Scaled dot-product attention applies a softmax function on the scaled dot-product of queries and keys to calculate weights and then multiplies the weights and values. In this work, we study how to improve the learning of scaled dot-product attention to improve the accuracy of DETR. Our method is based on the following observations: using ground truth foreground-background mask (GT Fg-Bg Mask) as additional cues in the weights/values learning enables learning much better weights/values; with better weights/values, better values/weights can be learned. We propose a triple-attention module in which the first attention is a plain scaled dot-product attention, the second/third attention generates high-quality weights/values (with the assistance of GT Fg-Bg Mask) and shares the values/weights with the first attention to improve the quality of values/weights. The second and third attentions are removed during inference. We call our method knowledge-sharing DETR (KS-DETR), which is an extension of knowledge distillation (KD) in the way that the improved weights and values of the teachers (the second and third attentions) are directly shared, instead of mimicked, by the student (the first attention) to enable more efficient knowledge transfer from the teachers to the student. Experiments on various DETR-like methods show consistent improvements over the baseline methods on the MS COCO benchmark. Code is available at https://github.com/edocanonymous/KS-DETR.
Deep models are susceptible to learning spurious correlations, even during the post-processing. We take a closer look at the knowledge distillation -- a popular post-processing technique for model compression -- and find that distilling with biased training data gives rise to a biased student, even when the teacher is debiased. To address this issue, we propose a simple knowledge distillation algorithm, coined DeTT (Debiasing by Teacher Transplanting). Inspired by a recent observation that the last neural net layer plays an overwhelmingly important role in debiasing, DeTT directly transplants the teacher's last layer to the student. Remaining layers are distilled by matching the feature map outputs of the student and the teacher, where the samples are reweighted to mitigate the dataset bias. Importantly, DeTT does not rely on the availability of extensive annotations on the bias-related attribute, which is typically not available during the post-processing phase. Throughout our experiments, DeTT successfully debiases the student model, consistently outperforming the baselines in terms of the worst-group accuracy.
Knowledge distillation is a common technique for improving the performance of a shallow student network by transferring information from a teacher network, which in general, is comparatively large and deep. These teacher networks are pre-trained and often uncalibrated, as no calibration technique is applied to the teacher model while training. Calibration of a network measures the probability of correctness for any of its predictions, which is critical in high-risk domains. In this paper, we study how to obtain a calibrated student from an uncalibrated teacher. Our approach relies on the fusion of the data-augmentation techniques, including but not limited to cutout, mixup, and CutMix, with knowledge distillation. We extend our approach beyond traditional knowledge distillation and find it suitable for Relational Knowledge Distillation and Contrastive Representation Distillation as well. The novelty of the work is that it provides a framework to distill a calibrated student from an uncalibrated teacher model without compromising the accuracy of the distilled student. We perform extensive experiments to validate our approach on various datasets, including CIFAR-10, CIFAR-100, CINIC-10 and TinyImageNet, and obtained calibrated student models. We also observe robust performance of our approach while evaluating it on corrupted CIFAR-100C data.
Knowledge distillation is an effective method for training lightweight vision models. However, acquiring teacher supervision for training samples is often costly, especially from large-scale models like vision transformers (ViTs). In this paper, we develop a simple framework to reduce the supervision cost of ViT distillation: masking out a fraction of input tokens given to the teacher. By masking input tokens, one can skip the computations associated with the masked tokens without requiring any change to teacher parameters or architecture. We find that masking patches with the lowest student attention scores is highly effective, saving up to 50% of teacher FLOPs without any drop in student accuracy, while other masking criterion leads to suboptimal efficiency gains. Through in-depth analyses, we reveal that the student-guided masking provides a good curriculum to the student, making teacher supervision easier to follow during the early stage and challenging in the later stage.
Facial forgery detection is a crucial but extremely challenging topic, with the fast development of forgery techniques making the synthetic artefact highly indistinguishable. Prior works show that by mining both spatial and frequency information the forgery detection performance of deep learning models can be vastly improved. However, leveraging multiple types of information usually requires more than one branch in the neural network, which makes the model heavy and cumbersome. Knowledge distillation, as an important technique for efficient modelling, could be a possible remedy. We find that existing knowledge distillation methods have difficulties distilling a dual-branch model into a single-branch model. More specifically, knowledge distillation on both the spatial and frequency branches has degraded performance than distillation only on the spatial branch. To handle such problem, we propose a novel two-in-one knowledge distillation framework which can smoothly merge the information from a large dual-branch network into a small single-branch network, with the help of different dedicated feature projectors and the gradient homogenization technique. Experimental analysis on two datasets, FaceForensics++ and Celeb-DF, shows that our proposed framework achieves superior performance for facial forgery detection with much fewer parameters.
Federated learning enables edge devices to train a global model collaboratively without exposing their data. Despite achieving outstanding advantages in computing efficiency and privacy protection, federated learning faces a significant challenge when dealing with non-IID data, i.e., data generated by clients that are typically not independent and identically distributed. In this paper, we tackle a new type of Non-IID data, called cluster-skewed non-IID, discovered in actual data sets. The cluster-skewed non-IID is a phenomenon in which clients can be grouped into clusters with similar data distributions. By performing an in-depth analysis of the behavior of a classification model's penultimate layer, we introduce a metric that quantifies the similarity between two clients' data distributions without violating their privacy. We then propose an aggregation scheme that guarantees equality between clusters. In addition, we offer a novel local training regularization based on the knowledge-distillation technique that reduces the overfitting problem at clients and dramatically boosts the training scheme's performance. We theoretically prove the superiority of the proposed aggregation over the benchmark FedAvg. Extensive experimental results on both standard public datasets and our in-house real-world dataset demonstrate that the proposed approach improves accuracy by up to 16% compared to the FedAvg algorithm.
The rise of mobile AI accelerators allows latency-sensitive applications to execute lightweight Deep Neural Networks (DNNs) on the client side. However, critical applications require powerful models that edge devices cannot host and must therefore offload requests, where the high-dimensional data will compete for limited bandwidth. This work proposes shifting away from focusing on executing shallow layers of partitioned DNNs. Instead, it advocates concentrating the local resources on variational compression optimized for machine interpretability. We introduce a novel framework for resource-conscious compression models and extensively evaluate our method in an environment reflecting the asymmetric resource distribution between edge devices and servers. Our method achieves 60% lower bitrate than a state-of-the-art SC method without decreasing accuracy and is up to 16x faster than offloading with existing codec standards.
Despite recommender systems play a key role in network content platforms, mining the user's interests is still a significant challenge. Existing works predict the user interest by utilizing user behaviors, i.e., clicks, views, etc., but current solutions are ineffective when users perform unsettled activities. The latter ones involve new users, which have few activities of any kind, and sparse users who have low-frequency behaviors. We uniformly describe both these user-types as "cold users", which are very common but often neglected in network content platforms. To address this issue, we enhance the representation of the user interest by combining his social interest, e.g., friendship, following bloggers, interest groups, etc., with the activity behaviors. Thus, in this work, we present a novel algorithm entitled SocialNet, which adopts a two-stage method to progressively extract the coarse-grained and fine-grained social interest. Our technique then concatenates SocialNet's output with the original user representation to get the final user representation that combines behavior interests and social interests. Offline experiments on Tencent video's recommender system demonstrate the superiority over the baseline behavior-based model. The online experiment also shows a significant performance improvement in clicks and view time in the real-world recommendation system. The source code is available at https://github.com/Social4Rec/SocialNet.
Self-supervised speech pre-training enables deep neural network models to capture meaningful and disentangled factors from raw waveform signals. The learned universal speech representations can then be used across numerous downstream tasks. These representations, however, are sensitive to distribution shifts caused by environmental factors, such as noise and/or room reverberation. Their large sizes, in turn, make them unfeasible for edge applications. In this work, we propose a knowledge distillation methodology termed RobustDistiller which compresses universal representations while making them more robust against environmental artifacts via a multi-task learning objective. The proposed layer-wise distillation recipe is evaluated on top of three well-established universal representations, as well as with three downstream tasks. Experimental results show the proposed methodology applied on top of the WavLM Base+ teacher model outperforming all other benchmarks across noise types and levels, as well as reverberation times. Oftentimes, the obtained results with the student model (24M parameters) achieved results inline with those of the teacher model (95M).
Knowledge distillation has been shown to be a powerful model compression approach to facilitate the deployment of pre-trained language models in practice. This paper focuses on task-agnostic distillation. It produces a compact pre-trained model that can be easily fine-tuned on various tasks with small computational costs and memory footprints. Despite the practical benefits, task-agnostic distillation is challenging. Since the teacher model has a significantly larger capacity and stronger representation power than the student model, it is very difficult for the student to produce predictions that match the teacher's over a massive amount of open-domain training data. Such a large prediction discrepancy often diminishes the benefits of knowledge distillation. To address this challenge, we propose Homotopic Distillation (HomoDistil), a novel task-agnostic distillation approach equipped with iterative pruning. Specifically, we initialize the student model from the teacher model, and iteratively prune the student's neurons until the target width is reached. Such an approach maintains a small discrepancy between the teacher's and student's predictions throughout the distillation process, which ensures the effectiveness of knowledge transfer. Extensive experiments demonstrate that HomoDistil achieves significant improvements on existing baselines.
Liver transplant is an essential therapy performed for severe liver diseases. The fact of scarce liver resources makes the organ assigning crucial. Model for End-stage Liver Disease (MELD) score is a widely adopted criterion when making organ distribution decisions. However, it ignores post-transplant outcomes and organ/donor features. These limitations motivate the emergence of machine learning (ML) models. Unfortunately, ML models could be unfair and trigger bias against certain groups of people. To tackle this problem, this work proposes a fair machine learning framework targeting graft failure prediction in liver transplant. Specifically, knowledge distillation is employed to handle dense and sparse features by combining the advantages of tree models and neural networks. A two-step debiasing method is tailored for this framework to enhance fairness. Experiments are conducted to analyze unfairness issues in existing models and demonstrate the superiority of our method in both prediction and fairness performance.
In autonomous driving, the novel objects and lack of annotations challenge the traditional 3D LiDAR semantic segmentation based on deep learning. Few-shot learning is a feasible way to solve these issues. However, currently few-shot semantic segmentation methods focus on camera data, and most of them only predict the novel classes without considering the base classes. This setting cannot be directly applied to autonomous driving due to safety concerns. Thus, we propose a few-shot 3D LiDAR semantic segmentation method that predicts both novel classes and base classes simultaneously. Our method tries to solve the background ambiguity problem in generalized few-shot semantic segmentation. We first review the original cross-entropy and knowledge distillation losses, then propose a new loss function that incorporates the background information to achieve 3D LiDAR few-shot semantic segmentation. Extensive experiments on SemanticKITTI demonstrate the effectiveness of our method.
Data-free knowledge distillation is a challenging model lightweight task for scenarios in which the original dataset is not available. Previous methods require a lot of extra computational costs to update one or more generators and their naive imitate-learning lead to lower distillation efficiency. Based on these observations, we first propose an efficient unlabeled sample selection method to replace high computational generators and focus on improving the training efficiency of the selected samples. Then, a class-dropping mechanism is designed to suppress the label noise caused by the data domain shifts. Finally, we propose a distillation method that incorporates explicit features and implicit structured relations to improve the effect of distillation. Experimental results show that our method can quickly converge and obtain higher accuracy than other state-of-the-art methods.
Currently, one of the major challenges in deep learning-based video frame interpolation (VFI) is the large model sizes and high computational complexity associated with many high performance VFI approaches. In this paper, we present a distillation-based two-stage workflow for obtaining compressed VFI models which perform competitively to the state of the arts, at a greatly reduced model size and complexity. Specifically, an optimisation-based network pruning method is first applied to a recently proposed frame interpolation model, ST-MFNet, which outperforms many other VFI methods but suffers from large model size. The resulting new network architecture achieves a 91% reduction in parameters and 35% increase in speed. Secondly, the performance of the new network is further enhanced through a teacher-student knowledge distillation training process using a Laplacian distillation loss. The final low complexity model, ST-MFNet Mini, achieves a comparable performance to most existing high-complex VFI methods, only outperformed by the original ST-MFNet. Our source code is available at https://github.com/crispianm/ST-MFNet-Mini
Large-scale language-agnostic sentence embedding models such as LaBSE (Feng et al., 2022) obtain state-of-the-art performance for parallel sentence alignment. However, these large-scale models can suffer from inference speed and computation overhead. This study systematically explores learning language-agnostic sentence embeddings with lightweight models. We demonstrate that a thin-deep encoder can construct robust low-dimensional sentence embeddings for 109 languages. With our proposed distillation methods, we achieve further improvements by incorporating knowledge from a teacher model. Empirical results on Tatoeba, United Nations, and BUCC show the effectiveness of our lightweight models. We release our lightweight language-agnostic sentence embedding models LEALLA on TensorFlow Hub.
Knowledge distillation has been widely adopted in a variety of tasks and has achieved remarkable successes. Since its inception, many researchers have been intrigued by the dark knowledge hidden in the outputs of the teacher model. Recently, a study has demonstrated that knowledge distillation and label smoothing can be unified as learning from soft labels. Consequently, how to measure the effectiveness of the soft labels becomes an important question. Most existing theories have stringent constraints on the teacher model or data distribution, and many assumptions imply that the soft labels are close to the ground-truth labels. This paper studies whether biased soft labels are still effective. We present two more comprehensive indicators to measure the effectiveness of such soft labels. Based on the two indicators, we give sufficient conditions to ensure biased soft label based learners are classifier-consistent and ERM learnable. The theory is applied to three weakly-supervised frameworks. Experimental results validate that biased soft labels can also teach good students, which corroborates the soundness of the theory.
High-order Takagi-Sugeno-Kang (TSK) fuzzy classifiers possess powerful classification performance yet have fewer fuzzy rules, but always be impaired by its exponential growth training time and poorer interpretability owing to High-order polynomial used in consequent part of fuzzy rule, while Low-order TSK fuzzy classifiers run quickly with high interpretability, however they usually require more fuzzy rules and perform relatively not very well. Address this issue, a novel TSK fuzzy classifier embeded with knowledge distillation in deep learning called HTSK-LLM-DKD is proposed in this study. HTSK-LLM-DKD achieves the following distinctive characteristics: 1) It takes High-order TSK classifier as teacher model and Low-order TSK fuzzy classifier as student model, and leverages the proposed LLM-DKD (Least Learning Machine based Decoupling Knowledge Distillation) to distill the fuzzy dark knowledge from High-order TSK fuzzy classifier to Low-order TSK fuzzy classifier, which resulting in Low-order TSK fuzzy classifier endowed with enhanced performance surpassing or at least comparable to High-order TSK classifier, as well as high interpretability; specifically 2) The Negative Euclidean distance between the output of teacher model and each class is employed to obtain the teacher logits, and then it compute teacher/student soft labels by the softmax function with distillating temperature parameter; 3) By reformulating the Kullback-Leibler divergence, it decouples fuzzy dark knowledge into target class knowledge and non-target class knowledge, and transfers them to student model. The advantages of HTSK-LLM-DKD are verified on the benchmarking UCI datasets and a real dataset Cleveland heart disease, in terms of classification performance and model interpretability.
To imitate the ability of keeping learning of human, continual learning which can learn from a never-ending data stream has attracted more interests recently. In all settings, the online class incremental learning (OCIL), where incoming samples from data stream can be used only once, is more challenging and can be encountered more frequently in real world. Actually, all continual learning models face a stability-plasticity dilemma, where the stability means the ability to preserve old knowledge while the plasticity denotes the ability to incorporate new knowledge. Although replay-based methods have shown exceptional promise, most of them concentrate on the strategy for updating and retrieving memory to keep stability at the expense of plasticity. To strike a preferable trade-off between stability and plasticity, we propose an Adaptive Focus Shifting algorithm (AFS), which dynamically adjusts focus to ambiguous samples and non-target logits in model learning. Through a deep analysis of the task-recency bias caused by class imbalance, we propose a revised focal loss to mainly keep stability. \Rt{By utilizing a new weight function, the revised focal loss will pay more attention to current ambiguous samples, which are the potentially valuable samples to make model progress quickly.} To promote plasticity, we introduce a virtual knowledge distillation. By designing a virtual teacher, it assigns more attention to non-target classes, which can surmount overconfidence and encourage model to focus on inter-class information. Extensive experiments on three popular datasets for OCIL have shown the effectiveness of AFS. The code will be available at \url{https://github.com/czjghost/AFS}.
The increasing intensity and frequency of floods is one of the many consequences of our changing climate. In this work, we explore ML techniques that improve the flood detection module of an operational early flood warning system. Our method exploits an unlabelled dataset of paired multi-spectral and Synthetic Aperture Radar (SAR) imagery to reduce the labeling requirements of a purely supervised learning method. Prior works have used unlabelled data by creating weak labels out of them. However, from our experiments we noticed that such a model still ends up learning the label mistakes in those weak labels. Motivated by knowledge distillation and semi supervised learning, we explore the use of a teacher to train a student with the help of a small hand labelled dataset and a large unlabelled dataset. Unlike the conventional self distillation setup, we propose a cross modal distillation framework that transfers supervision from a teacher trained on richer modality (multi-spectral images) to a student model trained on SAR imagery. The trained models are then tested on the Sen1Floods11 dataset. Our model outperforms the Sen1Floods11 baseline model trained on the weak labeled SAR imagery by an absolute margin of 6.53% Intersection-over-Union (IoU) on the test split.
In this paper, we present offline-to-online knowledge distillation (OOKD) for video instance segmentation (VIS), which transfers a wealth of video knowledge from an offline model to an online model for consistent prediction. Unlike previous methods that having adopting either an online or offline model, our single online model takes advantage of both models by distilling offline knowledge. To transfer knowledge correctly, we propose query filtering and association (QFA), which filters irrelevant queries to exact instances. Our KD with QFA increases the robustness of feature matching by encoding object-centric features from a single frame supplemented by long-range global information. We also propose a simple data augmentation scheme for knowledge distillation in the VIS task that fairly transfers the knowledge of all classes into the online model. Extensive experiments show that our method significantly improves the performance in video instance segmentation, especially for challenging datasets including long, dynamic sequences. Our method also achieves state-of-the-art performance on YTVIS-21, YTVIS-22, and OVIS datasets, with mAP scores of 46.1%, 43.6%, and 31.1%, respectively.