PyTorchKR Papers

최신 논문

총 4,526편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2204.06130

Impossible Triangle: What's Next for Pre-trained Language Models?

Recent development of large-scale pre-trained language models (PLM) have significantly improved the capability of models in various NLP tasks, in terms of performance after task-specific fine-tuning and zero-shot / few-shot learning. However, many of such models come with a dauntingly huge size that few institutions can afford to pre-train, fine-tune or even deploy, while moderate-sized models usually lack strong generalized few-shot learning capabilities. In this paper, we first elaborate the current obstacles of using PLM models in terms of the Impossible Triangle: 1) moderate model size, 2) state-of-the-art few-shot learning capability, and 3) state-of-the-art fine-tuning capability. We argue that all existing PLM models lack one or more properties from the Impossible Triangle. To remedy these missing properties of PLMs, various techniques have been proposed, such as knowledge distillation, data augmentation and prompt learning, which inevitably brings additional work to the application of PLMs in real scenarios. We then offer insights into future research directions of PLMs to achieve the Impossible Triangle, and break down the task into several key phases.

arXiv:2204.05957공식 코드 · Papers with Code

Localization Distillation for Object Detection

Previous knowledge distillation (KD) methods for object detection mostly focus on feature imitation instead of mimicking the prediction logits due to its inefficiency in distilling the localization information. In this paper, we investigate whether logit mimicking always lags behind feature imitation. Towards this goal, we first present a novel localization distillation (LD) method which can efficiently transfer the localization knowledge from the teacher to the student. Second, we introduce the concept of valuable localization region that can aid to selectively distill the classification and localization knowledge for a certain region. Combining these two new components, for the first time, we show that logit mimicking can outperform feature imitation and the absence of localization distillation is a critical reason for why logit mimicking underperforms for years. The thorough studies exhibit the great potential of logit mimicking that can significantly alleviate the localization ambiguity, learn robust feature representation, and ease the training difficulty in the early stage. We also provide the theoretical connection between the proposed LD and the classification KD, that they share the equivalent optimization effect. Our distillation scheme is simple as well as effective and can be easily applied to both dense horizontal object detectors and rotated object detectors. Extensive experiments on the MS COCO, PASCAL VOC, and DOTA benchmarks demonstrate that our method can achieve considerable AP improvement without any sacrifice on the inference speed. Our source code and pretrained models are publicly available at https://github.com/HikariTJU/LD.

arXiv:2204.05547

DistPro: Searching A Fast Knowledge Distillation Process via Meta Optimization

Recent Knowledge distillation (KD) studies show that different manually designed schemes impact the learned results significantly. Yet, in KD, automatically searching an optimal distillation scheme has not yet been well explored. In this paper, we propose DistPro, a novel framework which searches for an optimal KD process via differentiable meta-learning. Specifically, given a pair of student and teacher networks, DistPro first sets up a rich set of KD connection from the transmitting layers of the teacher to the receiving layers of the student, and in the meanwhile, various transforms are also proposed for comparing feature maps along its pathway for the distillation. Then, each combination of a connection and a transform choice (pathway) is associated with a stochastic weighting process which indicates its importance at every step during the distillation. In the searching stage, the process can be effectively learned through our proposed bi-level meta-optimization strategy. In the distillation stage, DistPro adopts the learned processes for knowledge distillation, which significantly improves the student accuracy especially when faster training is required. Lastly, we find the learned processes can be generalized between similar tasks and networks. In our experiments, DistPro produces state-of-the-art (SoTA) accuracy under varying number of learning epochs on popular datasets, i.e. CIFAR100 and ImageNet, which demonstrate the effectiveness of our framework.

arXiv:2204.05502

CoupleFace: Relation Matters for Face Recognition Distillation

Knowledge distillation is an effective method to improve the performance of a lightweight neural network (i.e., student model) by transferring the knowledge of a well-performed neural network (i.e., teacher model), which has been widely applied in many computer vision tasks, including face recognition. Nevertheless, the current face recognition distillation methods usually utilize the Feature Consistency Distillation (FCD) (e.g., L2 distance) on the learned embeddings extracted by the teacher and student models for each sample, which is not able to fully transfer the knowledge from the teacher to the student for face recognition. In this work, we observe that mutual relation knowledge between samples is also important to improve the discriminative ability of the learned representation of the student model, and propose an effective face recognition distillation method called CoupleFace by additionally introducing the Mutual Relation Distillation (MRD) into existing distillation framework. Specifically, in MRD, we first propose to mine the informative mutual relations, and then introduce the Relation-Aware Distillation (RAD) loss to transfer the mutual relation knowledge of the teacher model to the student model. Extensive experimental results on multiple benchmark datasets demonstrate the effectiveness of our proposed CoupleFace for face recognition. Moreover, based on our proposed CoupleFace, we have won the first place in the ICCV21 Masked Face Recognition Challenge (MS1M track).

arXiv:2204.03541공식 코드 · Papers with Code

End-to-End Zero-Shot HOI Detection via Vision and Language Knowledge Distillation

Most existing Human-Object Interaction~(HOI) Detection methods rely heavily on full annotations with predefined HOI categories, which is limited in diversity and costly to scale further. We aim at advancing zero-shot HOI detection to detect both seen and unseen HOIs simultaneously. The fundamental challenges are to discover potential human-object pairs and identify novel HOI categories. To overcome the above challenges, we propose a novel end-to-end zero-shot HOI Detection (EoID) framework via vision-language knowledge distillation. We first design an Interactive Score module combined with a Two-stage Bipartite Matching algorithm to achieve interaction distinguishment for human-object pairs in an action-agnostic manner. Then we transfer the distribution of action probability from the pretrained vision-language teacher as well as the seen ground truth to the HOI model to attain zero-shot HOI classification. Extensive experiments on HICO-Det dataset demonstrate that our model discovers potential interactive pairs and enables the recognition of unseen HOIs. Finally, our method outperforms the previous SOTA by 8.92% on unseen mAP and 10.18% on overall mAP under UA setting, by 6.02% on unseen mAP and 9.1% on overall mAP under UC setting. Moreover, our method is generalizable to large-scale object detection data to further scale up the action sets. The source code will be available at: https://github.com/mrwu-mac/EoID.

arXiv:2204.03475공식 코드 · Papers with Code

Solving ImageNet: a Unified Scheme for Training any Backbone to Top Results

ImageNet serves as the primary dataset for evaluating the quality of computer-vision models. The common practice today is training each architecture with a tailor-made scheme, designed and tuned by an expert. In this paper, we present a unified scheme for training any backbone on ImageNet. The scheme, named USI (Unified Scheme for ImageNet), is based on knowledge distillation and modern tricks. It requires no adjustments or hyper-parameters tuning between different models, and is efficient in terms of training times. We test USI on a wide variety of architectures, including CNNs, Transformers, Mobile-oriented and MLP-only. On all models tested, USI outperforms previous state-of-the-art results. Hence, we are able to transform training on ImageNet from an expert-oriented task to an automatic seamless routine. Since USI accepts any backbone and trains it to top results, it also enables to perform methodical comparisons, and identify the most efficient backbones along the speed-accuracy Pareto curve. Implementation is available at:https://github.com/Alibaba-MIIL/Solving_ImageNet

arXiv:2204.02136공식 코드 · Papers with Code

Overcoming Catastrophic Forgetting in Incremental Object Detection via Elastic Response Distillation

Traditional object detectors are ill-equipped for incremental learning. However, fine-tuning directly on a well-trained detection model with only new data will lead to catastrophic forgetting. Knowledge distillation is a flexible way to mitigate catastrophic forgetting. In Incremental Object Detection (IOD), previous work mainly focuses on distilling for the combination of features and responses. However, they under-explore the information that contains in responses. In this paper, we propose a response-based incremental distillation method, dubbed Elastic Response Distillation (ERD), which focuses on elastically learning responses from the classification head and the regression head. Firstly, our method transfers category knowledge while equipping student detector with the ability to retain localization information during incremental learning. In addition, we further evaluate the quality of all locations and provide valuable responses by the Elastic Response Selection (ERS) strategy. Finally, we elucidate that the knowledge from different responses should be assigned with different importance during incremental distillation. Extensive experiments conducted on MS COCO demonstrate our method achieves state-of-the-art result, which substantially narrows the performance gap towards full training.

arXiv:2204.01952

Towards On-Board Panoptic Segmentation of Multispectral Satellite Images

With tremendous advancements in low-power embedded computing devices and remote sensing instruments, the traditional satellite image processing pipeline which includes an expensive data transfer step prior to processing data on the ground is being replaced by on-board processing of captured data. This paradigm shift enables critical and time-sensitive analytic intelligence to be acquired in a timely manner on-board the satellite itself. However, at present, the on-board processing of multi-spectral satellite images is limited to classification and segmentation tasks. Extending this processing to its next logical level, in this paper we propose a lightweight pipeline for on-board panoptic segmentation of multi-spectral satellite images. Panoptic segmentation offers major economic and environmental insights, ranging from yield estimation from agricultural lands to intelligence for complex military applications. Nevertheless, the on-board intelligence extraction raises several challenges due to the loss of temporal observations and the need to generate predictions from a single image sample. To address this challenge, we propose a multimodal teacher network based on a cross-modality attention-based fusion strategy to improve the segmentation accuracy by exploiting data from multiple modes. We also propose an online knowledge distillation framework to transfer the knowledge learned by this multi-modal teacher network to a uni-modal student which receives only a single frame input, and is more appropriate for an on-board environment. We benchmark our approach against existing state-of-the-art panoptic segmentation models using the PASTIS multi-spectral panoptic segmentation dataset considering an on-board processing setting. Our evaluations demonstrate a substantial increase in accuracy metrics compared to the existing state-of-the-art models.

arXiv:2204.00796

A Dual-Contrastive Framework for Low-Resource Cross-Lingual Named Entity Recognition

Cross-lingual Named Entity Recognition (NER) has recently become a research hotspot because it can alleviate the data-hungry problem for low-resource languages. However, few researches have focused on the scenario where the source-language labeled data is also limited in some specific domains. A common approach for this scenario is to generate more training data through translation or generation-based data augmentation method. Unfortunately, we find that simply combining source-language data and the corresponding translation cannot fully exploit the translated data and the improvements obtained are somewhat limited. In this paper, we describe our novel dual-contrastive framework ConCNER for cross-lingual NER under the scenario of limited source-language labeled data. Specifically, based on the source-language samples and their translations, we design two contrastive objectives for cross-language NER at different grammatical levels, namely Translation Contrastive Learning (TCL) to close sentence representations between translated sentence pairs and Label Contrastive Learning (LCL) to close token representations within the same labels. Furthermore, we utilize knowledge distillation method where the NER model trained above is used as the teacher to train a student model on unlabeled target-language data to better fit the target language. We conduct extensive experiments on a wide variety of target languages, and the results demonstrate that ConCNER tends to outperform multiple baseline methods. For reproducibility, our code for this paper is available at https://github.com/GKLMIP/ConCNER.

arXiv:2204.00791공식 코드 · Papers with Code

CL-XABSA: Contrastive Learning for Cross-lingual Aspect-based Sentiment Analysis

As an extensive research in the field of natural language processing (NLP), aspect-based sentiment analysis (ABSA) is the task of predicting the sentiment expressed in a text relative to the corresponding aspect. Unfortunately, most languages lack sufficient annotation resources, thus more and more recent researchers focus on cross-lingual aspect-based sentiment analysis (XABSA). However, most recent researches only concentrate on cross-lingual data alignment instead of model alignment. To this end, we propose a novel framework, CL-XABSA: Contrastive Learning for Cross-lingual Aspect-Based Sentiment Analysis. Based on contrastive learning, we close the distance between samples with the same label in different semantic spaces, thus achieving a convergence of semantic spaces of different languages. Specifically, we design two contrastive strategies, token level contrastive learning of token embeddings (TL-CTE) and sentiment level contrastive learning of token embeddings (SL-CTE), to regularize the semantic space of source and target language to be more uniform. Since our framework can receive datasets in multiple languages during training, our framework can be adapted not only for XABSA task but also for multilingual aspect-based sentiment analysis (MABSA). To further improve the performance of our model, we perform knowledge distillation technology leveraging data from unlabeled target language. In the distillation XABSA task, we further explore the comparative effectiveness of different data (source dataset, translated dataset, and code-switched dataset). The results demonstrate that the proposed method has a certain improvement in the three tasks of XABSA, distillation XABSA and MABSA. For reproducibility, our code for this paper is available at https://github.com/GKLMIP/CL-XABSA.

arXiv:2204.01689

Using Explainable Boosting Machine to Compare Idiographic and Nomothetic Approaches for Ecological Momentary Assessment Data

Previous research on EMA data of mental disorders was mainly focused on multivariate regression-based approaches modeling each individual separately. This paper goes a step further towards exploring the use of non-linear interpretable machine learning (ML) models in classification problems. ML models can enhance the ability to accurately predict the occurrence of different behaviors by recognizing complicated patterns between variables in data. To evaluate this, the performance of various ensembles of trees are compared to linear models using imbalanced synthetic and real-world datasets. After examining the distributions of AUC scores in all cases, non-linear models appear to be superior to baseline linear models. Moreover, apart from personalized approaches, group-level prediction models are also likely to offer an enhanced performance. According to this, two different nomothetic approaches to integrate data of more than one individuals are examined, one using directly all data during training and one based on knowledge distillation. Interestingly, it is observed that in one of the two real-world datasets, knowledge distillation method achieves improved AUC scores (mean relative change of +17\% compared to personalized) showing how it can benefit EMA data classification and performance.

arXiv:2204.01542

CDKT-FL: Cross-Device Knowledge Transfer using Proxy Dataset in Federated Learning

In a practical setting towards better generalization abilities of client models for realizing robust personalized Federated Learning (FL) systems, efficient model aggregation methods have been considered as a critical research objective. It is a challenging issue due to the consequences of non-i.i.d. properties of client's data, often referred to as statistical heterogeneity and small local data samples from the various data distributions. Therefore, to develop robust generalized global and personalized models, conventional FL methods need redesigning the knowledge aggregation from biased local models while considering huge divergence of learning parameters due to skewed client data. In this work, we demonstrate that the knowledge transfer mechanism is a de facto technique to achieve these objectives and develop a novel knowledge distillation-based approach to study the extent of knowledge transfer between the global model and local models. Henceforth, our method considers the suitability of transferring the outcome distribution and (or) the embedding vector of representation from trained models during cross-device knowledge transfer using a small proxy dataset in heterogeneous FL. In doing so, we alternatively perform cross-device knowledge transfer following general formulations as 1) global knowledge transfer and 2) on-device knowledge transfer. Through simulations on four federated datasets, we show the proposed method achieves significant speedups and high personalized performance of local models. Furthermore, the proposed approach offers a more stable algorithm than FedAvg during the training, with minimal communication data load when exchanging the trained model's outcomes and representation.

arXiv:2204.00895공식 코드 · Papers with Code

Class-Incremental Learning by Knowledge Distillation with Adaptive Feature Consolidation

We present a novel class incremental learning approach based on deep neural networks, which continually learns new tasks with limited memory for storing examples in the previous tasks. Our algorithm is based on knowledge distillation and provides a principled way to maintain the representations of old models while adjusting to new tasks effectively. The proposed method estimates the relationship between the representation changes and the resulting loss increases incurred by model updates. It minimizes the upper bound of the loss increases using the representations, which exploits the estimated importance of each feature map within a backbone model. Based on the importance, the model restricts updates of important features for robustness while allowing changes in less critical features for flexibility. This optimization strategy effectively alleviates the notorious catastrophic forgetting problem despite the limited accessibility of data in the previous tasks. The experimental results show significant accuracy improvement of the proposed algorithm over the existing methods on the standard datasets. Code is available.

arXiv:2204.00649

Knowledge distillation with error-correcting transfer learning for wind power prediction

Wind power prediction, especially for turbines, is vital for the operation, controllability, and economy of electricity companies. Hybrid methodologies combining advanced data science with weather forecasting have been incrementally applied to the predictions. Nevertheless, individually modeling massive turbines from scratch and downscaling weather forecasts to turbine size are neither easy nor economical. Aiming at it, this paper proposes a novel framework with mathematical underpinnings for turbine power prediction. This framework is the first time to incorporate knowledge distillation into energy forecasting, enabling accurate and economical constructions of turbine models by learning knowledge from the well-established park model. Besides, park-scale weather forecasts non-explicitly are mapped to turbines by transfer learning of predicted power errors, achieving model correction for better performance. The proposed framework is deployed on five turbines featuring various terrains in an Arctic wind park, the results are evaluated against the competitors of ablation investigation. The major findings reveal that the proposed framework, developed on favorable knowledge distillation and transfer learning parameters tuning, yields performance boosts from 3.3 % to 23.9 % over its competitors. This advantage also exists in terms of wind energy physics and computing efficiency, which are verified by the prediction quality rate and calculation time.

arXiv:2204.01210공식 코드 · Papers with Code

Co-Teaching for Unsupervised Domain Adaptation and Expansion

Unsupervised Domain Adaptation (UDA) essentially trades a model's performance on a source domain for improving its performance on a target domain. To resolve the issue, Unsupervised Domain Expansion (UDE) has been proposed recently. UDE tries to adapt the model for the target domain as UDA does, and in the meantime maintains its source-domain performance. In both UDA and UDE settings, a model tailored to a given domain, let it be the source or the target domain, is assumed to well handle samples from the given domain. We question the assumption by reporting the existence of cross-domain visual ambiguity: Given the lack of a crystally clear boundary between the two domains, samples from one domain can be visually close to the other domain. Such sorts of samples are typically in minority in their host domain, so they tend to be overlooked by the domain-specific model, but can be better handled by a model from the other domain. We exploit this finding, and accordingly propose Co-Teaching (CT). The CT method is instantiated with knowledge distillation based CT (kdCT) plus mixup based CT (miCT). Specifically, kdCT transfers knowledge from a leading-teacher network and an assistant-teacher network to a student network, so the cross-domain ambiguity will be better handled by the student. Meanwhile, miCT further enhances the generalization ability of the student. Extensive experiments on two image classification datasets and two driving-scene segmentation datasets justify the viability of CT for UDA and UDE.

arXiv:2204.00972

DST: Dynamic Substitute Training for Data-free Black-box Attack

With the wide applications of deep neural network models in various computer vision tasks, more and more works study the model vulnerability to adversarial examples. For data-free black box attack scenario, existing methods are inspired by the knowledge distillation, and thus usually train a substitute model to learn knowledge from the target model using generated data as input. However, the substitute model always has a static network structure, which limits the attack ability for various target models and tasks. In this paper, we propose a novel dynamic substitute training attack method to encourage substitute model to learn better and faster from the target model. Specifically, a dynamic substitute structure learning strategy is proposed to adaptively generate optimal substitute model structure via a dynamic gate according to different target models and tasks. Moreover, we introduce a task-driven graph-based structure information learning constrain to improve the quality of generated training data, and facilitate the substitute model learning structural relationships from the target model multiple outputs. Extensive experiments have been conducted to verify the efficacy of the proposed attack method, which can achieve better performance compared with the state-of-the-art competitors on several datasets.

arXiv:2204.00771

Fast Real-time Personalized Speech Enhancement: End-to-End Enhancement Network (E3Net) and Knowledge Distillation

This paper investigates how to improve the runtime speed of personalized speech enhancement (PSE) networks while maintaining the model quality. Our approach includes two aspects: architecture and knowledge distillation (KD). We propose an end-to-end enhancement (E3Net) model architecture, which is $3\times$ faster than a baseline STFT-based model. Besides, we use KD techniques to develop compressed student models without significantly degrading quality. In addition, we investigate using noisy data without reference clean signals for training the student models, where we combine KD with multi-task learning (MTL) using automatic speech recognition (ASR) loss. Our results show that E3Net provides better speech and transcription quality with a lower target speaker over-suppression (TSOS) rate than the baseline model. Furthermore, we show that the KD methods can yield student models that are $2-4\times$ faster than the teacher and provides reasonable quality. Combining KD and MTL improves the ASR and TSOS metrics without degrading the speech quality.

arXiv:2204.00548

Unified and Effective Ensemble Knowledge Distillation

Ensemble knowledge distillation can extract knowledge from multiple teacher models and encode it into a single student model. Many existing methods learn and distill the student model on labeled data only. However, the teacher models are usually learned on the same labeled data, and their predictions have high correlations with groudtruth labels. Thus, they cannot provide sufficient knowledge complementary to task labels for student teaching. Distilling on unseen unlabeled data has the potential to enhance the knowledge transfer from the teachers to the student. In this paper, we propose a unified and effective ensemble knowledge distillation method that distills a single student model from an ensemble of teacher models on both labeled and unlabeled data. Since different teachers may have diverse prediction correctness on the same sample, on labeled data we weight the predictions of different teachers according to their correctness. In addition, we weight the distillation loss based on the overall prediction correctness of the teacher ensemble to distill high-quality knowledge. On unlabeled data, there is no groundtruth to evaluate prediction correctness. Fortunately, the disagreement among teachers is an indication of sample hardness, and thereby we weight the distillation loss based on teachers' disagreement to emphasize knowledge distillation on important samples. Extensive experiments on four datasets show the effectiveness of our proposed ensemble distillation method.

arXiv:2204.00292

Preventing Distillation-based Attacks on Neural Network IP

Neural networks (NNs) are already deployed in hardware today, becoming valuable intellectual property (IP) as many hours are invested in their training and optimization. Therefore, attackers may be interested in copying, reverse engineering, or even modifying this IP. The current practices in hardware obfuscation, including the widely studied logic locking technique, are insufficient to protect the actual IP of a well-trained NN: its weights. Simply hiding the weights behind a key-based scheme is inefficient (resource-hungry) and inadequate (attackers can exploit knowledge distillation). This paper proposes an intuitive method to poison the predictions that prevent distillation-based attacks; this is the first work to consider such a poisoning approach in hardware-implemented NNs. The proposed technique obfuscates a NN so an attacker cannot train the NN entirely or accurately. We elaborate a threat model which highlights the difference between random logic obfuscation and the obfuscation of NN IP. Based on this threat model, our security analysis shows that the poisoning successfully and significantly reduces the accuracy of the stolen NN model on various representative datasets. Moreover, the accuracy and prediction distributions are maintained, no functionality is disturbed, nor are high overheads incurred. Finally, we highlight that our proposed approach is flexible and does not require manipulation of the NN toolchain.

arXiv:2204.00270

Rethinking Position Bias Modeling with Knowledge Distillation for CTR Prediction

Click-through rate (CTR) Prediction is of great importance in real-world online ads systems. One challenge for the CTR prediction task is to capture the real interest of users from their clicked items, which is inherently biased by presented positions of items, i.e., more front positions tend to obtain higher CTR values. A popular line of existing works focuses on explicitly estimating position bias by result randomization which is expensive and inefficient, or by inverse propensity weighting (IPW) which relies heavily on the quality of the propensity estimation. Another common solution is modeling position as features during offline training and simply adopting fixed value or dropout tricks when serving. However, training-inference inconsistency can lead to sub-optimal performance. Furthermore, post-click information such as position values is informative while less exploited in CTR prediction. This work proposes a simple yet efficient knowledge distillation framework to alleviate the impact of position bias and leverage position information to improve CTR prediction. We demonstrate the performance of our proposed method on a real-world production dataset and online A/B tests, achieving significant improvements over competing baseline models. The proposed method has been deployed in the real world online ads systems, serving main traffic on one of the world's largest e-commercial platforms.