As the performance and popularity of deep neural networks has increased, so too has their computational cost. There are many effective techniques for reducing a network's computational footprint (quantisation, pruning, knowledge distillation), but these lead to models whose computational cost is the same regardless of their input. Our human reaction times vary with the complexity of the tasks we perform: easier tasks (e.g. telling apart dogs from boat) are executed much faster than harder ones (e.g. telling apart two similar looking breeds of dogs). Driven by this observation, we develop a method for adaptive network complexity by attaching a small classification layer, which we call SideNet, to a large pretrained network, which we call MainNet. Given an input, the SideNet returns a classification if its confidence level, obtained via softmax, surpasses a user determined threshold, and only passes it along to the large MainNet for further processing if its confidence is too low. This allows us to flexibly trade off the network's performance with its computational cost. Experimental results show that simple single hidden layer perceptron SideNets added onto pretrained ResNet and BERT MainNets allow for substantial decreases in compute with minimal drops in performance on image and text classification tasks. We also highlight three other desirable properties of our method, namely that the classifications obtained by SideNets are calibrated, complementary to other compute reduction techniques, and that they enable the easy exploration of compute accuracy space.
In real world applications like healthcare, it is usually difficult to build a machine learning prediction model that works universally well across different institutions. At the same time, the available model is often proprietary, i.e., neither the model parameter nor the data set used for model training is accessible. In consequence, leveraging the knowledge hidden in the available model (aka. the hypothesis) and adapting it to a local data set becomes extremely challenging. Motivated by this situation, in this paper we aim to address such a specific case within the hypothesis transfer learning framework, in which 1) the source hypothesis is a black-box model and 2) the source domain data is unavailable. In particular, we introduce a novel algorithm called dynamic knowledge distillation for hypothesis transfer learning (dkdHTL). In this method, we use knowledge distillation with instance-wise weighting mechanism to adaptively transfer the "dark" knowledge from the source hypothesis to the target domain.The weighting coefficients of the distillation loss and the standard loss are determined by the consistency between the predicted probability of the source hypothesis and the target ground-truth label.Empirical results on both transfer learning benchmark datasets and a healthcare dataset demonstrate the effectiveness of our method.
Deep learning methods show promising results for overlapping cervical cell instance segmentation. However, in order to train a model with good generalization ability, voluminous pixel-level annotations are demanded which is quite expensive and time-consuming for acquisition. In this paper, we propose to leverage both labeled and unlabeled data for instance segmentation with improved accuracy by knowledge distillation. We propose a novel Mask-guided Mean Teacher framework with Perturbation-sensitive Sample Mining (MMT-PSM), which consists of a teacher and a student network during training. Two networks are encouraged to be consistent both in feature and semantic level under small perturbations. The teacher's self-ensemble predictions from $K$-time augmented samples are used to construct the reliable pseudo-labels for optimizing the student. We design a novel strategy to estimate the sensitivity to perturbations for each proposal and select informative samples from massive cases to facilitate fast and effective semantic distillation. In addition, to eliminate the unavoidable noise from the background region, we propose to use the predicted segmentation mask as guidance to enforce the feature distillation in the foreground region. Experiments show that the proposed method improves the performance significantly compared with the supervised method learned from labeled data only, and outperforms state-of-the-art semi-supervised methods.
We propose a general method to train a single convolutional neural network which is capable of switching image resolutions at inference. Thus the running speed can be selected to meet various computational resource limits. Networks trained with the proposed method are named Resolution Switchable Networks (RS-Nets). The basic training framework shares network parameters for handling images which differ in resolution, yet keeps separate batch normalization layers. Though it is parameter-efficient in design, it leads to inconsistent accuracy variations at different resolutions, for which we provide a detailed analysis from the aspect of the train-test recognition discrepancy. A multi-resolution ensemble distillation is further designed, where a teacher is learnt on the fly as a weighted ensemble over resolutions. Thanks to the ensemble and knowledge distillation, RS-Nets enjoy accuracy improvements at a wide range of resolutions compared with individually trained models. Extensive experiments on the ImageNet dataset are provided, and we additionally consider quantization problems. Code and models are available at https://github.com/yikaiw/RS-Nets.
This paper proposes a knowledge distillation method for foreground object search (FoS). Given a background and a rectangle specifying the foreground location and scale, FoS retrieves compatible foregrounds in a certain category for later image composition. Foregrounds within the same category can be grouped into a small number of patterns. Instances within each pattern are compatible with any query input interchangeably. These instances are referred to as interchangeable foregrounds. We first present a pipeline to build pattern-level FoS dataset containing labels of interchangeable foregrounds. We then establish a benchmark dataset for further training and testing following the pipeline. As for the proposed method, we first train a foreground encoder to learn representations of interchangeable foregrounds. We then train a query encoder to learn query-foreground compatibility following a knowledge distillation framework. It aims to transfer knowledge from interchangeable foregrounds to supervise representation learning of compatibility. The query feature representation is projected to the same latent space as interchangeable foregrounds, enabling very efficient and interpretable instance-level search. Furthermore, pattern-level search is feasible to retrieve more controllable, reasonable and diverse foregrounds. The proposed method outperforms the previous state-of-the-art by 10.42% in absolute difference and 24.06% in relative improvement evaluated by mean average precision (mAP). Extensive experimental results also demonstrate its efficacy from various aspects. The benchmark dataset and code will be release shortly.
Variational mutual information (MI) estimators are widely used in unsupervised representation learning methods such as contrastive predictive coding (CPC). A lower bound on MI can be obtained from a multi-class classification problem, where a critic attempts to distinguish a positive sample drawn from the underlying joint distribution from $(m-1)$ negative samples drawn from a suitable proposal distribution. Using this approach, MI estimates are bounded above by $\log m$, and could thus severely underestimate unless $m$ is very large. To overcome this limitation, we introduce a novel estimator based on a multi-label classification problem, where the critic needs to jointly identify multiple positive samples at the same time. We show that using the same amount of negative samples, multi-label CPC is able to exceed the $\log m$ bound, while still being a valid lower bound of mutual information. We demonstrate that the proposed approach is able to lead to better mutual information estimation, gain empirical improvements in unsupervised representation learning, and beat a current state-of-the-art knowledge distillation method over 10 out of 13 tasks.
Due to the characteristics of COVID-19, the epidemic develops rapidly and overwhelms health service systems worldwide. Many patients suffer from systemic life-threatening problems and need to be carefully monitored in ICUs. Thus the intelligent prognosis is in an urgent need to assist physicians to take an early intervention, prevent the adverse outcome, and optimize the medical resource allocation. However, in the early stage of the epidemic outbreak, the data available for analysis is limited due to the lack of effective diagnostic mechanisms, rarity of the cases, and privacy concerns. In this paper, we propose a deep-learning-based approach, CovidCare, which leverages the existing electronic medical records to enhance the prognosis for inpatients with emerging infectious diseases. It learns to embed the COVID-19-related medical features based on massive existing EMR data via transfer learning. The transferred parameters are further trained to imitate the teacher model's representation behavior based on knowledge distillation, which embeds the health status more comprehensively in the source dataset. We conduct the length of stay prediction experiments for patients on a real-world COVID-19 dataset. The experiment results indicate that our proposed model consistently outperforms the comparative baseline methods. CovidCare also reveals that, 1) hs-cTnI, hs-CRP and Platelet Counts are the most fatal biomarkers, whose abnormal values usually indicate emergency adverse outcome. 2) Normal values of gamma-GT, AP and eGFR indicate the overall improvement of health. The medical findings extracted by CovidCare are empirically confirmed by human experts and medical literatures.
Deep learning based models are relatively large, and it is hard to deploy such models on resource-limited devices such as mobile phones and embedded devices. One possible solution is knowledge distillation whereby a smaller model (student model) is trained by utilizing the information from a larger model (teacher model). In this paper, we present a survey of knowledge distillation techniques applied to deep learning models. To compare the performances of different techniques, we propose a new metric called distillation metric. Distillation metric compares different knowledge distillation algorithms based on sizes and accuracy scores. Based on the survey, some interesting conclusions are drawn and presented in this paper.
2D/3D human pose estimation is needed to develop novel intelligent tools for the operating room that can analyze and support the clinical activities. The lack of annotated data and the complexity of state-of-the-art pose estimation approaches limit, however, the deployment of such techniques inside the OR. In this work, we propose to use knowledge distillation in a teacher/student framework to harness the knowledge present in a large-scale non-annotated dataset and in an accurate but complex multi-stage teacher network to train a lightweight network for joint 2D/3D pose estimation. The teacher network also exploits the unlabeled data to generate both hard and soft labels useful in improving the student predictions. The easily deployable network trained using this effective self-supervision strategy performs on par with the teacher network on \emph{MVOR+}, an extension of the public MVOR dataset where all persons have been fully annotated, thus providing a viable solution for real-time 2D/3D human pose estimation in the OR.
Defocus Blur Detection(DBD) aims to separate in-focus and out-of-focus regions from a single image pixel-wisely. This task has been paid much attention since bokeh effects are widely used in digital cameras and smartphone photography. However, identifying obscure homogeneous regions and borderline transitions in partially defocus images is still challenging. To solve these problems, we introduce depth information into DBD for the first time. When the camera parameters are fixed, we argue that the accuracy of DBD is highly related to scene depth. Hence, we consider the depth information as the approximate soft label of DBD and propose a joint learning framework inspired by knowledge distillation. In detail, we learn the defocus blur from ground truth and the depth distilled from a well-trained depth estimation network at the same time. Thus, the sharp region will provide a strong prior for depth estimation while the blur detection also gains benefits from the distilled depth. Besides, we propose a novel decoder in the fully convolutional network(FCN) as our network structure. In each level of the decoder, we design the Selective Reception Field Block(SRFB) for merging multi-scale features efficiently and reuse the side outputs as Supervision-guided Attention Block(SAB). Unlike previous methods, the proposed decoder builds reception field pyramids and emphasizes salient regions simply and efficiently. Experiments show that our approach outperforms 11 other state-of-the-art methods on two popular datasets. Our method also runs at over 30 fps on a single GPU, which is 2x faster than previous works. The code is available at: https://github.com/vinthony/depth-distillation
Prior works in cross-lingual named entity recognition (NER) with no/little labeled data fall into two primary categories: model transfer based and data transfer based methods. In this paper we find that both method types can complement each other, in the sense that, the former can exploit context information via language-independent features but sees no task-specific information in the target language; while the latter generally generates pseudo target-language training data via translation but its exploitation of context information is weakened by inaccurate translations. Moreover, prior works rarely leverage unlabeled data in the target language, which can be effortlessly collected and potentially contains valuable information for improved results. To handle both problems, we propose a novel approach termed UniTrans to Unify both model and data Transfer for cross-lingual NER, and furthermore, to leverage the available information from unlabeled target-language data via enhanced knowledge distillation. We evaluate our proposed UniTrans over 4 target languages on benchmark datasets. Our experimental results show that it substantially outperforms the existing state-of-the-art methods.
Learning generic representations with deep networks requires massive training samples and significant computer resources. To learn a new specific task, an important issue is to transfer the generic teacher's representation to a student network. In this paper, we propose to use a metric between representations that is based on a functional view of neurons. We use optimal transport to quantify the match between two representations, yielding a distance that embeds some invariances inherent to the representation of deep networks. This distance defines a regularizer promoting the similarity of the student's representation with that of the teacher. Our approach can be used in any learning context where representation transfer is applicable. We experiment here on two standard settings: inductive transfer learning, where the teacher's representation is transferred to a student network of same architecture for a new related task, and knowledge distillation, where the teacher's representation is transferred to a student of simpler architecture for the same task (model compression). Our approach also lends itself to solving new learning problems; we demonstrate this by showing how to directly transfer the teacher's representation to a simpler architecture student for a new related task.
Unsupervised domain adaptation (UDA) seeks to alleviate the problem of domain shift between the distribution of unlabeled data from the target domain w.r.t. labeled data from the source domain. While the single-target UDA scenario is well studied in the literature, Multi-Target Domain Adaptation (MTDA) remains largely unexplored despite its practical importance, e.g., in multi-camera video-surveillance applications. The MTDA problem can be addressed by adapting one specialized model per target domain, although this solution is too costly in many real-world applications. Blending multiple targets for MTDA has been proposed, yet this solution may lead to a reduction in model specificity and accuracy. In this paper, we propose a novel unsupervised MTDA approach to train a CNN that can generalize well across multiple target domains. Our Multi-Teacher MTDA (MT-MTDA) method relies on multi-teacher knowledge distillation (KD) to iteratively distill target domain knowledge from multiple teachers to a common student. The KD process is performed in a progressive manner, where the student is trained by each teacher on how to perform UDA for a specific target, instead of directly learning domain adapted features. Finally, instead of combining the knowledge from each teacher, MT-MTDA alternates between teachers that distill knowledge, thereby preserving the specificity of each target (teacher) when learning to adapt to the student. MT-MTDA is compared against state-of-the-art methods on several challenging UDA benchmarks, and empirical results show that our proposed model can provide a considerably higher level of accuracy across multiple target domains. Our code is available at: https://github.com/LIVIAETS/MT-MTDA
One-class novelty detection is to identify anomalous instances that do not conform to the expected normal instances. In this paper, the Generative Adversarial Networks (GANs) based on encoder-decoder-encoder pipeline are used for detection and achieve state-of-the-art performance. However, deep neural networks are too over-parameterized to deploy on resource-limited devices. Therefore, Progressive Knowledge Distillation with GANs (PKDGAN) is proposed to learn compact and fast novelty detection networks. The P-KDGAN is a novel attempt to connect two standard GANs by the designed distillation loss for transferring knowledge from the teacher to the student. The progressive learning of knowledge distillation is a two-step approach that continuously improves the performance of the student GAN and achieves better performance than single step methods. In the first step, the student GAN learns the basic knowledge totally from the teacher via guiding of the pretrained teacher GAN with fixed weights. In the second step, joint fine-training is adopted for the knowledgeable teacher and student GANs to further improve the performance and stability. The experimental results on CIFAR-10, MNIST, and FMNIST show that our method improves the performance of the student GAN by 2.44%, 1.77%, and 1.73% when compressing the computation at ratios of 24.45:1, 311.11:1, and 700:1, respectively.
Multi-task learning (MTL) is to learn one single model that performs multiple tasks for achieving good performance on all tasks and lower cost on computation. Learning such a model requires to jointly optimize losses of a set of tasks with different difficulty levels, magnitudes, and characteristics (e.g. cross-entropy, Euclidean loss), leading to the imbalance problem in multi-task learning. To address the imbalance problem, we propose a knowledge distillation based method in this work. We first learn a task-specific model for each task. We then learn the multi-task model for minimizing task-specific loss and for producing the same feature with task-specific models. As the task-specific network encodes different features, we introduce small task-specific adaptors to project multi-task features to the task-specific features. In this way, the adaptors align the task-specific feature and the multi-task feature, which enables a balanced parameter sharing across tasks. Extensive experimental results demonstrate that our method can optimize a multi-task learning model in a more balanced way and achieve better overall performance.
Convolutional Neural Networks (CNNs) do not have a predictable recognition behavior with respect to the input resolution change. This prevents the feasibility of deployment on different input image resolutions for a specific model. To achieve efficient and flexible image classification at runtime, we employ meta learners to generate convolutional weights of main networks for various input scales and maintain privatized Batch Normalization layers per scale. For improved training performance, we further utilize knowledge distillation on the fly over model predictions based on different input resolutions. The learned meta network could dynamically parameterize main networks to act on input images of arbitrary size with consistently better accuracy compared to individually trained models. Extensive experiments on the ImageNet demonstrate that our method achieves an improved accuracy-efficiency trade-off during the adaptive inference process. By switching executable input resolutions, our method could satisfy the requirement of fast adaption in different resource-constrained environments. Code and models are available at https://github.com/d-li14/SAN.
This paper focuses on Semi-Supervised Object Detection (SSOD). Knowledge Distillation (KD) has been widely used for semi-supervised image classification. However, adapting these methods for SSOD has the following obstacles. (1) The teacher model serves a dual role as a teacher and a student, such that the teacher predictions on unlabeled images may be very close to those of student, which limits the upper-bound of the student. (2) The class imbalance issue in SSOD hinders an efficient knowledge transfer from teacher to student. To address these problems, we propose a novel method Temporal Self-Ensembling Teacher (TSE-T) for SSOD. Differently from previous KD based methods, we devise a temporally evolved teacher model. First, our teacher model ensembles its temporal predictions for unlabeled images under stochastic perturbations. Second, our teacher model ensembles its temporal model weights with the student model weights by an exponential moving average (EMA) which allows the teacher gradually learn from the student. These self-ensembling strategies increase data and model diversity, thus improving teacher predictions on unlabeled images. Finally, we use focal loss to formulate consistency regularization term to handle the data imbalance problem, which is a more efficient manner to utilize the useful information from unlabeled images than a simple hard-thresholding method which solely preserves confident predictions. Evaluated on the widely used VOC and COCO benchmarks, the mAP of our method has achieved 80.73% and 40.52% on the VOC2007 test set and the COCO2014 minval5k set respectively, which outperforms a strong fully-supervised detector by 2.37% and 1.49%. Furthermore, our method sets the new state-of-the-art in SSOD on VOC2007 test set which outperforms the baseline SSOD method by 1.44%. The source code of this work is publicly available at http://github.com/syangdong/tse-t.
Lipreading has witnessed a lot of progress due to the resurgence of neural networks. Recent works have placed emphasis on aspects such as improving performance by finding the optimal architecture or improving generalization. However, there is still a significant gap between the current methodologies and the requirements for an effective deployment of lipreading in practical scenarios. In this work, we propose a series of innovations that significantly bridge that gap: first, we raise the state-of-the-art performance by a wide margin on LRW and LRW-1000 to 88.5% and 46.6%, respectively using self-distillation. Secondly, we propose a series of architectural changes, including a novel Depthwise Separable Temporal Convolutional Network (DS-TCN) head, that slashes the computational cost to a fraction of the (already quite efficient) original model. Thirdly, we show that knowledge distillation is a very effective tool for recovering performance of the lightweight models. This results in a range of models with different accuracy-efficiency trade-offs. However, our most promising lightweight models are on par with the current state-of-the-art while showing a reduction of 8.2x and 3.9x in terms of computational cost and number of parameters, respectively, which we hope will enable the deployment of lipreading models in practical applications.
Medical image annotations are prohibitively time-consuming and expensive to obtain. To alleviate annotation scarcity, many approaches have been developed to efficiently utilize extra information, e.g.,semi-supervised learning further exploring plentiful unlabeled data, domain adaptation including multi-modality learning and unsupervised domain adaptation resorting to the prior knowledge from additional modality. In this paper, we aim to investigate the feasibility of simultaneously leveraging abundant unlabeled data and well-established cross-modality data for annotation-efficient medical image segmentation. To this end, we propose a novel semi-supervised domain adaptation approach, namely Dual-Teacher, where the student model not only learns from labeled target data (e.g., CT), but also explores unlabeled target data and labeled source data (e.g., MR) by two teacher models. Specifically, the student model learns the knowledge of unlabeled target data from intra-domain teacher by encouraging prediction consistency, as well as the shape priors embedded in labeled source data from inter-domain teacher via knowledge distillation. Consequently, the student model can effectively exploit the information from all three data resources and comprehensively integrate them to achieve improved performance. We conduct extensive experiments on MM-WHS 2017 dataset and demonstrate that our approach is able to concurrently utilize unlabeled data and cross-modality data with superior performance, outperforming semi-supervised learning and domain adaptation methods with a large margin.
Research on continual learning has led to a variety of approaches to mitigating catastrophic forgetting in feed-forward classification networks. Until now surprisingly little attention has been focused on continual learning of recurrent models applied to problems like image captioning. In this paper we take a systematic look at continual learning of LSTM-based models for image captioning. We propose an attention-based approach that explicitly accommodates the transient nature of vocabularies in continual image captioning tasks -- i.e. that task vocabularies are not disjoint. We call our method Recurrent Attention to Transient Tasks (RATT), and also show how to adapt continual learning approaches based on weight egularization and knowledge distillation to recurrent continual learning problems. We apply our approaches to incremental image captioning problem on two new continual learning benchmarks we define using the MS-COCO and Flickr30 datasets. Our results demonstrate that RATT is able to sequentially learn five captioning tasks while incurring no forgetting of previously learned ones.