Existing state-of-the-art human pose estimation methods require heavy computational resources for accurate predictions. One promising technique to obtain an accurate yet lightweight pose estimator is knowledge distillation, which distills the pose knowledge from a powerful teacher model to a less-parameterized student model. However, existing pose distillation works rely on a heavy pre-trained estimator to perform knowledge transfer and require a complex two-stage learning procedure. In this work, we investigate a novel Online Knowledge Distillation framework by distilling Human Pose structure knowledge in a one-stage manner to guarantee the distillation efficiency, termed OKDHP. Specifically, OKDHP trains a single multi-branch network and acquires the predicted heatmaps from each, which are then assembled by a Feature Aggregation Unit (FAU) as the target heatmaps to teach each branch in reverse. Instead of simply averaging the heatmaps, FAU which consists of multiple parallel transformations with different receptive fields, leverages the multi-scale information, thus obtains target heatmaps with higher-quality. Specifically, the pixel-wise Kullback-Leibler (KL) divergence is utilized to minimize the discrepancy between the target heatmaps and the predicted ones, which enables the student network to learn the implicit keypoint relationship. Besides, an unbalanced OKDHP scheme is introduced to customize the student networks with different compression rates. The effectiveness of our approach is demonstrated by extensive experiments on two common benchmark datasets, MPII and COCO.
Achieving backward compatibility when rolling out new models can highly reduce costs or even bypass feature re-encoding of existing gallery images for in-production visual retrieval systems. Previous related works usually leverage losses used in knowledge distillation which can cause performance degradations or not guarantee compatibility. To address these issues, we propose a general framework called Learning Compatible Embeddings (LCE) which is applicable for both cross model compatibility and compatible training in direct/forward/backward manners. Our compatibility is achieved by aligning class centers between models directly or via a transformation, and restricting more compact intra-class distributions for the new model. Experiments are conducted in extensive scenarios such as changes of training dataset, loss functions, network architectures as well as feature dimensions, and demonstrate that LCE efficiently enables model compatibility with marginal sacrifices of accuracies. The code will be available at https://github.com/IrvingMeng/LCE.
Recent breakthroughs in the field of semi-supervised learning have achieved results that match state-of-the-art traditional supervised learning methods. Most successful semi-supervised learning approaches in computer vision focus on leveraging huge amount of unlabeled data, learning the general representation via data augmentation and transformation, creating pseudo labels, implementing different loss functions, and eventually transferring this knowledge to more task-specific smaller models. In this paper, we aim to conduct our analyses on three different aspects of SimCLR, the current state-of-the-art semi-supervised learning framework for computer vision. First, we analyze properties of contrast learning on fine-tuning, as we understand that contrast learning is what makes this method so successful. Second, we research knowledge distillation through teacher-forcing paradigm. We observe that when the teacher and the student share the same base model, knowledge distillation will achieve better result. Finally, we study how transfer learning works and its relationship with the number of classes on different data sets. Our results indicate that transfer learning performs better when number of classes are smaller.
Occluded person re-identification (ReID) aims to match person images with occlusion. It is fundamentally challenging because of the serious occlusion which aggravates the misalignment problem between images. At the cost of incorporating a pose estimator, many works introduce pose information to alleviate the misalignment in both training and testing. To achieve high accuracy while preserving low inference complexity, we propose a network named Pose-Guided Feature Learning with Knowledge Distillation (PGFL-KD), where the pose information is exploited to regularize the learning of semantics aligned features but is discarded in testing. PGFL-KD consists of a main branch (MB), and two pose-guided branches, \ieno, a foreground-enhanced branch (FEB), and a body part semantics aligned branch (SAB). The FEB intends to emphasise the features of visible body parts while excluding the interference of obstructions and background (\ieno, foreground feature alignment). The SAB encourages different channel groups to focus on different body parts to have body part semantics aligned representation. To get rid of the dependency on pose information when testing, we regularize the MB to learn the merits of the FEB and SAB through knowledge distillation and interaction-based training. Extensive experiments on occluded, partial, and holistic ReID tasks show the effectiveness of our proposed network.
It is a consensus that small models perform quite poorly under the paradigm of self-supervised contrastive learning. Existing methods usually adopt a large off-the-shelf model to transfer knowledge to the small one via distillation. Despite their effectiveness, distillation-based methods may not be suitable for some resource-restricted scenarios due to the huge computational expenses of deploying a large model. In this paper, we study the issue of training self-supervised small models without distillation signals. We first evaluate the representation spaces of the small models and make two non-negligible observations: (i) the small models can complete the pretext task without overfitting despite their limited capacity and (ii) they universally suffer the problem of over clustering. Then we verify multiple assumptions that are considered to alleviate the over-clustering phenomenon. Finally, we combine the validated techniques and improve the baseline performances of five small architectures with considerable margins, which indicates that training small self-supervised contrastive models is feasible even without distillation signals. The code is available at \textit{https://github.com/WOWNICE/ssl-small}.
Non-autoregressive neural machine translation (NAT) usually employs sequence-level knowledge distillation using autoregressive neural machine translation (AT) as its teacher model. However, a NAT model often outputs shorter sentences than an AT model. In this work, we propose sequence-level knowledge distillation (SKD) using perturbed length-aware positional encoding and apply it to a student model, the Levenshtein Transformer. Our method outperformed a standard Levenshtein Transformer by 2.5 points in bilingual evaluation understudy (BLEU) at maximum in a WMT14 German to English translation. The NAT model output longer sentences than the baseline NAT models.
Traditionally, federated learning (FL) aims to train a single global model while collaboratively using multiple clients and a server. Two natural challenges that FL algorithms face are heterogeneity in data across clients and collaboration of clients with {\em diverse resources}. In this work, we introduce a \textit{quantized} and \textit{personalized} FL algorithm QuPeD that facilitates collective (personalized model compression) training via \textit{knowledge distillation} (KD) among clients who have access to heterogeneous data and resources. For personalization, we allow clients to learn \textit{compressed personalized models} with different quantization parameters and model dimensions/structures. Towards this, first we propose an algorithm for learning quantized models through a relaxed optimization problem, where quantization values are also optimized over. When each client participating in the (federated) learning process has different requirements for the compressed model (both in model dimension and precision), we formulate a compressed personalization framework by introducing knowledge distillation loss for local client objectives collaborating through a global model. We develop an alternating proximal gradient update for solving this compressed personalization problem, and analyze its convergence properties. Numerically, we validate that QuPeD outperforms competing personalized FL methods, FedAvg, and local training of clients in various heterogeneous settings.
Knowledge distillation often involves how to define and transfer knowledge from teacher to student effectively. Although recent self-supervised contrastive knowledge achieves the best performance, forcing the network to learn such knowledge may damage the representation learning of the original class recognition task. We therefore adopt an alternative self-supervised augmented task to guide the network to learn the joint distribution of the original recognition task and self-supervised auxiliary task. It is demonstrated as a richer knowledge to improve the representation power without losing the normal classification capability. Moreover, it is incomplete that previous methods only transfer the probabilistic knowledge between the final layers. We propose to append several auxiliary classifiers to hierarchical intermediate feature maps to generate diverse self-supervised knowledge and perform the one-to-one transfer to teach the student network thoroughly. Our method significantly surpasses the previous SOTA SSKD with an average improvement of 2.56\% on CIFAR-100 and an improvement of 0.77\% on ImageNet across widely used network pairs. Codes are available at https://github.com/winycg/HSAKD.
Federated learning allows a large number of devices to jointly learn a model without sharing data. In this work, we enable clients with limited computing power to perform action recognition, a computationally heavy task. We first perform model compression at the central server through knowledge distillation on a large dataset. This allows the model to learn complex features and serves as an initialization for model fine-tuning. The fine-tuning is required because the limited data present in smaller datasets is not adequate for action recognition models to learn complex spatio-temporal features. Because the clients present are often heterogeneous in their computing resources, we use an asynchronous federated optimization and we further show a convergence bound. We compare our approach to two baseline approaches: fine-tuning at the central server (no clients) and fine-tuning using (heterogeneous) clients using synchronous federated averaging. We empirically show on a testbed of heterogeneous embedded devices that we can perform action recognition with comparable accuracy to the two baselines above, while our asynchronous learning strategy reduces the training time by 40%, relative to synchronous learning.
Catastrophic forgetting is one of the major challenges on the road for continual learning systems, which are presented with an on-line stream of tasks. The field has attracted considerable interest and a diverse set of methods have been presented for overcoming this challenge. Learning without Forgetting (LwF) is one of the earliest and most frequently cited methods. It has the advantages of not requiring the storage of samples from the previous tasks, of implementation simplicity, and of being well-grounded by relying on knowledge distillation. However, the prevailing view is that while it shows a relatively small amount of forgetting when only two tasks are introduced, it fails to scale to long sequences of tasks. This paper challenges this view, by showing that using the right architecture along with a standard set of augmentations, the results obtained by LwF surpass the latest algorithms for task incremental scenario. This improved performance is demonstrated by an extensive set of experiments over CIFAR-100 and Tiny-ImageNet, where it is also shown that other methods cannot benefit as much from similar improvements.
Text recognition remains a fundamental and extensively researched topic in computer vision, largely owing to its wide array of commercial applications. The challenging nature of the very problem however dictated a fragmentation of research efforts: Scene Text Recognition (STR) that deals with text in everyday scenes, and Handwriting Text Recognition (HTR) that tackles hand-written text. In this paper, for the first time, we argue for their unification -- we aim for a single model that can compete favourably with two separate state-of-the-art STR and HTR models. We first show that cross-utilisation of STR and HTR models trigger significant performance drops due to differences in their inherent challenges. We then tackle their union by introducing a knowledge distillation (KD) based framework. This is however non-trivial, largely due to the variable-length and sequential nature of text sequences, which renders off-the-shelf KD techniques that mostly works with global fixed-length data inadequate. For that, we propose three distillation losses all of which are specifically designed to cope with the aforementioned unique characteristics of text recognition. Empirical evidence suggests that our proposed unified model performs on par with individual models, even surpassing them in certain cases. Ablative studies demonstrate that naive baselines such as a two-stage framework, and domain adaption/generalisation alternatives do not work as well, further verifying the appropriateness of our design.
Graph neural networks (GNNs) have been widely used in many graph-based tasks such as node classification, link prediction, and node clustering. However, GNNs gain their performance benefits mainly from performing the feature propagation and smoothing across the edges of the graph, thus requiring sufficient connectivity and label information for effective propagation. Unfortunately, many real-world networks are sparse in terms of both edges and labels, leading to sub-optimal performance of GNNs. Recent interest in this sparse problem has focused on the self-training approach, which expands supervised signals with pseudo labels. Nevertheless, the self-training approach inherently cannot realize the full potential of refining the learning performance on sparse graphs due to the unsatisfactory quality and quantity of pseudo labels. In this paper, we propose ROD, a novel reception-aware online knowledge distillation approach for sparse graph learning. We design three supervision signals for ROD: multi-scale reception-aware graph knowledge, task-based supervision, and rich distilled knowledge, allowing online knowledge transfer in a peer-teaching manner. To extract knowledge concealed in the multi-scale reception fields, ROD explicitly requires individual student models to preserve different levels of locality information. For a given task, each student would predict based on its reception-scale knowledge, while simultaneously a strong teacher is established on-the-fly by combining multi-scale knowledge. Our approach has been extensively evaluated on 9 datasets and a variety of graph-based tasks, including node classification, link prediction, and node clustering. The result demonstrates that ROD achieves state-of-art performance and is more robust for the graph sparsity.
This paper describes the University of Sydney& JD's joint submission of the IWSLT 2021 low resource speech translation task. We participated in the Swahili-English direction and got the best scareBLEU (25.3) score among all the participants. Our constrained system is based on a pipeline framework, i.e. ASR and NMT. We trained our models with the officially provided ASR and MT datasets. The ASR system is based on the open-sourced tool Kaldi and this work mainly explores how to make the most of the NMT models. To reduce the punctuation errors generated by the ASR model, we employ our previous work SlotRefine to train a punctuation correction model. To achieve better translation performance, we explored the most recent effective strategies, including back translation, knowledge distillation, multi-feature reranking and transductive finetuning. For model structure, we tried auto-regressive and non-autoregressive models, respectively. In addition, we proposed two novel pre-train approaches, i.e. \textit{de-noising training} and \textit{bidirectional training} to fully exploit the data. Extensive experiments show that adding the above techniques consistently improves the BLEU scores, and the final submission system outperforms the baseline (Transformer ensemble model trained with the original parallel data) by approximately 10.8 BLEU score, achieving the SOTA performance.
A medical dialogue system is essential for healthcare service as providing primary clinical advice and diagnoses. It has been gradually adopted and practiced in medical organizations in the form of a conversational bot, largely due to the advancement of NLP. In recent years, the introduction of state-of-the-art deep learning models and transfer learning techniques like Universal Language Model Fine Tuning (ULMFiT) and Knowledge Distillation (KD) largely contributes to the performance of NLP tasks. However, some deep neural networks are poorly calibrated and wrongly estimate the uncertainty. Hence the model is not trustworthy, especially in sensitive medical decision-making systems and safety tasks. In this paper, we investigate the well-calibrated model for ULMFiT and self-distillation (SD) in a medical dialogue system. The calibrated ULMFiT (CULMFiT) is obtained by incorporating label smoothing (LS), a commonly used regularization technique to achieve a well-calibrated model. Moreover, we apply the technique to recalibrate the confidence score called temperature scaling (TS) with KD to observe its correlation with network calibration. To further understand the relation between SD and calibration, we use both fixed and optimal temperatures to fine-tune the whole model. All experiments are conducted on the consultation backpain dataset collected by experts then further validated using a large publicly medial dialogue corpus. We empirically show that our proposed methodologies outperform conventional methods in terms of accuracy and robustness.
Deep neural networks often have a huge number of parameters, which posts challenges in deployment in application scenarios with limited memory and computation capacity. Knowledge distillation is one approach to derive compact models from bigger ones. However, it has been observed that a converged heavy teacher model is strongly constrained for learning a compact student network and could make the optimization subject to poor local optima. In this paper, we propose ProKT, a new model-agnostic method by projecting the supervision signals of a teacher model into the student's parameter space. Such projection is implemented by decomposing the training objective into local intermediate targets with an approximate mirror descent technique. The proposed method could be less sensitive with the quirks during optimization which could result in a better local optimum. Experiments on both image and text datasets show that our proposed ProKT consistently achieves superior performance compared to other existing knowledge distillation methods.
The balance between high accuracy and high speed has always been a challenging task in semantic image segmentation. Compact segmentation networks are more widely used in the case of limited resources, while their performances are constrained. In this paper, motivated by the residual learning and global aggregation, we propose a simple yet general and effective knowledge distillation framework called double similarity distillation (DSD) to improve the classification accuracy of all existing compact networks by capturing the similarity knowledge in pixel and category dimensions, respectively. Specifically, we propose a pixel-wise similarity distillation (PSD) module that utilizes residual attention maps to capture more detailed spatial dependencies across multiple layers. Compared with exiting methods, the PSD module greatly reduces the amount of calculation and is easy to expand. Furthermore, considering the differences in characteristics between semantic segmentation task and other computer vision tasks, we propose a category-wise similarity distillation (CSD) module, which can help the compact segmentation network strengthen the global category correlation by constructing the correlation matrix. Combining these two modules, DSD framework has no extra parameters and only a minimal increase in FLOPs. Extensive experiments on four challenging datasets, including Cityscapes, CamVid, ADE20K, and Pascal VOC 2012, show that DSD outperforms current state-of-the-art methods, proving its effectiveness and generality. The code and models will be publicly available.
In this paper, a novel confidence conditioned knowledge distillation (CCKD) scheme for transferring the knowledge from a teacher model to a student model is proposed. Existing state-of-the-art methods employ fixed loss functions for this purpose and ignore the different levels of information that need to be transferred for different samples. In addition to that, these methods are also inefficient in terms of data usage. CCKD addresses these issues by leveraging the confidence assigned by the teacher model to the correct class to devise sample-specific loss functions (CCKD-L formulation) and targets (CCKD-T formulation). Further, CCKD improves the data efficiency by employing self-regulation to stop those samples from participating in the distillation process on which the student model learns faster. Empirical evaluations on several benchmark datasets show that CCKD methods achieve at least as much generalization performance levels as other state-of-the-art methods while being data efficient in the process. Student models trained through CCKD methods do not retain most of the misclassifications commited by the teacher model on the training set. Distillation through CCKD methods improves the resilience of the student models against adversarial attacks compared to the conventional KD method. Experiments show at least 3% increase in performance against adversarial attacks for the MNIST and the Fashion MNIST datasets, and at least 6% increase for the CIFAR10 dataset.
Inertial sensor-based human activity recognition (HAR) is the base of many human-centered mobile applications. Deep learning-based fine-grained HAR models enable accurate classification in various complex application scenarios. Nevertheless, the large storage and computational overhead of the existing fine-grained deep HAR models hinder their widespread deployment on resource-limited platforms. Inspired by the knowledge distillation's reasonable model compression and potential performance improvement capability, we design a multi-level HAR modeling pipeline called Stage-Logits-Memory Distillation (SMLDist) based on the widely-used MobileNet. By paying more attention to the frequency-related features during the distillation process, the SMLDist improves the HAR classification robustness of the students. We also propose an auto-search mechanism in the heterogeneous classifiers to improve classification performance. Extensive simulation results demonstrate that SMLDist outperforms various state-of-the-art HAR frameworks in accuracy and F1 macro score. The practical evaluation of the Jetson Xavier AGX platform shows that the SMLDist model is both energy-efficient and computation-efficient. These experiments validate the reasonable balance between the robustness and efficiency of the proposed model. The comparative experiments of knowledge distillation on six public datasets also demonstrate that the SMLDist outperforms other advanced knowledge distillation methods of students' performance, which verifies the good generalization of the SMLDist on other classification tasks, including but not limited to HAR.
Sequential recommender systems (SRS) have become a research hotspot due to its power in modeling user dynamic interests and sequential behavioral patterns. To maximize model expressive ability, a default choice is to apply a larger and deeper network architecture, which, however, often brings high network latency when generating online recommendations. Naturally, we argue that compressing the heavy recommendation models into middle- or light- weight neural networks is of great importance for practical production systems. To realize such a goal, we propose AdaRec, a knowledge distillation (KD) framework which compresses knowledge of a teacher model into a student model adaptively according to its recommendation scene by using differentiable Neural Architecture Search (NAS). Specifically, we introduce a target-oriented distillation loss to guide the structure search process for finding the student network architecture, and a cost-sensitive loss as constraints for model size, which achieves a superior trade-off between recommendation effectiveness and efficiency. In addition, we leverage Earth Mover's Distance (EMD) to realize many-to-many layer mapping during knowledge distillation, which enables each intermediate student layer to learn from other intermediate teacher layers adaptively. Extensive experiments on real-world recommendation datasets demonstrate that our model achieves competitive or better accuracy with notable inference speedup comparing to strong counterparts, while discovering diverse neural architectures for sequential recommender models under different recommendation scenes.
Pretraining and multitask learning are widely used to improve the speech to text translation performance. In this study, we are interested in training a speech to text translation model along with an auxiliary text to text translation task. We conduct a detailed analysis to understand the impact of the auxiliary task on the primary task within the multitask learning framework. Our analysis confirms that multitask learning tends to generate similar decoder representations from different modalities and preserve more information from the pretrained text translation modules. We observe minimal negative transfer effect between the two tasks and sharing more parameters is helpful to transfer knowledge from the text task to the speech task. The analysis also reveals that the modality representation difference at the top decoder layers is still not negligible, and those layers are critical for the translation quality. Inspired by these findings, we propose three methods to improve translation quality. First, a parameter sharing and initialization strategy is proposed to enhance information sharing between the tasks. Second, a novel attention-based regularization is proposed for the encoders and pulls the representations from different modalities closer. Third, an online knowledge distillation is proposed to enhance the knowledge transfer from the text to the speech task. Our experiments show that the proposed approach improves translation performance by more than 2 BLEU over a strong baseline and achieves state-of-the-art results on the \textsc{MuST-C} English-German, English-French and English-Spanish language pairs.