Recently, much progress has been made for self-supervised action recognition. Most existing approaches emphasize the contrastive relations among videos, including appearance and motion consistency. However, two main issues remain for existing pre-training methods: 1) the learned representation is neutral and not informative for a specific task; 2) multi-task learning-based pre-training sometimes leads to sub-optimal solutions due to inconsistent domains of different tasks. To address the above issues, we propose a novel action recognition pre-training framework, which exploits human-centered prior knowledge that generates more informative representation, and avoids the conflict between multiple tasks by using task-dependent representations. Specifically, we distill knowledge from a human parsing model to enrich the semantic capability of representation. In addition, we combine knowledge distillation with contrastive learning to constitute a task-dependent multi-task framework. We achieve state-of-the-art performance on two popular benchmarks for action recognition task, i.e., UCF101 and HMDB51, verifying the effectiveness of our method.
Modern recommender systems operate in a fully server-based fashion. To cater to millions of users, the frequent model maintaining and the high-speed processing for concurrent user requests are required, which comes at the cost of a huge carbon footprint. Meanwhile, users need to upload their behavior data even including the immediate environmental context to the server, raising the public concern about privacy. On-device recommender systems circumvent these two issues with cost-conscious settings and local inference. However, due to the limited memory and computing resources, on-device recommender systems are confronted with two fundamental challenges: (1) how to reduce the size of regular models to fit edge devices? (2) how to retain the original capacity? Previous research mostly adopts tensor decomposition techniques to compress the regular recommendation model with limited compression ratio so as to avoid drastic performance degradation. In this paper, we explore ultra-compact models for next-item recommendation, by loosing the constraint of dimensionality consistency in tensor decomposition. Meanwhile, to compensate for the capacity loss caused by compression, we develop a self-supervised knowledge distillation framework which enables the compressed model (student) to distill the essential information lying in the raw data, and improves the long-tail item recommendation through an embedding-recombination strategy with the original model (teacher). The extensive experiments on two benchmarks demonstrate that, with 30x model size reduction, the compressed model almost comes with no accuracy loss, and even outperforms its uncompressed counterpart in most cases.
Prototypical methods have recently gained a lot of attention due to their intrinsic interpretable nature, which is obtained through the prototypes. With growing use cases of model reuse and distillation, there is a need to also study transfer of interpretability from one model to another. We present Proto2Proto, a novel method to transfer interpretability of one prototypical part network to another via knowledge distillation. Our approach aims to add interpretability to the "dark" knowledge transferred from the teacher to the shallower student model. We propose two novel losses: "Global Explanation" loss and "Patch-Prototype Correspondence" loss to facilitate such a transfer. Global Explanation loss forces the student prototypes to be close to teacher prototypes, and Patch-Prototype Correspondence loss enforces the local representations of the student to be similar to that of the teacher. Further, we propose three novel metrics to evaluate the student's proximity to the teacher as measures of interpretability transfer in our settings. We qualitatively and quantitatively demonstrate the effectiveness of our method on CUB-200-2011 and Stanford Cars datasets. Our experiments show that the proposed method indeed achieves interpretability transfer from teacher to student while simultaneously exhibiting competitive performance.
The outpouring of various pre-trained models empowers knowledge distillation by providing abundant teacher resources, but there lacks a developed mechanism to utilize these teachers adequately. With a massive model repository composed of teachers pre-trained on diverse tasks, we must surmount two obstacles when using knowledge distillation to learn a new task. First, given a fixed computing budget, it is not affordable to try each teacher and train the student repeatedly, making it necessary to seek out the most contributive teacher precisely and efficiently. Second, semantic gaps exist between the teachers and the target student since they are trained on different tasks. Thus, we need to extract knowledge from a general label space that may be different from the student's. Faced with these two challenges, we study a new setting named selective cross-task distillation that includes teacher assessment and generalized knowledge reuse. We bridge the teacher's label space and the student's label space through optimal transport. The transportation cost from the teacher's prediction to the student's prediction measures the relatedness between two tasks and acts as an objective for distillation. Our method reuses cross-task knowledge from a distinct label space and efficiently assesses teachers without enumerating the model repository. Experiments demonstrate the effectiveness of our proposed method.
Structured pruning compresses neural networks by reducing channels (filters) for fast inference and low footprint at run-time. To restore accuracy after pruning, fine-tuning is usually applied to pruned networks. However, too few remaining parameters in pruned networks inevitably bring a great challenge to fine-tuning to restore accuracy. To address this challenge, we propose a novel method that first linearly over-parameterizes the compact layers in pruned networks to enlarge the number of fine-tuning parameters and then re-parameterizes them to the original layers after fine-tuning. Specifically, we equivalently expand the convolution/linear layer with several consecutive convolution/linear layers that do not alter the current output feature maps. Furthermore, we utilize similarity-preserving knowledge distillation that encourages the over-parameterized block to learn the immediate data-to-data similarities of the corresponding dense layer to maintain its feature learning ability. The proposed method is comprehensively evaluated on CIFAR-10 and ImageNet which significantly outperforms the vanilla fine-tuning strategy, especially for large pruning ratio.
Face recognition for visible light (VIS) images achieve high accuracy thanks to the recent development of deep learning. However, heterogeneous face recognition (HFR), which is a face matching in different domains, is still a difficult task due to the domain discrepancy and lack of large HFR dataset. Several methods have attempted to reduce the domain discrepancy by means of fine-tuning, which causes significant degradation of the performance in the VIS domain because it loses the highly discriminative VIS representation. To overcome this problem, we propose joint feature distribution alignment learning (JFDAL) which is a joint learning approach utilizing knowledge distillation. It enables us to achieve high HFR performance with retaining the original performance for the VIS domain. Extensive experiments demonstrate that our proposed method delivers statistically significantly better performances compared with the conventional fine-tuning approach on a public HFR dataset Oulu-CASIA NIR&VIS and popular verification datasets in VIS domain such as FLW, CFP, AgeDB. Furthermore, comparative experiments with existing state-of-the-art HFR methods show that our method achieves a comparable HFR performance on the Oulu-CASIA NIR&VIS dataset with less degradation of VIS performance.
Due to the prosperity of Artificial Intelligence (AI) techniques, more and more backdoors are designed by adversaries to attack Deep Neural Networks (DNNs).Although the state-of-the-art method Neural Attention Distillation (NAD) can effectively erase backdoor triggers from DNNs, it still suffers from non-negligible Attack Success Rate (ASR) together with lowered classification ACCuracy (ACC), since NAD focuses on backdoor defense using attention features (i.e., attention maps) of the same order. In this paper, we introduce a novel backdoor defense framework named Attention Relation Graph Distillation (ARGD), which fully explores the correlation among attention features with different orders using our proposed Attention Relation Graphs (ARGs). Based on the alignment of ARGs between both teacher and student models during knowledge distillation, ARGD can eradicate more backdoor triggers than NAD. Comprehensive experimental results show that, against six latest backdoor attacks, ARGD outperforms NAD by up to 94.85% reduction in ASR, while ACC can be improved by up to 3.23%.
Segmenting unseen objects is a crucial ability for the robot since it may encounter new environments during the operation. Recently, a popular solution is leveraging RGB-D features of large-scale synthetic data and directly applying the model to unseen real-world scenarios. However, the domain shift caused by the sim2real gap is inevitable, posing a crucial challenge to the segmentation model. In this paper, we emphasize the adaptation process across sim2real domains and model it as a learning problem on the BatchNorm parameters of a simulation-trained model. Specifically, we propose a novel non-parametric entropy objective, which formulates the learning objective for the test-time adaptation in an open-world manner. Then, a cross-modality knowledge distillation objective is further designed to encourage the test-time knowledge transfer for feature enhancement. Our approach can be efficiently implemented with only test images, without requiring annotations or revisiting the large-scale synthetic training data. Besides significant time savings, the proposed method consistently improves segmentation results on the overlap and boundary metrics, achieving state-of-the-art performance on unseen object instance segmentation.
Unsupervised person re-identification is a challenging and promising task in computer vision. Nowadays unsupervised person re-identification methods have achieved great progress by training with pseudo labels. However, how to purify feature and label noise is less explicitly studied in the unsupervised manner. To purify the feature, we take into account two types of additional features from different local views to enrich the feature representation. The proposed multi-view features are carefully integrated into our cluster contrast learning to leverage more discriminative cues that the global feature easily ignored and biased. To purify the label noise, we propose to take advantage of the knowledge of teacher model in an offline scheme. Specifically, we first train a teacher model from noisy pseudo labels, and then use the teacher model to guide the learning of our student model. In our setting, the student model could converge fast with the supervision of the teacher model thus reduce the interference of noisy labels as the teacher model greatly suffered. After carefully handling the noise and bias in the feature learning, our purification modules are proven to be very effective for unsupervised person re-identification. Extensive experiments on three popular person re-identification datasets demonstrate the superiority of our method. Especially, our approach achieves a state-of-the-art accuracy 85.8\% @mAP and 94.5\% @Rank-1 on the challenging Market-1501 benchmark with ResNet-50 under the fully unsupervised setting. The code will be released.
Real-time 6D object pose estimation is essential for many real-world applications, such as robotic grasping and augmented reality. To achieve an accurate object pose estimation from RGB images in real-time, we propose an effective and lightweight model, namely High-Resolution 6D Pose Estimation Network (HRPose). We adopt the efficient and small HRNetV2-W18 as a feature extractor to reduce computational burdens while generating accurate 6D poses. With only 33\% of the model size and lower computational costs, our HRPose achieves comparable performance compared with state-of-the-art models. Moreover, by transferring knowledge from a large model to our proposed HRPose through output and feature-similarity distillations, the performance of our HRPose is improved in effectiveness and efficiency. Numerical experiments on the widely-used benchmark LINEMOD demonstrate the superiority of our proposed HRPose against state-of-the-art methods.
Task-oriented dialogue generation is challenging since the underlying knowledge is often dynamic and effectively incorporating knowledge into the learning process is hard. It is particularly challenging to generate both human-like and informative responses in this setting. Recent research primarily focused on various knowledge distillation methods where the underlying relationship between the facts in a knowledge base is not effectively captured. In this paper, we go one step further and demonstrate how the structural information of a knowledge graph can improve the system's inference capabilities. Specifically, we propose DialoKG, a novel task-oriented dialogue system that effectively incorporates knowledge into a language model. Our proposed system views relational knowledge as a knowledge graph and introduces (1) a structure-aware knowledge embedding technique, and (2) a knowledge graph-weighted attention masking strategy to facilitate the system selecting relevant information during the dialogue generation. An empirical evaluation demonstrates the effectiveness of DialoKG over state-of-the-art methods on several standard benchmark datasets.
Knowledge distillation is an approach to transfer information on representations from a teacher to a student by reducing their difference. A challenge of this approach is to reduce the flexibility of the student's representations inducing inaccurate learning of the teacher's knowledge. To resolve it in transferring, we investigate distillation of structures of representations specified to three types: intra-feature, local inter-feature, global inter-feature structures. To transfer them, we introduce feature structure distillation methods based on the Centered Kernel Alignment, which assigns a consistent value to similar features structures and reveals more informative relations. In particular, a memory-augmented transfer method with clustering is implemented for the global structures. The methods are empirically analyzed on the nine tasks for language understanding of the GLUE dataset with Bidirectional Encoder Representations from Transformers (BERT), which is a representative neural language model. In the results, the proposed methods effectively transfer the three types of structures and improve performance compared to state-of-the-art distillation methods. Indeed, the code for the methods is available in https://github.com/maroo-sky/FSD.
Despite the recent advances in the field of object detection, common architectures are still ill-suited to incrementally detect new categories over time. They are vulnerable to catastrophic forgetting: they forget what has been already learned while updating their parameters in absence of the original training data. Previous works extended standard classification methods in the object detection task, mainly adopting the knowledge distillation framework. However, we argue that object detection introduces an additional problem, which has been overlooked. While objects belonging to new classes are learned thanks to their annotations, if no supervision is provided for other objects that may still be present in the input, the model learns to associate them to background regions. We propose to handle these missing annotations by revisiting the standard knowledge distillation framework. Our approach outperforms current state-of-the-art methods in every setting of the Pascal-VOC dataset. We further propose an extension to instance segmentation, outperforming the other baselines. Code can be found here: https://github.com/fcdl94/MMA
An electrocardiogram (ECG) monitors the electrical activity generated by the heart and is used to detect fatal cardiovascular diseases (CVDs). Conventionally, to capture the precise electrical activity, clinical experts use multiple-lead ECGs (typically 12 leads). But in recent times, large-size deep learning models have been used to detect these diseases. However, such models require heavy compute resources like huge memory and long inference time. To alleviate these shortcomings, we propose a low-parameter model, named Low Resource Heart-Network (LRH-Net), which uses fewer leads to detect ECG anomalies in a resource-constrained environment. A multi-level knowledge distillation process is used on top of that to get better generalization performance on our proposed model. The multi-level knowledge distillation process distills the knowledge to LRH-Net trained on a reduced number of leads from higher parameter (teacher) models trained on multiple leads to reduce the performance gap. The proposed model is evaluated on the PhysioNet-2020 challenge dataset with constrained input. The parameters of the LRH-Net are 106x less than our teacher model for detecting CVDs. The performance of the LRH-Net was scaled up to 3.2% and the inference time scaled down by 75% compared to the teacher model. In contrast to the compute- and parameter-intensive deep learning techniques, the proposed methodology uses a subset of ECG leads using the low resource LRH-Net, making it eminently suitable for deployment on edge devices.
Pre-trained language models have demonstrated superior performance in various natural language processing tasks. However, these models usually contain hundreds of millions of parameters, which limits their practicality because of latency requirements in real-world applications. Existing methods train small compressed models via knowledge distillation. However, performance of these small models drops significantly compared with the pre-trained models due to their reduced model capacity. We propose MoEBERT, which uses a Mixture-of-Experts structure to increase model capacity and inference speed. We initialize MoEBERT by adapting the feed-forward neural networks in a pre-trained model into multiple experts. As such, representation power of the pre-trained model is largely retained. During inference, only one of the experts is activated, such that speed can be improved. We also propose a layer-wise distillation method to train MoEBERT. We validate the efficiency and effectiveness of MoEBERT on natural language understanding and question answering tasks. Results show that the proposed method outperforms existing task-specific distillation algorithms. For example, our method outperforms previous approaches by over 2% on the MNLI (mismatched) dataset. Our code is publicly available at https://github.com/SimiaoZuo/MoEBERT.
Knowledge distillation (KD) is an efficient framework for compressing large-scale pre-trained language models. Recent years have seen a surge of research aiming to improve KD by leveraging Contrastive Learning, Intermediate Layer Distillation, Data Augmentation, and Adversarial Training. In this work, we propose a learning based data augmentation technique tailored for knowledge distillation, called CILDA. To the best of our knowledge, this is the first time that intermediate layer representations of the main task are used in improving the quality of augmented samples. More precisely, we introduce an augmentation technique for KD based on intermediate layer matching using contrastive loss to improve masked adversarial data augmentation. CILDA outperforms existing state-of-the-art KD approaches on the GLUE benchmark, as well as in an out-of-domain evaluation.
We study multi-modal few-shot object detection (FSOD) in this paper, using both few-shot visual examples and class semantic information for detection, which are complementary to each other by definition. Most of the previous works on multi-modal FSOD are fine-tuning-based which are inefficient for online applications. Moreover, these methods usually require expertise like class names to extract class semantic embedding, which are hard to get for rare classes. Our approach is motivated by the high-level conceptual similarity of (metric-based) meta-learning and prompt-based learning to learn generalizable few-shot and zero-shot object detection models respectively without fine-tuning. Specifically, we combine the few-shot visual classifier and text classifier learned via meta-learning and prompt-based learning respectively to build the multi-modal classifier and detection models. In addition, to fully exploit the pre-trained language models, we propose meta-learning-based cross-modal prompting to generate soft prompts for novel classes present in few-shot visual examples, which are then used to learn the text classifier. Knowledge distillation is introduced to learn the soft prompt generator without using human prior knowledge of class names, which may not be available for rare classes. Our insight is that the few-shot support images naturally include related context information and semantics of the class. We comprehensively evaluate the proposed multi-modal FSOD models on multiple few-shot object detection benchmarks, achieving promising results.
Recognizing human locomotion intent and activities is important for controlling the wearable robots while walking in complex environments. However, human-robot interface signals are usually user-dependent, which causes that the classifier trained on source subjects performs poorly on new subjects. To address this issue, this paper designs the ensemble diverse hypotheses and knowledge distillation (EDHKD) method to realize unsupervised cross-subject adaptation. EDH mitigates the divergence between labeled data of source subjects and unlabeled data of target subjects to accurately classify the locomotion modes of target subjects without labeling data. Compared to previous domain adaptation methods based on the single learner, which may only learn a subset of features from input signals, EDH can learn diverse features by incorporating multiple diverse feature generators and thus increases the accuracy and decreases the variance of classifying target data, but it sacrifices the efficiency. To solve this problem, EDHKD (student) distills the knowledge from the EDH (teacher) to a single network to remain efficient and accurate. The performance of the EDHKD is theoretically proved and experimentally validated on a 2D moon dataset and two public human locomotion datasets. Experimental results show that the EDHKD outperforms all other methods. The EDHKD can classify target data with 96.9%, 94.4%, and 97.4% average accuracy on the above three datasets with a short computing time (1 ms). Compared to a benchmark (BM) method, the EDHKD increases 1.3% and 7.1% average accuracy for classifying the locomotion modes of target subjects. The EDHKD also stabilizes the learning curves. Therefore, the EDHKD is significant for increasing the generalization ability and efficiency of the human intent prediction and human activity recognition system, which will improve human-robot interactions.
Current Knowledge Distillation (KD) methods for semantic segmentation often guide the student to mimic the teacher's structured information generated from individual data samples. However, they ignore the global semantic relations among pixels across various images that are valuable for KD. This paper proposes a novel Cross-Image Relational KD (CIRKD), which focuses on transferring structured pixel-to-pixel and pixel-to-region relations among the whole images. The motivation is that a good teacher network could construct a well-structured feature space in terms of global pixel dependencies. CIRKD makes the student mimic better structured semantic relations from the teacher, thus improving the segmentation performance. Experimental results over Cityscapes, CamVid and Pascal VOC datasets demonstrate the effectiveness of our proposed approach against state-of-the-art distillation methods. The code is available at https://github.com/winycg/CIRKD.
Weakly supervised object detection (WSOD), which is an effective way to train an object detection model using only image-level annotations, has attracted considerable attention from researchers. However, most of the existing methods, which are based on multiple instance learning (MIL), tend to localize instances to the discriminative parts of salient objects instead of the entire content of all objects. In this paper, we propose a WSOD framework called the Spatial Likelihood Voting with Self-knowledge Distillation Network (SLV-SD Net). In this framework, we introduce a spatial likelihood voting (SLV) module to converge region proposal localization without bounding box annotations. Specifically, in every iteration during training, all the region proposals in a given image act as voters voting for the likelihood of each category in the spatial dimensions. After dilating the alignment on the area with large likelihood values, the voting results are regularized as bounding boxes, which are then used for the final classification and localization. Based on SLV, we further propose a self-knowledge distillation (SD) module to refine the feature representations of the given image. The likelihood maps generated by the SLV module are used to supervise the feature learning of the backbone network, encouraging the network to attend to wider and more diverse areas of the image. Extensive experiments on the PASCAL VOC 2007/2012 and MS-COCO datasets demonstrate the excellent performance of SLV-SD Net. In addition, SLV-SD Net produces new state-of-the-art results on these benchmarks.