Mutual knowledge distillation (MKD) improves a model by distilling knowledge from another model. However, \textit{not all knowledge is certain and correct}, especially under adverse conditions. For example, label noise usually leads to less reliable models due to undesired memorization \cite{zhang2017understanding,arpit2017closer}. Wrong knowledge misleads the learning rather than helps. This problem can be handled by two aspects: (i) improving the reliability of a model where the knowledge is from (i.e., knowledge source's reliability); (ii) selecting reliable knowledge for distillation. In the literature, making a model more reliable is widely studied while selective MKD receives little attention. Therefore, we focus on studying selective MKD. Concretely, a generic MKD framework, \underline{C}onfident knowledge selection followed by \underline{M}utual \underline{D}istillation (CMD), is designed. The key component of CMD is a generic knowledge selection formulation, making the selection threshold either static (CMD-S) or progressive (CMD-P). Additionally, CMD covers two special cases: zero-knowledge and all knowledge, leading to a unified MKD framework. Extensive experiments are present to demonstrate the effectiveness of CMD and thoroughly justify the design of CMD. For example, CMD-P obtains new state-of-the-art results in robustness against label noise.
Pre-trained language models (PLMs) achieve great success in NLP. However, their huge model sizes hinder their applications in many practical systems. Knowledge distillation is a popular technique to compress PLMs, which learns a small student model from a large teacher PLM. However, the knowledge learned from a single teacher may be limited and even biased, resulting in low-quality student model. In this paper, we propose a multi-teacher knowledge distillation framework named MT-BERT for pre-trained language model compression, which can train high-quality student model from multiple teacher PLMs. In MT-BERT we design a multi-teacher co-finetuning method to jointly finetune multiple teacher PLMs in downstream tasks with shared pooling and prediction layers to align their output space for better collaborative teaching. In addition, we propose a multi-teacher hidden loss and a multi-teacher distillation loss to transfer the useful knowledge in both hidden states and soft labels from multiple teacher PLMs to the student model. Experiments on three benchmark datasets validate the effectiveness of MT-BERT in compressing PLMs.
Manual fact-checking does not scale well to serve the needs of the internet. This issue is further compounded in non-English contexts. In this paper, we discuss claim matching as a possible solution to scale fact-checking. We define claim matching as the task of identifying pairs of textual messages containing claims that can be served with one fact-check. We construct a novel dataset of WhatsApp tipline and public group messages alongside fact-checked claims that are first annotated for containing "claim-like statements" and then matched with potentially similar items and annotated for claim matching. Our dataset contains content in high-resource (English, Hindi) and lower-resource (Bengali, Malayalam, Tamil) languages. We train our own embedding model using knowledge distillation and a high-quality "teacher" model in order to address the imbalance in embedding quality between the low- and high-resource languages in our dataset. We provide evaluations on the performance of our solution and compare with baselines and existing state-of-the-art multilingual embedding models, namely LASER and LaBSE. We demonstrate that our performance exceeds LASER and LaBSE in all settings. We release our annotated datasets, codebooks, and trained embedding model to allow for further research.
Knowledge distillation (KD) is commonly used to construct synthetic data for training non-autoregressive translation (NAT) models. However, there exists a discrepancy on low-frequency words between the distilled and the original data, leading to more errors on predicting low-frequency words. To alleviate the problem, we directly expose the raw data into NAT by leveraging pretraining. By analyzing directed alignments, we found that KD makes low-frequency source words aligned with targets more deterministically but fails to align sufficient low-frequency words from target to source. Accordingly, we propose reverse KD to rejuvenate more alignments for low-frequency target words. To make the most of authentic and synthetic data, we combine these complementary approaches as a new training strategy for further boosting NAT performance. We conduct experiments on five translation benchmarks over two advanced architectures. Results demonstrate that the proposed approach can significantly and universally improve translation quality by reducing translation errors on low-frequency words. Encouragingly, our approach achieves 28.2 and 33.9 BLEU points on the WMT14 English-German and WMT16 Romanian-English datasets, respectively. Our code, data, and trained models are available at \url{https://github.com/alphadl/RLFW-NAT}.
Automatic dialogue coherence evaluation has attracted increasing attention and is crucial for developing promising dialogue systems. However, existing metrics have two major limitations: (a) they are mostly trained in a simplified two-level setting (coherent vs. incoherent), while humans give Likert-type multi-level coherence scores, dubbed as "quantifiable"; (b) their predicted coherence scores cannot align with the actual human rating standards due to the absence of human guidance during training. To address these limitations, we propose Quantifiable Dialogue Coherence Evaluation (QuantiDCE), a novel framework aiming to train a quantifiable dialogue coherence metric that can reflect the actual human rating standards. Specifically, QuantiDCE includes two training stages, Multi-Level Ranking (MLR) pre-training and Knowledge Distillation (KD) fine-tuning. During MLR pre-training, a new MLR loss is proposed for enabling the model to learn the coarse judgement of coherence degrees. Then, during KD fine-tuning, the pretrained model is further finetuned to learn the actual human rating standards with only very few human-annotated data. To advocate the generalizability even with limited fine-tuning data, a novel KD regularization is introduced to retain the knowledge learned at the pre-training stage. Experimental results show that the model trained by QuantiDCE presents stronger correlations with human judgements than the other state-of-the-art metrics.
In a matter that is analog to the study of natural image statistics, we study the natural statistics of the deep neural network activations at various layers. As we show, these statistics, similar to image statistics, follow a power law. We also show, both analytically and empirically, that with depth the exponent of this power law increases at a linear rate. As a direct implication of our discoveries, we present a method for performing Knowledge Distillation (KD). While classical KD methods consider the logits of the teacher network, more recent methods obtain a leap in performance by considering the activation maps. This, however, uses metrics that are suitable for comparing images. We propose to employ two additional loss terms that are based on the spectral properties of the intermediate activation maps. The proposed method obtains state of the art results on multiple image recognition KD benchmarks.
Named entity recognition (NER) is a fundamental component in many applications, such as Web Search and Voice Assistants. Although deep neural networks greatly improve the performance of NER, due to the requirement of large amounts of training data, deep neural networks can hardly scale out to many languages in an industry setting. To tackle this challenge, cross-lingual NER transfers knowledge from a rich-resource language to languages with low resources through pre-trained multilingual language models. Instead of using training data in target languages, cross-lingual NER has to rely on only training data in source languages, and optionally adds the translated training data derived from source languages. However, the existing cross-lingual NER methods do not make good use of rich unlabeled data in target languages, which is relatively easy to collect in industry applications. To address the opportunities and challenges, in this paper we describe our novel practice in Microsoft to leverage such large amounts of unlabeled data in target languages in real production settings. To effectively extract weak supervision signals from the unlabeled data, we develop a novel approach based on the ideas of semi-supervised learning and reinforcement learning. The empirical study on three benchmark data sets verifies that our approach establishes the new state-of-the-art performance with clear edges. Now, the NER techniques reported in this paper are on their way to become a fundamental component for Web ranking, Entity Pane, Answers Triggering, and Question Answering in the Microsoft Bing search engine. Moreover, our techniques will also serve as part of the Spoken Language Understanding module for a commercial voice assistant. We plan to open source the code of the prototype framework after deployment.
Fine-tuning transformer models after unsupervised pre-training reaches a very high performance on many different natural language processing tasks. Unfortunately, transformers suffer from long inference times which greatly increases costs in production. One possible solution is to use knowledge distillation, which solves this problem by transferring information from large teacher models to smaller student models. Knowledge distillation maintains high performance and reaches high compression rates, nevertheless, the size of the student model is fixed after pre-training and can not be changed individually for a given downstream task and use-case to reach a desired performance/speedup ratio. Another solution to reduce the size of models in a much more fine-grained and computationally cheaper fashion is to prune layers after the pre-training. The price to pay is that the performance of layer-wise pruning algorithms is not on par with state-of-the-art knowledge distillation methods. In this paper, Greedy-layer pruning is introduced to (1) outperform current state-of-the-art for layer-wise pruning, (2) close the performance gap when compared to knowledge distillation, while (3) providing a method to adapt the model size dynamically to reach a desired performance/speedup tradeoff without the need of additional pre-training phases. Our source code is available on https://github.com/deepopinion/greedy-layer-pruning.
In this paper, we study the task of source-free domain adaptation (SFDA), where the source data are not available during target adaptation. Previous works on SFDA mainly focus on aligning the cross-domain distributions. However, they ignore the generalization ability of the pretrained source model, which largely influences the initial target outputs that are vital to the target adaptation stage. To address this, we make the interesting observation that the model accuracy is highly correlated with whether or not attention is focused on the objects in an image. To this end, we propose a generic and effective framework based on Transformer, named TransDA, for learning a generalized model for SFDA. Specifically, we apply the Transformer as the attention module and inject it into a convolutional network. By doing so, the model is encouraged to turn attention towards the object regions, which can effectively improve the model's generalization ability on the target domains. Moreover, a novel self-supervised knowledge distillation approach is proposed to adapt the Transformer with target pseudo-labels, thus further encouraging the network to focus on the object regions. Experiments on three domain adaptation tasks, including closed-set, partial-set, and open-set adaption, demonstrate that TransDA can greatly improve the adaptation accuracy and produce state-of-the-art results. The source code and trained models are available at https://github.com/ygjwd12345/TransDA.
In Natural Language Processing (NLP), finding data augmentation techniques that can produce high-quality human-interpretable examples has always been challenging. Recently, leveraging kNN such that augmented examples are retrieved from large repositories of unlabelled sentences has made a step toward interpretable augmentation. Inspired by this paradigm, we introduce Minimax-kNN, a sample efficient data augmentation strategy tailored for Knowledge Distillation (KD). We exploit a semi-supervised approach based on KD to train a model on augmented data. In contrast to existing kNN augmentation techniques that blindly incorporate all samples, our method dynamically selects a subset of augmented samples that maximizes KL-divergence between the teacher and student models. This step aims to extract the most efficient samples to ensure our augmented data covers regions in the input space with maximum loss value. We evaluated our technique on several text classification tasks and demonstrated that Minimax-kNN consistently outperforms strong baselines. Our results show that Minimax-kNN requires fewer augmented examples and less computation to achieve superior performance over the state-of-the-art kNN-based augmentation techniques.
Recent explorations of large-scale pre-trained language models (PLMs) have revealed the power of PLMs with huge amounts of parameters, setting off a wave of training ever-larger PLMs. However, it requires tremendous computational resources to train a large-scale PLM, which may be practically unaffordable. In addition, existing large-scale PLMs are mainly trained from scratch individually, ignoring that many well-trained PLMs are available. To this end, we explore the question how could existing PLMs benefit training large-scale PLMs in future. Specifically, we introduce a pre-training framework named "knowledge inheritance" (KI) and explore how could knowledge distillation serve as auxiliary supervision during pre-training to efficiently learn larger PLMs. Experimental results demonstrate the superiority of KI in training efficiency. We also conduct empirical analyses to explore the effects of teacher PLMs' pre-training settings, including model architecture, pre-training data, etc. Finally, we show that KI could be applied to domain adaptation and knowledge transfer.
As GAN-based video and image manipulation technologies become more sophisticated and easily accessible, there is an urgent need for effective deepfake detection technologies. Moreover, various deepfake generation techniques have emerged over the past few years. While many deepfake detection methods have been proposed, their performance suffers from new types of deepfake methods on which they are not sufficiently trained. To detect new types of deepfakes, the model should learn from additional data without losing its prior knowledge about deepfakes (catastrophic forgetting), especially when new deepfakes are significantly different. In this work, we employ the Representation Learning (ReL) and Knowledge Distillation (KD) paradigms to introduce a transfer learning-based Feature Representation Transfer Adaptation Learning (FReTAL) method. We use FReTAL to perform domain adaptation tasks on new deepfake datasets while minimizing catastrophic forgetting. Our student model can quickly adapt to new types of deepfake by distilling knowledge from a pre-trained teacher model and applying transfer learning without using source domain data during domain adaptation. Through experiments on FaceForensics++ datasets, we demonstrate that FReTAL outperforms all baselines on the domain adaptation task with up to 86.97% accuracy on low-quality deepfakes.
Knowledge distillation, i.e., one classifier being trained on the outputs of another classifier, is an empirically very successful technique for knowledge transfer between classifiers. It has even been observed that classifiers learn much faster and more reliably if trained with the outputs of another classifier as soft labels, instead of from ground truth data. So far, however, there is no satisfactory theoretical explanation of this phenomenon. In this work, we provide the first insights into the working mechanisms of distillation by studying the special case of linear and deep linear classifiers. Specifically, we prove a generalization bound that establishes fast convergence of the expected risk of a distillation-trained linear classifier. From the bound and its proof we extract three key factors that determine the success of distillation: * data geometry -- geometric properties of the data distribution, in particular class separation, has a direct influence on the convergence speed of the risk; * optimization bias -- gradient descent optimization finds a very favorable minimum of the distillation objective; and * strong monotonicity -- the expected risk of the student classifier always decreases when the size of the training set grows.
While non-autoregressive (NAR) models are showing great promise for machine translation, their use is limited by their dependence on knowledge distillation from autoregressive models. To address this issue, we seek to understand why distillation is so effective. Prior work suggests that distilled training data is less complex than manual translations. Based on experiments with the Levenshtein Transformer and the Mask-Predict NAR models on the WMT14 German-English task, this paper shows that different types of complexity have different impacts: while reducing lexical diversity and decreasing reordering complexity both help NAR learn better alignment between source and target, and thus improve translation quality, lexical diversity is the main reason why distillation increases model confidence, which affects the calibration of different NAR models differently.
We propose Joint-DetNAS, a unified NAS framework for object detection, which integrates 3 key components: Neural Architecture Search, pruning, and Knowledge Distillation. Instead of naively pipelining these techniques, our Joint-DetNAS optimizes them jointly. The algorithm consists of two core processes: student morphism optimizes the student's architecture and removes the redundant parameters, while dynamic distillation aims to find the optimal matching teacher. For student morphism, weight inheritance strategy is adopted, allowing the student to flexibly update its architecture while fully utilize the predecessor's weights, which considerably accelerates the search; To facilitate dynamic distillation, an elastic teacher pool is trained via integrated progressive shrinking strategy, from which teacher detectors can be sampled without additional cost in subsequent searches. Given a base detector as the input, our algorithm directly outputs the derived student detector with high performance without additional training. Experiments demonstrate that our Joint-DetNAS outperforms the naive pipelining approach by a great margin. Given a classic R101-FPN as the base detector, Joint-DetNAS is able to boost its mAP from 41.4 to 43.9 on MS COCO and reduce the latency by 47%, which is on par with the SOTA EfficientDet while requiring less search cost. We hope our proposed method can provide the community with a new way of jointly optimizing NAS, KD and pruning.
Neural Machine Translation (NMT) models achieve state-of-the-art performance on many translation benchmarks. As an active research field in NMT, knowledge distillation is widely applied to enhance the model's performance by transferring teacher model's knowledge on each training sample. However, previous work rarely discusses the different impacts and connections among these samples, which serve as the medium for transferring teacher knowledge. In this paper, we design a novel protocol that can effectively analyze the different impacts of samples by comparing various samples' partitions. Based on above protocol, we conduct extensive experiments and find that the teacher's knowledge is not the more, the better. Knowledge over specific samples may even hurt the whole performance of knowledge distillation. Finally, to address these issues, we propose two simple yet effective strategies, i.e., batch-level and global-level selections, to pick suitable samples for distillation. We evaluate our approaches on two large-scale machine translation tasks, WMT'14 English->German and WMT'19 Chinese->English. Experimental results show that our approaches yield up to +1.28 and +0.89 BLEU points improvements over the Transformer baseline, respectively.
Recently, deep reinforcement learning (RL) algorithms have made great progress in multi-agent domain. However, due to characteristics of RL, training for complex tasks would be resource-intensive and time-consuming. To meet this challenge, mutual learning strategy between homogeneous agents is essential, which is under-explored in previous studies, because most existing methods do not consider to use the knowledge of agent models. In this paper, we present an adaptation method of the majority of multi-agent reinforcement learning (MARL) algorithms called KnowSR which takes advantage of the differences in learning between agents. We employ the idea of knowledge distillation (KD) to share knowledge among agents to shorten the training phase. To empirically demonstrate the robustness and effectiveness of KnowSR, we performed extensive experiments on state-of-the-art MARL algorithms in collaborative and competitive scenarios. The results demonstrate that KnowSR outperforms recently reported methodologies, emphasizing the importance of the proposed knowledge sharing for MARL.
Monocular (relative or metric) depth estimation is a critical task for various applications, such as autonomous vehicles, augmented reality and image editing. In recent years, with the increasing availability of mobile devices, accurate and mobile-friendly depth models have gained importance. Increasingly accurate models typically require more computational resources, which inhibits the use of such models on mobile devices. The mobile use case is arguably the most unrestricted one, which requires highly accurate yet mobile-friendly architectures. Therefore, we try to answer the following question: How can we improve a model without adding further complexity (i.e. parameters)? Towards this end, we systematically explore the design space of a relative depth estimation model from various dimensions and we show, with key design choices and ablation studies, even an existing architecture can reach highly competitive performance to the state of the art, with a fraction of the complexity. Our study spans an in-depth backbone model selection process, knowledge distillation, intermediate predictions, model pruning and loss rebalancing. We show that our model, using only DIW as the supervisory dataset, achieves 0.1156 WHDR on DIW with 2.6M parameters and reaches 37 FPS on a mobile GPU, without pruning or hardware-specific optimization. A pruned version of our model achieves 0.1208 WHDR on DIW with 1M parameters and reaches 44 FPS on a mobile GPU.
It is known that deep neural networks, trained for the classification of non-sensitive target attributes, can reveal sensitive attributes of their input data through internal representations extracted by the classifier. We take a step forward and show that deep classifiers can be trained to secretly encode a sensitive attribute of their input data into the classifier's outputs for the target attribute, at inference time. Our proposed attack works even if users have a full white-box view of the classifier, can keep all internal representations hidden, and only release the classifier's estimations for the target attribute. We introduce an information-theoretical formulation for such attacks and present efficient empirical implementations for training honest-but-curious (HBC) classifiers: classifiers that can be accurate in predicting their target attribute, but can also exploit their outputs to secretly encode a sensitive attribute. Our work highlights a vulnerability that can be exploited by malicious machine learning service providers to attack their user's privacy in several seemingly safe scenarios; such as encrypted inferences, computations at the edge, or private knowledge distillation. Experimental results on several attributes in two face-image datasets show that a semi-trusted server can train classifiers that are not only perfectly honest but also accurately curious. We conclude by showing the difficulties in distinguishing between standard and HBC classifiers, discussing challenges in defending against this vulnerability of deep classifiers, and enumerating related open directions for future studies.
Multi-modal magnetic resonance imaging (MRI) is a crucial method for analyzing the human brain. It is usually used for diagnosing diseases and for making valuable decisions regarding the treatments - for instance, checking for gliomas in the human brain. With varying degrees of severity and detection, properly diagnosing gliomas is one of the most daunting and significant analysis tasks in modern-day medicine. Our primary focus is on working with different approaches to perform the segmentation of brain tumors in multimodal MRI scans. Now, the quantity, variability of the data used for training has always been considered to be crucial for developing excellent models. Hence, we also want to experiment with Knowledge Distillation techniques.