Distillation Framework Research Articles

BERTtoCNN: Similarity-preserving enhanced knowledge distillation for stance detection.

In recent years, text sentiment analysis has attracted wide attention, and promoted the rise and development of stance detection research. The purpose of stance detection is to determine the author’s stance (favor or against) towards a specific target or proposition in the text. Pre-trained language models like BERT have been proven to perform well in this task. However, in many reality scenes, they are usually very expensive in computation, because such heavy models are difficult to implement with limited resources. To improve the efficiency while ensuring the performance, we propose a knowledge distillation model BERTtoCNN, which combines the classic distillation loss and similarity-preserving loss in a joint knowledge distillation framework. On the one hand, BERTtoCNN provides an efficient distillation process to train a novel ‘student’ CNN structure from a much larger ‘teacher’ language model BERT. On the other hand, based on the similarity-preserving loss function, BERTtoCNN guides the training of a student network, so that input pairs with similar (dissimilar) activation in the teacher network have similar (dissimilar) activation in the student network. We conduct experiments and test the proposed model on the open Chinese and English stance detection datasets. The experimental results show that our model outperforms the competitive baseline methods obviously.

Positive-parity baryon spectrum and the role of hybrid baryons

We calculate the low-lying spectra for the positive-parity $\Delta$ and $N$ at two pion masses of 358 and 278 MeV using an isotropic clover action with two degenerate light-quaark and one strange-quark flavors through the application of the generalized variational method within the distillation framework. The spectrum exhibits the general features observed in previous calculations using an anisotropic clover lattice, with a counting of states at least as rich as the quark model. Furthermore, we identify states that are hybrid in nature, where gluonic degrees of freedom play a structural role, indicatinng that such states appear a feature of the excited baryon spectrum, irrespective of the lattice action, or the precise details of the smearing of the lattice interpolating operators used to identify such states.

Top-aware recommender distillation with deep reinforcement learning

Most existing recommenders focus on providing users with a list of recommended products. In practical, users may only pay attention to the recommendations at the top positions. Our analysis, however, shows that the correctly recommended products from existing methods are not top posited; this would result in sacrificing users’ patience and engagements. To address this issue, this paper proposes a top-aware recommender distillation (TRD) framework, viz., the rank of recommendation lists from given state-of-the-art recommendation approaches are further reinforced and refined using reinforcement learning. Different from the traditional goal of knowledge distillation to mimic the behavior of its teacher, our recommender contrastingly goes step further with the goal of surpassing its teacher recommender. More importantly, our proposed TRD can be plugged on any existed recommender; thus being generic for real-world deployment. Theoretical analysis demonstrates that TRD is guaranteed to perform at least as well as the basic recommender. Extensive experiments on five state-of-the-art recommendation algorithms across three real-world datasets further show that TRD consistently improves the recommendations at top positions.

DnS: Distill-and-Select for Efficient and Accurate Video Indexing and Retrieval

In this paper, we address the problem of high performance and computationally efficient content-based video retrieval in large-scale datasets. Current methods typically propose either: (i) fine-grained approaches employing spatio-temporal representations and similarity calculations, achieving high performance at a high computational cost or (ii) coarse-grained approaches representing/indexing videos as global vectors, where the spatio-temporal structure is lost, providing low performance but also having low computational cost. In this work, we propose a Knowledge Distillation framework, which we call Distill-and-Select (DnS), that starting from a well-performing fine-grained Teacher Network learns: a) Student Networks at different retrieval performance and computational efficiency trade-offs and b) a Selection Network that at test time rapidly directs samples to the appropriate student to maintain both high retrieval performance and high computational efficiency. We train several students with different architectures and arrive at different trade-offs of performance and efficiency, i.e., speed and storage requirements, including fine-grained students that store index videos using binary representations. Importantly, the proposed scheme allows Knowledge Distillation in large, unlabelled datasets -- this leads to good students. We evaluate DnS on five public datasets on three different video retrieval tasks and demonstrate a) that our students achieve state-of-the-art performance in several cases and b) that our DnS framework provides an excellent trade-off between retrieval performance, computational speed, and storage space. In specific configurations, our method achieves similar mAP with the teacher but is 20 times faster and requires 240 times less storage space. Our collected dataset and implementation are publicly available: this https URL.

Classifier-adaptation knowledge distillation framework for relation extraction and event detection with imbalanced data

Fundamental information extraction tasks, such as relation extraction and event detection, suffer from a data imbalance problem. To alleviate this problem, existing methods rely mostly on well-designed loss functions to reduce the negative influence of imbalanced data. However, this approach requires additional hyper-parameters and limits scalability. Furthermore, these methods can only benefit specific tasks and do not provide a unified framework across relation extraction and event detection. In this paper, a Classifier-Adaptation Knowledge Distillation (CAKD) framework is proposed to address these issues, thus improving relation extraction and event detection performance. The first step is to exploit sentence-level identification information across relation extraction and event detection, which can reduce identification errors caused by the data imbalance problem without relying on additional hyper-parameters. Moreover, this sentence-level identification information is used by a teacher network to guide the baseline model’s training by sharing its classifier. Like an instructor, the classifier improves the baseline model’s ability to extract this sentence-level identification information from raw texts, thus benefiting overall performance. Experiments were conducted on both relation extraction and event detection using the Text Analysis Conference Relation Extraction Dataset (TACRED) and Automatic Content Extraction (ACE) 2005 English datasets, respectively. The results demonstrate the effectiveness of the proposed framework.

Continual Learning for Named Entity Recognition

Named Entity Recognition (NER) is a vital task in various NLP applications. However, in many real-world scenarios (e.g., voice-enabled assistants) new named entities are frequently introduced, entailing re-training NER models to support these new entities. Re-annotating the original training data for the new entities could be costly or even impossible when storage limitations or security concerns restrict access to that data, and annotating a new dataset for all of the entities becomes impractical and error-prone as the number of entities increases. To tackle this problem, we introduce a novel Continual Learning approach for NER, which requires new training material to be annotated only for the new entities. To preserve the existing knowledge previously learned by the model, we exploit the Knowledge Distillation (KD) framework, where the existing NER model acts as the teacher for a new NER model (i.e., the student), which learns the new entity by using the new training material and retains knowledge of old entities by imitating the teacher's outputs on this new training set. Our experiments show that this approach allows the student model to ``progressively'' learn to identify new entities without forgetting the previously learned ones. We also present a comparison with multiple strong baselines to demonstrate that our approach is superior for continually updating an NER model.

Lord of the Rings: Hanoi Pooling and Self-Knowledge Distillation for Fast and Accurate Vehicle Reidentification

Vehicle reidentification has seen increasing interest, thanks to its fundamental impact on intelligent surveillance systems and smart transportation. The visual data acquired from monitoring camera networks come with severe challenges, including occlusions, color and illumination changes, as well as orientation issues (a vehicle can be seen from the side/front/rear due to different camera viewpoints). To deal with such challenges, the community has spent much effort in learning robust feature representations that hinge on additional visual attributes and part-driven methods, but with the side effects of requiring extensive human annotation labor as well as increasing computational complexity. In this article, we propose an approach that learns a feature representation robust to vehicle orientation issues without the need for extra-labeled data and adding negligible computational overheads. The former objective is achieved through the introduction of a Hanoi pooling layer exploiting ring regions and the image pyramid approach yielding a multiscale representation of vehicle appearance. The latter is tackled by transferring the accuracy of a deep network to its first layers, thus reducing the inference effort by the early stop of a test example. This is obtained by means of a self-knowledge distillation framework encouraging multiexit network decisions to agree with each other. Results demonstrate that the proposed approach significantly improves the accuracy of early (i.e., very fast) exits while maintaining the same accuracy of a deep (slow) baseline. Moreover, our solution obtains the best existing performance on three benchmark datasets. <xref ref-type="fn" rid="fn1" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">¹</xref> <fn id="fn1" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><label>¹</label> [Online]. Available: <uri>https://github.com/iN1k1/</uri>. </fn>

Deep Cross-Modal Representation Learning and Distillation for Illumination-Invariant Pedestrian Detection

Integrating multispectral data has been demonstrated to be an effective solution for illumination-invariant pedestrian detection, in particular, RGB and thermal images can provide complementary information to handle light variations. However, most of the current multispectral detectors fuse the multimodal features by simple concatenation, without discovering their latent relationships. In this paper, we propose a cross-modal feature learning (CFL) module, based on a split-and-aggregation strategy, to explicitly explore both the shared and modality-specific representations between paired RGB and thermal images. We insert the proposed CFL module into multiple layers of a two-branch-based pedestrian detection network, to learn the cross-modal representations in diverse semantic levels. By introducing a segmentation-based auxiliary task, the multimodal network is trained end-to-end by jointly optimizing a multi-task loss. On the other hand, to alleviate the reliance of existing multispectral pedestrian detectors on thermal images, we propose a knowledge distillation framework to train a student detector, which only receives RGB images as input and distills the cross-modal representations guided by a well-trained multimodal teacher detector. In order to facilitate the cross-modal knowledge distillation, we design different distillation loss functions for the feature, detection and segmentation levels. Experimental results on the public KAIST multispectral pedestrian benchmark validate that the proposed cross-modal representation learning and distillation method achieves robust performance.

KDnet-RUL: A Knowledge Distillation Framework to Compress Deep Neural Networks for Machine Remaining Useful Life Prediction

Machine remaining useful life (RUL) prediction is vital in improving the reliability of industrial systems and reducing maintenance cost. Recently, long short-term memory (LSTM) based algorithms have achieved state-of-the-art performance for RUL prediction due to their strong capability of modeling sequential sensory data. In many cases, the RUL prediction algorithms are required to be deployed on edge devices to support real-time decision making, reduce the data communication cost, and preserve the data privacy. However, the powerful LSTM-based methods which have high complexity cannot be deployed to edge devices with limited computational power and memory. To solve this problem, we propose a knowledge distillation framework, entitled KDnet-RUL, to compress a complex LSTM-based method for RUL prediction. Specifically, it includes a generative adversarial network based knowledge distillation (GAN-KD) for disparate architecture knowledge transfer, a learning-during-teaching based knowledge distillation (LDT-KD) for identical architecture knowledge transfer, and a sequential distillation upon LDT-KD for complicated datasets. We leverage simple and complicated datasets to verify the effectiveness of the proposed KDnet-RUL. The results demonstrate that the proposed method significantly outperforms state-of-the-art KD methods. The compressed model with 12.8 times less weights and 46.2 times less total float point operations even achieves a comparable performance with the complex LSTM model for RUL prediction.

Distillation at high momentum

Extraction of hadronic observables at finite-momenta from Lattice QCD (LQCD) is constrained by the well-known signal-to-noise problems afflicting all such LQCD calculations. Traditional quark smearing algorithms are commonly used tools to improve the statistical quality of hadronic $n$-point functions, provided operator momenta are small. The momentum smearing algorithm of Bali et al. extends the range of momenta that are cleanly accessible, and has facilitated countless novel lattice calculations. Momentum smearing has, however, not been explicitly demonstrated within the framework of distillation. In this work we extend the momentum-smearing idea, by exploring a few modifications to the distillation framework. Together with enhanced time slice sampling and expanded operator bases engendered by distillation, we find ground-state nucleon energies can be extracted reliably for $\left|\vec{p}\right|\lesssim3\text{ GeV}$ and matrix elements featuring a large momentum dependence can be resolved.

Resolution-Aware Knowledge Distillation for Efficient Inference.

Minimizing the computation complexity is essential for the popularization of deep networks in practical applications. Nowadays, most researches attempt to accelerate deep networks by designing new network structure or compressing the network parameters. Meanwhile, transfer learning techniques such as knowledge distillation are utilized to keep the performance of deep models. In this paper, we focus on accelerating deep models and relieving the computation burden by using low-resolution (LR) images as inputs while maintaining competitive performance, which is rarely researched in the current literature. Deep networks may encounter serious performance degradation when using LR inputs because many details are unavailable from LR images. Besides, the existing approaches may fail to learn discriminative features for LR images because of the dramatic appearance variations between LR and high-resolution (HR) images. To tackle with the above problems, we propose a resolution-aware knowledge distillation (RKD) framework to narrow the cross-resolution variations by transferring knowledge from HR domain to LR domain. The proposed framework consists of a HR teacher network and a LR student network. First, we introduce a discriminator and propose an adversarial learning strategy to shrink the variations between inputs with changing resolution. Then we design a cross-resolution knowledge distillation (CRKD) loss to train discriminative student network by exploiting the knowledge of the teacher network. The CRKD loss is consisted of a resolution-aware distillation loss, a pair-wise constraint, and a maximum mean discrepancy loss. Experimental results on person re-identification, image classification, face recognition, and defect segmentation tasks demonstrate that RKD outperforms traditional knowledge distillation method by achieving better performance with lower computation complexities. Furthermore, CRKD surpasses the state-of-the-art knowledge distillation methods in transferring knowledge across different resolutions under RKD framework, especially when coping with large resolution differences.

Double Similarity Distillation for Semantic Image Segmentation.

The balance between high accuracy and high speed has always been a challenging task in semantic image segmentation. Compact segmentation networks are more widely used in the case of limited resources, while their performances are constrained. In this paper, motivated by the residual learning and global aggregation, we propose a simple yet general and effective knowledge distillation framework called double similarity distillation (DSD) to improve the classification accuracy of all existing compact networks by capturing the similarity knowledge in pixel and category dimensions, respectively. Specifically, we propose a pixel-wise similarity distillation (PSD) module that utilizes residual attention maps to capture more detailed spatial dependencies across multiple layers. Compared with exiting methods, the PSD module greatly reduces the amount of calculation and is easy to expand. Furthermore, considering the differences in characteristics between semantic segmentation task and other computer vision tasks, we propose a category-wise similarity distillation (CSD) module, which can help the compact segmentation network strengthen the global category correlation by constructing the correlation matrix. Combining these two modules, DSD framework has no extra parameters and only a minimal increase in FLOPs. Extensive experiments on four challenging datasets, including Cityscapes, CamVid, ADE20K, and Pascal VOC 2012, show that DSD outperforms current state-of-the-art methods, proving its effectiveness and generality. The code and models will be publicly available.

Mutual-learning sequence-level knowledge distillation for automatic speech recognition

Automatic speech recognition (ASR) is a crucial technology for man-machine interaction. End-to-end models have been studied recently in deep learning for ASR. However, these models are not suitable for the practical application of ASR due to their large model sizes and computation costs. To address this issue, we propose a novel mutual-learning sequence-level knowledge distillation framework enjoying distinct student structures for ASR. Trained mutually and simultaneously, each student learns not only from the pre-trained teacher but also from its distinct peers, which can improve the generalization capability of the whole network, through making up for the insufficiency of each student and bridging the gap between each student and the teacher. Extensive experiments on the TIMIT and large LibriSpeech corpuses show that, compared with the state-of-the-art methods, the proposed method achieves an excellent balance between recognition accuracy and model compression.

Topological Space Knowledge Distillation for Compact Road Extraction in Optical Remote Sensing Images

Road extraction from optical remote sensing images has drawn much attention in recent decades and has a wide range of applications. Most of the previous studies rarely take into account the unique topological characteristics of the road. It is the most apparent feature of linear structure that describes the variety of connection relationships of the road. However, designing a particular topological feature extraction network usually results in a model that is too heavy and impractical. To address the problems mentioned above, in this paper, we propose a lightweight topological space network for road extraction based on knowledge distillation (TSKD-Road). Specifically, (1) narrow and short roads easily influence topological features extracted directly in optical remote sensing images. Therefore, we propose a denser teacher network for extracting road structures; (2) to enhance the weight of topological features, we propose a topological space loss calculation model with multiple widths and depths; (3) based on the above innovations, a topological space knowledge distillation framework is proposed, which aims to transfer different kinds of knowledge acquired in a heavy net to a lightweight net, while significantly improving the lightweight net’s accuracy. Experiments were conducted on two publicly available benchmark datasets, which show the obvious superiority and effectiveness of our network.

Ensemble Learning of Lightweight Deep Learning Models Using Knowledge Distillation for Image Classification

In recent years, deep learning models have been used successfully in almost every field including both industry and academia, especially for computer vision tasks. However, these models are huge in size, with millions (and billions) of parameters, and thus cannot be deployed on the systems and devices with limited resources (e.g., embedded systems and mobile phones). To tackle this, several techniques on model compression and acceleration have been proposed. As a representative type of them, knowledge distillation suggests a way to effectively learn a small student model from large teacher model(s). It has attracted increasing attention since it showed its promising performance. In the work, we propose an ensemble model that combines feature-based, response-based, and relation-based lightweight knowledge distillation models for simple image classification tasks. In our knowledge distillation framework, we use ResNet−20 as a student network and ResNet−110 as a teacher network. Experimental results demonstrate that our proposed ensemble model outperforms other knowledge distillation models as well as the large teacher model for image classification tasks, with less computational power than the teacher model.

GENERALIZED KNOWLEDGE DISTILLATION FOR MULTI-SENSOR REMOTE SENSING CLASSIFICATION: AN APPLICATION TO LAND COVER MAPPING

Abstract. Due to the proliferation of Earth Observation programmes, information at different spatial, spectral and temporal resolution is collected by means of various sensors (optical, radar, hyperspectral, LiDAR, etc.). Despite such abundance of information, it is not always possible to obtain a complete coverage of the same area (especially for large ones) from all the different sensors due to: (i) atmospheric conditions and/or (ii) acquisition cost. In this context of data (or modalities) misalignment, only part of the area under consideration could be covered by the different sensors (modalities). Unfortunately, standard machine learning approaches commonly employed in operational Earth monitoring systems require consistency between training and test data (i.e., they need to match the same information schema). Such a constraint limits the use of additional fruitful information, i.e., information coming from a particular sensor that may be available at training but not at test time. Recently, a framework able to manage such information misalignment between training and test information is proposed under the name of Generalized Knowledge Distillation (GKD). With the aim to provide a proof of concept of GKD in the context of multi-source Earth Observation analysis, here we provide a Generalized Knowledge Distillation framework for land use land cover mapping involving radar (Sentinel-1) and optical (Sentinel-2) satellite image time series data (SITS). Considering that part of the optical information may not be available due to bad atmospheric conditions, we make the assumption that radar SITS are always available (at both training and test time) while optical SITS are only accessible when the model is learnt (i.e., it is considered as privileged information). Evaluations are carried out on a real-world study area in the southwest of France, namely Dordogne, considering a mapping task involving seven different land use land cover classes. Experimental results underline how the additional (privileged) information ameliorates the results of the radar based classification with a main gain on the agricultural classes.

Online Knowledge Distillation with Diverse Peers

Distillation is an effective knowledge-transfer technique that uses predicted distributions of a powerful teacher model as soft targets to train a less-parameterized student model. A pre-trained high capacity teacher, however, is not always available. Recently proposed online variants use the aggregated intermediate predictions of multiple student models as targets to train each student model. Although group-derived targets give a good recipe for teacher-free distillation, group members are homogenized quickly with simple aggregation functions, leading to early saturated solutions. In this work, we propose Online Knowledge Distillation with Diverse peers (OKDDip), which performs two-level distillation during training with multiple auxiliary peers and one group leader. In the first-level distillation, each auxiliary peer holds an individual set of aggregation weights generated with an attention-based mechanism to derive its own targets from predictions of other auxiliary peers. Learning from distinct target distributions helps to boost peer diversity for effectiveness of group-based distillation. The second-level distillation is performed to transfer the knowledge in the ensemble of auxiliary peers further to the group leader, i.e., the model used for inference. Experimental results show that the proposed framework consistently gives better performance than state-of-the-art approaches without sacrificing training or inference complexity, demonstrating the effectiveness of the proposed two-level distillation framework.

Knowledge Distillation in Acoustic Scene Classification

Common acoustic properties that different classes share degrades the performance of acoustic scene classification systems. This results in a phenomenon where a few confusing pairs of acoustic scenes dominate a significant proportion of all misclassified audio segments. In this article, we propose adopting a knowledge distillation framework that trains deep neural networks using soft labels. Soft labels, extracted from another pre-trained deep neural network, are used to reflect the similarity between different classes that share similar acoustic properties. We also propose utilizing specialist models to provide additional soft labels. Each specialist model in this study refers to a deep neural network that concentrates on discriminating a single pair of acoustic scenes that are frequently misclassified. Self multi-head attention is explored for training specialist deep neural networks to further concentrate on target pairs of classes. The goal of this article is to train a single deep neural network that demonstrates performance equivalent to, or higher than, an ensemble of multiple models, by distilling the knowledge from several models. Diverse experiments conducted using the detection and classification of acoustic scenes and events 2019 task 1-a dataset demonstrate that the knowledge distillation framework is effective in acoustic scene classification. Specialist models successfully decrease the number of misclassified audio segments in the target classes. The final single model with the proposed method that is trained by the proposed knowledge distillation from several models, including specialists trained using an attention mechanism, shows a classification accuracy of 77.63 %, higher than an ensemble of the baseline and multiple specialists.

Improving Low-Resource Neural Machine Translation With Teacher-Free Knowledge Distillation

Knowledge Distillation (KD) aims to distill the knowledge of a cumbersome teacher model into a lightweight student model. Its success is generally attributed to the privileged information on similarities among categories provided by the teacher model, and in this sense, only strong teacher models are deployed to teach weaker students in practice. However, in low-resource neural machine translation, a stronger teacher model is not available. To counteract this, We therefore propose a novel Teacher-free Knowledge Distillation framework for low-resource neural machine translation, where the model learns from manually designed regularization distribution as a virtual teacher model. The prior distribution of artificial design can not only obtain the similarity information between words, but also provide effective regularity for model training. Experimental results show that the proposed method has improved performance in low-resource language effectively.

Progressive Motion Representation Distillation With Two-Branch Networks for Egocentric Activity Recognition

Video-based egocentric activity recognition involves fine-grained spatio-temporal human-object interactions. State-of-the-art methods, based on the two-branch-based architecture, rely on pre-calculated optical flows to provide motion information. However, this two-stage strategy is computationally intensive, storage demanding, and not task-oriented, which hampers it from being deployed in real-world applications. Albeit there have been numerous attempts to explore other motion representations to replace optical flows, most of the methods were designed for third-person activities, without capturing fine-grained cues. To tackle these issues, in this letter, we propose a progressive motion representation distillation (PMRD) method, based on two-branch networks, for egocentric activity recognition. We exploit a generalized knowledge distillation framework to train a hallucination network, which receives RGB frames as input and produces motion cues guided by the optical-flow network. Specifically, we propose a progressive metric loss, which aims to distill local fine-grained motion patterns in terms of each temporal progress level. To further enforce the proposed distillation framework to concentrate on those informative frames, we integrate a temporal attention mechanism into the metric loss. Moreover, a multi-stage training procedure is employed for the efficient learning of the hallucination network. Experimental results on three egocentric activity benchmarks demonstrate the state-of-the-art performance of the proposed method.