channel coding problems, aimed at delivering DNNs with the goal of maximizing their accuracy at the receiver, rather than recovering them with high fidelity. In AirNet, we propose the direct mapping of the DNN parameters to transmitted channel symbols, while the network is trained to meet the channel constraints, and exhibit robustness against channel noise. AirNet achieves higher accuracy compared to separation-based alternatives. We further improve the performance of AirNet by pruning the network below the available bandwidth, and expanding it for improved robustness. We also benefit from unequal error protection by selectively expanding important layers of the network. Finally, we develop an approach, which simultaneously trains a spectrum of DNNs, each targeting a different channel condition, resolving the impractical memory requirements of training distinct networks for different channel conditions.

arXiv:2105.10123공식 코드 · Papers with Code

Backdoor Attacks on Self-Supervised Learning

Large-scale unlabeled data has spurred recent progress in self-supervised learning methods that learn rich visual representations. State-of-the-art self-supervised methods for learning representations from images (e.g., MoCo, BYOL, MSF) use an inductive bias that random augmentations (e.g., random crops) of an image should produce similar embeddings. We show that such methods are vulnerable to backdoor attacks - where an attacker poisons a small part of the unlabeled data by adding a trigger (image patch chosen by the attacker) to the images. The model performance is good on clean test images, but the attacker can manipulate the decision of the model by showing the trigger at test time. Backdoor attacks have been studied extensively in supervised learning and to the best of our knowledge, we are the first to study them for self-supervised learning. Backdoor attacks are more practical in self-supervised learning, since the use of large unlabeled data makes data inspection to remove poisons prohibitive. We show that in our targeted attack, the attacker can produce many false positives for the target category by using the trigger at test time. We also propose a defense method based on knowledge distillation that succeeds in neutralizing the attack. Our code is available here: https://github.com/UMBCvision/SSL-Backdoor .

arXiv:2105.10056공식 코드 · Papers with Code

Data-Free Knowledge Distillation for Heterogeneous Federated Learning

Federated Learning (FL) is a decentralized machine-learning paradigm, in which a global server iteratively averages the model parameters of local users without accessing their data. User heterogeneity has imposed significant challenges to FL, which can incur drifted global models that are slow to converge. Knowledge Distillation has recently emerged to tackle this issue, by refining the server model using aggregated knowledge from heterogeneous users, other than directly averaging their model parameters. This approach, however, depends on a proxy dataset, making it impractical unless such a prerequisite is satisfied. Moreover, the ensemble knowledge is not fully utilized to guide local model learning, which may in turn affect the quality of the aggregated model. Inspired by the prior art, we propose a data-free knowledge distillation} approach to address heterogeneous FL, where the server learns a lightweight generator to ensemble user information in a data-free manner, which is then broadcasted to users, regulating local training using the learned knowledge as an inductive bias. Empirical studies powered by theoretical implications show that, our approach facilitates FL with better generalization performance using fewer communication rounds, compared with the state-of-the-art.

arXiv:2105.09816공식 코드 · Papers with Code

Intra-Document Cascading: Learning to Select Passages for Neural Document Ranking

An emerging recipe for achieving state-of-the-art effectiveness in neural document re-ranking involves utilizing large pre-trained language models - e.g., BERT - to evaluate all individual passages in the document and then aggregating the outputs by pooling or additional Transformer layers. A major drawback of this approach is high query latency due to the cost of evaluating every passage in the document with BERT. To make matters worse, this high inference cost and latency varies based on the length of the document, with longer documents requiring more time and computation. To address this challenge, we adopt an intra-document cascading strategy, which prunes passages of a candidate document using a less expensive model, called ESM, before running a scoring model that is more expensive and effective, called ETM. We found it best to train ESM (short for Efficient Student Model) via knowledge distillation from the ETM (short for Effective Teacher Model) e.g., BERT. This pruning allows us to only run the ETM model on a smaller set of passages whose size does not vary by document length. Our experiments on the MS MARCO and TREC Deep Learning Track benchmarks suggest that the proposed Intra-Document Cascaded Ranking Model (IDCM) leads to over 400% lower query latency by providing essentially the same effectiveness as the state-of-the-art BERT-based document ranking models.

arXiv:2105.08919공식 코드 · Papers with Code

Comparing Kullback-Leibler Divergence and Mean Squared Error Loss in Knowledge Distillation

Knowledge distillation (KD), transferring knowledge from a cumbersome teacher model to a lightweight student model, has been investigated to design efficient neural architectures. Generally, the objective function of KD is the Kullback-Leibler (KL) divergence loss between the softened probability distributions of the teacher model and the student model with the temperature scaling hyperparameter tau. Despite its widespread use, few studies have discussed the influence of such softening on generalization. Here, we theoretically show that the KL divergence loss focuses on the logit matching when tau increases and the label matching when tau goes to 0 and empirically show that the logit matching is positively correlated to performance improvement in general. From this observation, we consider an intuitive KD loss function, the mean squared error (MSE) between the logit vectors, so that the student model can directly learn the logit of the teacher model. The MSE loss outperforms the KL divergence loss, explained by the difference in the penultimate layer representations between the two losses. Furthermore, we show that sequential distillation can improve performance and that KD, particularly when using the KL divergence loss with small tau, mitigates the label noise. The code to reproduce the experiments is publicly available online at https://github.com/jhoon-oh/kd_data/.

arXiv:2105.08369

Inplace knowledge distillation with teacher assistant for improved training of flexible deep neural networks

Deep neural networks (DNNs) have achieved great success in various machine learning tasks. However, most existing powerful DNN models are computationally expensive and memory demanding, hindering their deployment in devices with low memory and computational resources or in applications with strict latency requirements. Thus, several resource-adaptable or flexible approaches were recently proposed that train at the same time a big model and several resource-specific sub-models. Inplace knowledge distillation (IPKD) became a popular method to train those models and consists in distilling the knowledge from a larger model (teacher) to all other sub-models (students). In this work a novel generic training method called IPKD with teacher assistant (IPKD-TA) is introduced, where sub-models themselves become teacher assistants teaching smaller sub-models. We evaluated the proposed IPKD-TA training method using two state-of-the-art flexible models (MSDNet and Slimmable MobileNet-V1) with two popular image classification benchmarks (CIFAR-10 and CIFAR-100). Our results demonstrate that the IPKD-TA is on par with the existing state of the art while improving it in most cases.

arXiv:2105.08252

Weakly Supervised Dense Video Captioning via Jointly Usage of Knowledge Distillation and Cross-modal Matching

This paper proposes an approach to Dense Video Captioning (DVC) without pairwise event-sentence annotation. First, we adopt the knowledge distilled from relevant and well solved tasks to generate high-quality event proposals. Then we incorporate contrastive loss and cycle-consistency loss typically applied to cross-modal retrieval tasks to build semantic matching between the proposals and sentences, which are eventually used to train the caption generation module. In addition, the parameters of matching module are initialized via pre-training based on annotated images to improve the matching performance. Extensive experiments on ActivityNet-Caption dataset reveal the significance of distillation-based event proposal generation and cross-modal retrieval-based semantic matching to weakly supervised DVC, and demonstrate the superiority of our method to existing state-of-the-art methods.

arXiv:2105.08584공식 코드 · Papers with Code

Contrastive Model Inversion for Data-Free Knowledge Distillation

Model inversion, whose goal is to recover training data from a pre-trained model, has been recently proved feasible. However, existing inversion methods usually suffer from the mode collapse problem, where the synthesized instances are highly similar to each other and thus show limited effectiveness for downstream tasks, such as knowledge distillation. In this paper, we propose Contrastive Model Inversion~(CMI), where the data diversity is explicitly modeled as an optimizable objective, to alleviate the mode collapse issue. Our main observation is that, under the constraint of the same amount of data, higher data diversity usually indicates stronger instance discrimination. To this end, we introduce in CMI a contrastive learning objective that encourages the synthesizing instances to be distinguishable from the already synthesized ones in previous batches. Experiments of pre-trained models on CIFAR-10, CIFAR-100, and Tiny-ImageNet demonstrate that CMI not only generates more visually plausible instances than the state of the arts, but also achieves significantly superior performance when the generated data are used for knowledge distillation. Code is available at \url{https://github.com/zju-vipa/DataFree}.

arXiv:2105.07637

Class-Incremental Few-Shot Object Detection

Conventional detection networks usually need abundant labeled training samples, while humans can learn new concepts incrementally with just a few examples. This paper focuses on a more challenging but realistic class-incremental few-shot object detection problem (iFSD). It aims to incrementally transfer the model for novel objects from only a few annotated samples without catastrophically forgetting the previously learned ones. To tackle this problem, we propose a novel method LEAST, which can transfer with Less forgetting, fEwer training resources, And Stronger Transfer capability. Specifically, we first present the transfer strategy to reduce unnecessary weight adaptation and improve the transfer capability for iFSD. On this basis, we then integrate the knowledge distillation technique using a less resource-consuming approach to alleviate forgetting and propose a novel clustering-based exemplar selection process to preserve more discriminative features previously learned. Being a generic and effective method, LEAST can largely improve the iFSD performance on various benchmarks.

arXiv:2105.07381공식 코드 · Papers with Code

Undistillable: Making A Nasty Teacher That CANNOT teach students

Knowledge Distillation (KD) is a widely used technique to transfer knowledge from pre-trained teacher models to (usually more lightweight) student models. However, in certain situations, this technique is more of a curse than a blessing. For instance, KD poses a potential risk of exposing intellectual properties (IPs): even if a trained machine learning model is released in 'black boxes' (e.g., as executable software or APIs without open-sourcing code), it can still be replicated by KD through imitating input-output behaviors. To prevent this unwanted effect of KD, this paper introduces and investigates a concept called Nasty Teacher: a specially trained teacher network that yields nearly the same performance as a normal one, but would significantly degrade the performance of student models learned by imitating it. We propose a simple yet effective algorithm to build the nasty teacher, called self-undermining knowledge distillation. Specifically, we aim to maximize the difference between the output of the nasty teacher and a normal pre-trained network. Extensive experiments on several datasets demonstrate that our method is effective on both standard KD and data-free KD, providing the desirable KD-immunity to model owners for the first time. We hope our preliminary study can draw more awareness and interest in this new practical problem of both social and legal importance.

arXiv:2105.07519공식 코드 · Papers with Code

Graph-Free Knowledge Distillation for Graph Neural Networks

Knowledge distillation (KD) transfers knowledge from a teacher network to a student by enforcing the student to mimic the outputs of the pretrained teacher on training data. However, data samples are not always accessible in many cases due to large data sizes, privacy, or confidentiality. Many efforts have been made on addressing this problem for convolutional neural networks (CNNs) whose inputs lie in a grid domain within a continuous space such as images and videos, but largely overlook graph neural networks (GNNs) that handle non-grid data with different topology structures within a discrete space. The inherent differences between their inputs make these CNN-based approaches not applicable to GNNs. In this paper, we propose to our best knowledge the first dedicated approach to distilling knowledge from a GNN without graph data. The proposed graph-free KD (GFKD) learns graph topology structures for knowledge transfer by modeling them with multivariate Bernoulli distribution. We then introduce a gradient estimator to optimize this framework. Essentially, the gradients w.r.t. graph structures are obtained by only using GNN forward-propagation without back-propagation, which means that GFKD is compatible with modern GNN libraries such as DGL and Geometric. Moreover, we provide the strategies for handling different types of prior knowledge in the graph data or the GNNs. Extensive experiments demonstrate that GFKD achieves the state-of-the-art performance for distilling knowledge from GNNs without training data.

arXiv:2105.07239공식 코드 · Papers with Code

AgeFlow: Conditional Age Progression and Regression with Normalizing Flows

Age progression and regression aim to synthesize photorealistic appearance of a given face image with aging and rejuvenation effects, respectively. Existing generative adversarial networks (GANs) based methods suffer from the following three major issues: 1) unstable training introducing strong ghost artifacts in the generated faces, 2) unpaired training leading to unexpected changes in facial attributes such as genders and races, and 3) non-bijective age mappings increasing the uncertainty in the face transformation. To overcome these issues, this paper proposes a novel framework, termed AgeFlow, to integrate the advantages of both flow-based models and GANs. The proposed AgeFlow contains three parts: an encoder that maps a given face to a latent space through an invertible neural network, a novel invertible conditional translation module (ICTM) that translates the source latent vector to target one, and a decoder that reconstructs the generated face from the target latent vector using the same encoder network; all parts are invertible achieving bijective age mappings. The novelties of ICTM are two-fold. First, we propose an attribute-aware knowledge distillation to learn the manipulation direction of age progression while keeping other unrelated attributes unchanged, alleviating unexpected changes in facial attributes. Second, we propose to use GANs in the latent space to ensure the learned latent vector indistinguishable from the real ones, which is much easier than traditional use of GANs in the image domain. Experimental results demonstrate superior performance over existing GANs-based methods on two benchmarked datasets. The source code is available at https://github.com/Hzzone/AgeFlow.

arXiv:2105.05912

MATE-KD: Masked Adversarial TExt, a Companion to Knowledge Distillation

The advent of large pre-trained language models has given rise to rapid progress in the field of Natural Language Processing (NLP). While the performance of these models on standard benchmarks has scaled with size, compression techniques such as knowledge distillation have been key in making them practical. We present, MATE-KD, a novel text-based adversarial training algorithm which improves the performance of knowledge distillation. MATE-KD first trains a masked language model based generator to perturb text by maximizing the divergence between teacher and student logits. Then using knowledge distillation a student is trained on both the original and the perturbed training samples. We evaluate our algorithm, using BERT-based models, on the GLUE benchmark and demonstrate that MATE-KD outperforms competitive adversarial learning and data augmentation baselines. On the GLUE test set our 6 layer RoBERTa based model outperforms BERT-Large.

arXiv:2105.06152공식 코드 · Papers with Code

When Human Pose Estimation Meets Robustness: Adversarial Algorithms and Benchmarks

Human pose estimation is a fundamental yet challenging task in computer vision, which aims at localizing human anatomical keypoints. However, unlike human vision that is robust to various data corruptions such as blur and pixelation, current pose estimators are easily confused by these corruptions. This work comprehensively studies and addresses this problem by building rigorous robust benchmarks, termed COCO-C, MPII-C, and OCHuman-C, to evaluate the weaknesses of current advanced pose estimators, and a new algorithm termed AdvMix is proposed to improve their robustness in different corruptions. Our work has several unique benefits. (1) AdvMix is model-agnostic and capable in a wide-spectrum of pose estimation models. (2) AdvMix consists of adversarial augmentation and knowledge distillation. Adversarial augmentation contains two neural network modules that are trained jointly and competitively in an adversarial manner, where a generator network mixes different corrupted images to confuse a pose estimator, improving the robustness of the pose estimator by learning from harder samples. To compensate for the noise patterns by adversarial augmentation, knowledge distillation is applied to transfer clean pose structure knowledge to the target pose estimator. (3) Extensive experiments show that AdvMix significantly increases the robustness of pose estimations across a wide range of corruptions, while maintaining accuracy on clean data in various challenging benchmark datasets.

arXiv:2105.06143공식 코드 · Papers with Code

Boosting Light-Weight Depth Estimation Via Knowledge Distillation

Monocular depth estimation (MDE) methods are often either too computationally expensive or not accurate enough due to the trade-off between model complexity and inference performance. In this paper, we propose a lightweight network that can accurately estimate depth maps using minimal computing resources. We achieve this by designing a compact model architecture that maximally reduces model complexity. To improve the performance of our lightweight network, we adopt knowledge distillation (KD) techniques. We consider a large network as an expert teacher that accurately estimates depth maps on the target domain. The student, which is the lightweight network, is then trained to mimic the teacher's predictions. However, this KD process can be challenging and insufficient due to the large model capacity gap between the teacher and the student. To address this, we propose to use auxiliary unlabeled data to guide KD, enabling the student to better learn from the teacher's predictions. This approach helps fill the gap between the teacher and the student, resulting in improved data-driven learning. Our extensive experiments show that our method achieves comparable performance to state-of-the-art methods while using only 1% of their parameters. Furthermore, our method outperforms previous lightweight methods regarding inference accuracy, computational efficiency, and generalizability.

arXiv:2105.05752

Stacked Acoustic-and-Textual Encoding: Integrating the Pre-trained Models into Speech Translation Encoders

Encoder pre-training is promising in end-to-end Speech Translation (ST), given the fact that speech-to-translation data is scarce. But ST encoders are not simple instances of Automatic Speech Recognition (ASR) or Machine Translation (MT) encoders. For example, we find that ASR encoders lack the global context representation, which is necessary for translation, whereas MT encoders are not designed to deal with long but locally attentive acoustic sequences. In this work, we propose a Stacked Acoustic-and-Textual Encoding (SATE) method for speech translation. Our encoder begins with processing the acoustic sequence as usual, but later behaves more like an MT encoder for a global representation of the input sequence. In this way, it is straightforward to incorporate the pre-trained models into the system. Also, we develop an adaptor module to alleviate the representation inconsistency between the pre-trained ASR encoder and MT encoder, and develop a multi-teacher knowledge distillation method to preserve the pre-training knowledge. Experimental results on the LibriSpeech En-Fr and MuST-C En-De ST tasks show that our method achieves state-of-the-art BLEU scores of 18.3 and 25.2. To our knowledge, we are the first to develop an end-to-end ST system that achieves comparable or even better BLEU performance than the cascaded ST counterpart when large-scale ASR and MT data is available.

arXiv:2105.03544

Test-Time Adaptation Toward Personalized Speech Enhancement: Zero-Shot Learning with Knowledge Distillation

In realistic speech enhancement settings for end-user devices, we often encounter only a few speakers and noise types that tend to reoccur in the specific acoustic environment. We propose a novel personalized speech enhancement method to adapt a compact denoising model to the test-time specificity. Our goal in this test-time adaptation is to utilize no clean speech target of the test speaker, thus fulfilling the requirement for zero-shot learning. To complement the lack of clean utterance, we employ the knowledge distillation framework. Instead of the missing clean utterance target, we distill the more advanced denoising results from an overly large teacher model, and use it as the pseudo target to train the small student model. This zero-shot learning procedure circumvents the process of collecting users' clean speech, a process that users are reluctant to comply due to privacy concerns and technical difficulty of recording clean voice. Experiments on various test-time conditions show that the proposed personalization method achieves significant performance gains compared to larger baseline networks trained from a large speaker- and noise-agnostic datasets. In addition, since the compact personalized models can outperform larger general-purpose models, we claim that the proposed method performs model compression with no loss of denoising performance.

arXiv:2105.04181

KDExplainer: A Task-oriented Attention Model for Explaining Knowledge Distillation

Knowledge distillation (KD) has recently emerged as an efficacious scheme for learning compact deep neural networks (DNNs). Despite the promising results achieved, the rationale that interprets the behavior of KD has yet remained largely understudied. In this paper, we introduce a novel task-oriented attention model, termed as KDExplainer, to shed light on the working mechanism underlying the vanilla KD. At the heart of KDExplainer is a Hierarchical Mixture of Experts (HME), in which a multi-class classification is reformulated as a multi-task binary one. Through distilling knowledge from a free-form pre-trained DNN to KDExplainer, we observe that KD implicitly modulates the knowledge conflicts between different subtasks, and in reality has much more to offer than label smoothing. Based on such findings, we further introduce a portable tool, dubbed as virtual attention module (VAM), that can be seamlessly integrated with various DNNs to enhance their performance under KD. Experimental results demonstrate that with a negligible additional cost, student models equipped with VAM consistently outperform their non-VAM counterparts across different benchmarks. Furthermore, when combined with other KD methods, VAM remains competent in promoting results, even though it is only motivated by vanilla KD. The code is available at https://github.com/zju-vipa/KDExplainer.

arXiv:2105.03048

Regression Bugs Are In Your Model! Measuring, Reducing and Analyzing Regressions In NLP Model Updates

Behavior of deep neural networks can be inconsistent between different versions. Regressions during model update are a common cause of concern that often over-weigh the benefits in accuracy or efficiency gain. This work focuses on quantifying, reducing and analyzing regression errors in the NLP model updates. Using negative flip rate as regression measure, we show that regression has a prevalent presence across tasks in the GLUE benchmark. We formulate the regression-free model updates into a constrained optimization problem, and further reduce it into a relaxed form which can be approximately optimized through knowledge distillation training method. We empirically analyze how model ensemble reduces regression. Finally, we conduct CheckList behavioral testing to understand the distribution of regressions across linguistic phenomena, and the efficacy of ensemble and distillation methods.