Recent years have witnessed the significant progress of action recognition task with deep networks. However, most of current video networks require large memory and computational resources, which hinders their applications in practice. Existing knowledge distillation methods are limited to the image-level spatial domain, ignoring the temporal and frequency information which provide structural knowledge and are important for video analysis. This paper explores how to train small and efficient networks for action recognition. Specifically, we propose two distillation strategies in the frequency domain, namely the feature spectrum and parameter distribution distillations respectively. Our insight is that appealing performance of action recognition requires \textit{explicitly} modeling the temporal frequency spectrum of video features. Therefore, we introduce a spectrum loss that enforces the student network to mimic the temporal frequency spectrum from the teacher network, instead of \textit{implicitly} distilling features as many previous works. Second, the parameter frequency distribution is further adopted to guide the student network to learn the appearance modeling process from the teacher. Besides, a collaborative learning strategy is presented to optimize the training process from a probabilistic view. Extensive experiments are conducted on several action recognition benchmarks, such as Kinetics, Something-Something, and Jester, which consistently verify effectiveness of our approach, and demonstrate that our method can achieve higher performance than state-of-the-art methods with the same backbone.
The performance of autoregressive models on natural language generation tasks has dramatically improved due to the adoption of deep, self-attentive architectures. However, these gains have come at the cost of hindering inference speed, making state-of-the-art models cumbersome to deploy in real-world, time-sensitive settings. We develop a compression technique for autoregressive models that is driven by an imitation learning perspective on knowledge distillation. The algorithm is designed to address the exposure bias problem. On prototypical language generation tasks such as translation and summarization, our method consistently outperforms other distillation algorithms, such as sequence-level knowledge distillation. Student models trained with our method attain 1.4 to 4.8 BLEU/ROUGE points higher than those trained from scratch, while increasing inference speed by up to 14 times in comparison to the teacher model.
Domain adaptive person re-identification (re-ID) is a challenging task due to the large discrepancy between the source domain and the target domain. To reduce the domain discrepancy, existing methods mainly attempt to generate pseudo labels for unlabeled target images by clustering algorithms. However, clustering methods tend to bring noisy labels and the rich fine-grained details in unlabeled images are not sufficiently exploited. In this paper, we seek to improve the quality of labels by capturing feature representation from multiple augmented views of unlabeled images. To this end, we propose a Self-Supervised Knowledge Distillation (SSKD) technique containing two modules, the identity learning and the soft label learning. Identity learning explores the relationship between unlabeled samples and predicts their one-hot labels by clustering to give exact information for confidently distinguished images. Soft label learning regards labels as a distribution and induces an image to be associated with several related classes for training peer network in a self-supervised manner, where the slowly evolving network is a core to obtain soft labels as a gentle constraint for reliable images. Finally, the two modules can resist label noise for re-ID by enhancing each other and systematically integrating label information from unlabeled images. Extensive experiments on several adaptation tasks demonstrate that the proposed method outperforms the current state-of-the-art approaches by large margins.
Knowledge Graph Embedding (KGE) is a popular method for KG reasoning and training KGEs with higher dimension are usually preferred since they have better reasoning capability. However, high-dimensional KGEs pose huge challenges to storage and computing resources and are not suitable for resource-limited or time-constrained applications, for which faster and cheaper reasoning is necessary. To address this problem, we propose DualDE, a knowledge distillation method to build low-dimensional student KGE from pre-trained high-dimensional teacher KGE. DualDE considers the dual-influence between the teacher and the student. In DualDE, we propose a soft label evaluation mechanism to adaptively assign different soft label and hard label weights to different triples, and a two-stage distillation approach to improve the student's acceptance of the teacher. Our DualDE is general enough to be applied to various KGEs. Experimental results show that our method can successfully reduce the embedding parameters of a high-dimensional KGE by 7 times - 15 times and increase the inference speed by 2 times - 6 times while retaining a high performance. We also experimentally prove the effectiveness of our soft label evaluation mechanism and two-stage distillation approach via ablation study.
As a pre-trained Transformer model, BERT (Bidirectional Encoder Representations from Transformers) has achieved ground-breaking performance on multiple NLP tasks. On the other hand, Boosting is a popular ensemble learning technique which combines many base classifiers and has been demonstrated to yield better generalization performance in many machine learning tasks. Some works have indicated that ensemble of BERT can further improve the application performance. However, current ensemble approaches focus on bagging or stacking and there has not been much effort on exploring the boosting. In this work, we proposed a novel Boosting BERT model to integrate multi-class boosting into the BERT. Our proposed model uses the pre-trained Transformer as the base classifier to choose harder training sets to fine-tune and gains the benefits of both the pre-training language knowledge and boosting ensemble in NLP tasks. We evaluate the proposed model on the GLUE dataset and 3 popular Chinese NLU benchmarks. Experimental results demonstrate that our proposed model significantly outperforms BERT on all datasets and proves its effectiveness in many NLP tasks. Replacing the BERT base with RoBERTa as base classifier, BoostingBERT achieves new state-of-the-art results in several NLP Tasks. We also use knowledge distillation within the "teacher-student" framework to reduce the computational overhead and model storage of BoostingBERT while keeping its performance for practical application.
Inspired by the strong correlation between the Label Smoothing Regularization(LSR) and Knowledge distillation(KD), we propose an algorithm LsrKD for training boost by extending the LSR method to the KD regime and applying a softer temperature. Then we improve the LsrKD by a Teacher Correction(TC) method, which manually sets a constant larger proportion for the right class in the uniform distribution teacher. To further improve the performance of LsrKD, we develop a self-distillation method named Memory-replay Knowledge Distillation (MrKD) that provides a knowledgeable teacher to replace the uniform distribution one in LsrKD. The MrKD method penalizes the KD loss between the current model's output distributions and its copies' on the training trajectory. By preventing the model learning so far from its historical output distribution space, MrKD can stabilize the learning and find a more robust minimum. Our experiments show that LsrKD can improve LSR performance consistently at no cost, especially on several deep neural networks where LSR is ineffectual. Also, MrKD can significantly improve single model training. The experiment results confirm that the TC can help LsrKD and MrKD to boost training, especially on the networks they are failed. Overall, LsrKD, MrKD, and their TC variants are comparable to or outperform the LSR method, suggesting the broad applicability of these KD methods.
To put a state-of-the-art neural network to practical use, it is necessary to design a model that has a good trade-off between the resource consumption and performance on the test set. Many researchers and engineers are developing methods that enable training or designing a model more efficiently. Developing an efficient model includes several strategies such as network architecture search, pruning, quantization, knowledge distillation, utilizing cheap convolution, regularization, and also includes any craft that leads to a better performance-resource trade-off. When combining these technologies together, it would be ideal if one source of performance improvement does not conflict with others. We call this property as the orthogonality in model efficiency. In this paper, we focus on knowledge distillation and demonstrate that knowledge distillation methods are orthogonal to other efficiency-enhancing methods both analytically and empirically. Analytically, we claim that knowledge distillation functions analogous to a ensemble method, bootstrap aggregating. This analytical explanation is provided from the perspective of implicit data augmentation property of knowledge distillation. Empirically, we verify knowledge distillation as a powerful apparatus for practical deployment of efficient neural network, and also introduce ways to integrate it with other methods effectively.
Knowledge distillation allows transferring knowledge from a pre-trained model to another. However, it suffers from limitations, and constraints related to the two models need to be architecturally similar. Knowledge distillation addresses some of the shortcomings associated with transfer learning by generalizing a complex model to a lighter model. However, some parts of the knowledge may not be distilled by knowledge distillation sufficiently. In this paper, a novel knowledge distillation approach using transfer learning is proposed. The proposed method transfers the entire knowledge of a model to a new smaller one. To accomplish this, unlabeled data are used in an unsupervised manner to transfer the maximum amount of knowledge to the new slimmer model. The proposed method can be beneficial in medical image analysis, where labeled data are typically scarce. The proposed approach is evaluated in the context of classification of images for diagnosing Diabetic Retinopathy on two publicly available datasets, including Messidor and EyePACS. Simulation results demonstrate that the approach is effective in transferring knowledge from a complex model to a lighter one. Furthermore, experimental results illustrate that the performance of different small models is improved significantly using unlabeled data and knowledge distillation.
Many adversarial attacks have been proposed to investigate the security issues of deep neural networks. In the black-box setting, current model stealing attacks train a substitute model to counterfeit the functionality of the target model. However, the training requires querying the target model. Consequently, the query complexity remains high, and such attacks can be defended easily. This study aims to train a generalized substitute model called "Simulator", which can mimic the functionality of any unknown target model. To this end, we build the training data with the form of multiple tasks by collecting query sequences generated during the attacks of various existing networks. The learning process uses a mean square error-based knowledge-distillation loss in the meta-learning to minimize the difference between the Simulator and the sampled networks. The meta-gradients of this loss are then computed and accumulated from multiple tasks to update the Simulator and subsequently improve generalization. When attacking a target model that is unseen in training, the trained Simulator can accurately simulate its functionality using its limited feedback. As a result, a large fraction of queries can be transferred to the Simulator, thereby reducing query complexity. Results of the comprehensive experiments conducted using the CIFAR-10, CIFAR-100, and TinyImageNet datasets demonstrate that the proposed approach reduces query complexity by several orders of magnitude compared to the baseline method. The implementation source code is released at https://github.com/machanic/SimulatorAttack.
This paper studies learning node representations with graph neural networks (GNNs) for unsupervised scenario. Specifically, we derive a theoretical analysis and provide an empirical demonstration about the non-steady performance of GNNs over different graph datasets, when the supervision signals are not appropriately defined. The performance of GNNs depends on both the node feature smoothness and the locality of graph structure. To smooth the discrepancy of node proximity measured by graph topology and node feature, we proposed SAIL - a novel \underline{S}elf-\underline{A}ugmented graph contrast\underline{i}ve \underline{L}earning framework, with two complementary self-distilling regularization modules, \emph{i.e.}, intra- and inter-graph knowledge distillation. We demonstrate the competitive performance of SAIL on a variety of graph applications. Even with a single GNN layer, SAIL has consistently competitive or even better performance on various benchmark datasets, comparing with state-of-the-art baselines.
Selecting radiology examination protocol is a repetitive, and time-consuming process. In this paper, we present a deep learning approach to automatically assign protocols to computer tomography examinations, by pre-training a domain-specific BERT model ($BERT_{rad}$). To handle the high data imbalance across exam protocols, we used a knowledge distillation approach that up-sampled the minority classes through data augmentation. We compared classification performance of the described approach with the statistical n-gram models using Support Vector Machine (SVM), Gradient Boosting Machine (GBM), and Random Forest (RF) classifiers, as well as the Google's $BERT_{base}$ model. SVM, GBM and RF achieved macro-averaged F1 scores of 0.45, 0.45, and 0.6 while $BERT_{base}$ and $BERT_{rad}$ achieved 0.61 and 0.63. Knowledge distillation improved overall performance on the minority classes, achieving a F1 score of 0.66.
Existing vision-based action recognition is susceptible to occlusion and appearance variations, while wearable sensors can alleviate these challenges by capturing human motion with one-dimensional time-series signal. For the same action, the knowledge learned from vision sensors and wearable sensors, may be related and complementary. However, there exists significantly large modality difference between action data captured by wearable-sensor and vision-sensor in data dimension, data distribution and inherent information content. In this paper, we propose a novel framework, named Semantics-aware Adaptive Knowledge Distillation Networks (SAKDN), to enhance action recognition in vision-sensor modality (videos) by adaptively transferring and distilling the knowledge from multiple wearable sensors. The SAKDN uses multiple wearable-sensors as teacher modalities and uses RGB videos as student modality. To preserve local temporal relationship and facilitate employing visual deep learning model, we transform one-dimensional time-series signals of wearable sensors to two-dimensional images by designing a gramian angular field based virtual image generation model. Then, we build a novel Similarity-Preserving Adaptive Multi-modal Fusion Module to adaptively fuse intermediate representation knowledge from different teacher networks. Finally, to fully exploit and transfer the knowledge of multiple well-trained teacher networks to the student network, we propose a novel Graph-guided Semantically Discriminative Mapping loss, which utilizes graph-guided ablation analysis to produce a good visual explanation highlighting the important regions across modalities and concurrently preserving the interrelations of original data. Experimental results on Berkeley-MHAD, UTD-MHAD and MMAct datasets well demonstrate the effectiveness of our proposed SAKDN.
Recent advances in object detection have benefited significantly from rapid developments in deep neural networks. However, neural networks suffer from the well-known issue of catastrophic forgetting, which makes continual or lifelong learning problematic. In this paper, we leverage the fact that new training classes arrive in a sequential manner and incrementally refine the model so that it additionally detects new object classes in the absence of previous training data. Specifically, we consider the representative object detector, Faster R-CNN, for both accurate and efficient prediction. To prevent abrupt performance degradation due to catastrophic forgetting, we propose to apply knowledge distillation on both the region proposal network and the region classification network, to retain the detection of previously trained classes. A pseudo-positive-aware sampling strategy is also introduced for distillation sample selection. We evaluate the proposed method on PASCAL VOC 2007 and MS COCO benchmarks and show competitive mAP and 6x inference speed improvement, which makes the approach more suitable for real-time applications. Our implementation will be publicly available.
We investigate the task of learning blind image denoising networks from an unpaired set of clean and noisy images. Such problem setting generally is practical and valuable considering that it is feasible to collect unpaired noisy and clean images in most real-world applications. And we further assume that the noise can be signal dependent but is spatially uncorrelated. In order to facilitate unpaired learning of denoising network, this paper presents a two-stage scheme by incorporating self-supervised learning and knowledge distillation. For self-supervised learning, we suggest a dilated blind-spot network (D-BSN) to learn denoising solely from real noisy images. Due to the spatial independence of noise, we adopt a network by stacking 1x1 convolution layers to estimate the noise level map for each image. Both the D-BSN and image-specific noise model (CNN\_est) can be jointly trained via maximizing the constrained log-likelihood. Given the output of D-BSN and estimated noise level map, improved denoising performance can be further obtained based on the Bayes' rule. As for knowledge distillation, we first apply the learned noise models to clean images to synthesize a paired set of training images, and use the real noisy images and the corresponding denoising results in the first stage to form another paired set. Then, the ultimate denoising model can be distilled by training an existing denoising network using these two paired sets. Experiments show that our unpaired learning method performs favorably on both synthetic noisy images and real-world noisy photographs in terms of quantitative and qualitative evaluation.
Incremental Learning (IL) is useful when artificial systems need to deal with streams of data and do not have access to all data at all times. The most challenging setting requires a constant complexity of the deep model and an incremental model update without access to a bounded memory of past data. Then, the representations of past classes are strongly affected by catastrophic forgetting. To mitigate its negative effect, an adapted fine tuning which includes knowledge distillation is usually deployed. We propose a different approach based on a vanilla fine tuning backbone. It leverages initial classifier weights which provide a strong representation of past classes because they are trained with all class data. However, the magnitude of classifiers learned in different states varies and normalization is needed for a fair handling of all classes. Normalization is performed by standardizing the initial classifier weights, which are assumed to be normally distributed. In addition, a calibration of prediction scores is done by using state level statistics to further improve classification fairness. We conduct a thorough evaluation with four public datasets in a memoryless incremental learning setting. Results show that our method outperforms existing techniques by a large margin for large-scale datasets.
Knowledge Distillation (KD) has been one of the most popu-lar methods to learn a compact model. However, it still suffers from highdemand in time and computational resources caused by sequential train-ing pipeline. Furthermore, the soft targets from deeper models do notoften serve as good cues for the shallower models due to the gap of com-patibility. In this work, we consider these two problems at the same time.Specifically, we propose that better soft targets with higher compatibil-ity can be generated by using a label generator to fuse the feature mapsfrom deeper stages in a top-down manner, and we can employ the meta-learning technique to optimize this label generator. Utilizing the softtargets learned from the intermediate feature maps of the model, we canachieve better self-boosting of the network in comparison with the state-of-the-art. The experiments are conducted on two standard classificationbenchmarks, namely CIFAR-100 and ILSVRC2012. We test various net-work architectures to show the generalizability of our MetaDistiller. Theexperiments results on two datasets strongly demonstrate the effective-ness of our method.
Federated learning is a privacy-preserving machine learning technique that learns a shared model across decentralized clients. It can alleviate privacy concerns of personal re-identification, an important computer vision task. In this work, we implement federated learning to person re-identification (FedReID) and optimize its performance affected by statistical heterogeneity in the real-world scenario. We first construct a new benchmark to investigate the performance of FedReID. This benchmark consists of (1) nine datasets with different volumes sourced from different domains to simulate the heterogeneous situation in reality, (2) two federated scenarios, and (3) an enhanced federated algorithm for FedReID. The benchmark analysis shows that the client-edge-cloud architecture, represented by the federated-by-dataset scenario, has better performance than client-server architecture in FedReID. It also reveals the bottlenecks of FedReID under the real-world scenario, including poor performance of large datasets caused by unbalanced weights in model aggregation and challenges in convergence. Then we propose two optimization methods: (1) To address the unbalanced weight problem, we propose a new method to dynamically change the weights according to the scale of model changes in clients in each training round; (2) To facilitate convergence, we adopt knowledge distillation to refine the server model with knowledge generated from client models on a public dataset. Experiment results demonstrate that our strategies can achieve much better convergence with superior performance on all datasets. We believe that our work will inspire the community to further explore the implementation of federated learning on more computer vision tasks in real-world scenarios.
In this paper, we propose a simple yet effective approach, named Point Adversarial Self Mining (PASM), to improve the recognition accuracy in facial expression recognition. Unlike previous works focusing on designing specific architectures or loss functions to solve this problem, PASM boosts the network capability by simulating human learning processes: providing updated learning materials and guidance from more capable teachers. Specifically, to generate new learning materials, PASM leverages a point adversarial attack method and a trained teacher network to locate the most informative position related to the target task, generating harder learning samples to refine the network. The searched position is highly adaptive since it considers both the statistical information of each sample and the teacher network capability. Other than being provided new learning materials, the student network also receives guidance from the teacher network. After the student network finishes training, the student network changes its role and acts as a teacher, generating new learning materials and providing stronger guidance to train a better student network. The adaptive learning materials generation and teacher/student update can be conducted more than one time, improving the network capability iteratively. Extensive experimental results validate the efficacy of our method over the existing state of the arts for facial expression recognition.
Knowledge Transfer has been applied in solving a wide variety of problems. For example, knowledge can be transferred between tasks (e.g., learning to handle novel situations by leveraging prior knowledge) or between agents (e.g., learning from others without direct experience). Without loss of generality, we relate knowledge transfer to KL-divergence minimization, i.e., matching the (belief) distributions of learners and teachers. The equivalence gives us a new perspective in understanding variants of the KL-divergence by looking at how learners structure their interaction with teachers in order to acquire knowledge. In this paper, we provide an in-depth analysis of KL-divergence minimization in Forward and Backward orders, which shows that learners are reinforced via on-policy learning in Backward. In contrast, learners are supervised in Forward. Moreover, our analysis is gradient-based, so it can be generalized to arbitrary tasks and help to decide which order to minimize given the property of the task. By replacing Forward with Backward in Knowledge Distillation, we observed +0.7-1.1 BLEU gains on the WMT'17 De-En and IWSLT'15 Th-En machine translation tasks.
Multi-person pose estimation plays an important role in many fields. Although previous works have researched a lot on different parts of human pose estimation, full body pose estimation for multi-person still needs further research. Our work has developed an integrated model through knowledge distillation which can estimate full body poses. Trained based on the AlphaPose system and MSCOCO2017 dataset, our model achieves 51.5 mAP on the validation dataset annotated manually by ourselves. Related resources are available at https://esflfei.github.io/esflfei.gethub.io/website.html.