End-to-end approaches open a new way for more accurate and efficient spoken language understanding (SLU) systems by alleviating the drawbacks of traditional pipeline systems. Previous works exploit textual information for an SLU model via pre-training with automatic speech recognition or fine-tuning with knowledge distillation. To utilize textual information more effectively, this work proposes a two-stage textual knowledge distillation method that matches utterance-level representations and predicted logits of two modalities during pre-training and fine-tuning, sequentially. We use vq-wav2vec BERT as a speech encoder because it captures general and rich features. Furthermore, we improve the performance, especially in a low-resource scenario, with data augmentation methods by randomly masking spans of discrete audio tokens and contextualized hidden representations. Consequently, we push the state-of-the-art on the Fluent Speech Commands, achieving 99.7% test accuracy in the full dataset setting and 99.5% in the 10% subset setting. Throughout the ablation studies, we empirically verify that all used methods are crucial to the final performance, providing the best practice for spoken language understanding. Code is available at https://github.com/clovaai/textual-kd-slu.
Recent state-of-the-art approaches to summarization utilize large pre-trained Transformer models. Distilling these models to smaller student models has become critically important for practical use; however there are many different distillation methods proposed by the NLP literature. Recent work on distilling BERT for classification and regression tasks shows strong performance using direct knowledge distillation. Alternatively, machine translation practitioners distill using pseudo-labeling, where a small model is trained on the translations of a larger model. A third, simpler approach is to 'shrink and fine-tune' (SFT), which avoids any explicit distillation by copying parameters to a smaller student model and then fine-tuning. We compare these three approaches for distillation of Pegasus and BART, the current and former state of the art, pre-trained summarization models, and find that SFT outperforms knowledge distillation and pseudo-labeling on the CNN/DailyMail dataset, but under-performs pseudo-labeling on the more abstractive XSUM dataset. PyTorch Code and checkpoints of different sizes are available through Hugging Face transformers here http://tiny.cc/4iy0tz.
Model compression becomes a recent trend due to the requirement of deploying neural networks on embedded and mobile devices. Hence, both accuracy and efficiency are of critical importance. To explore a balance between them, a knowledge distillation strategy is proposed for general visual representation learning. It utilizes our well-designed activation map adaptive module to replace some blocks of the teacher network, exploring the most appropriate supervisory features adaptively during the training process. Using the teacher's hidden layer output to prompt the student network to train so as to transfer effective semantic information.To verify the effectiveness of our strategy, this paper applied our method to cifar-10 dataset. Results demonstrate that the method can boost the accuracy of the student network by 0.6% with 6.5% loss reduction, and significantly improve its training speed.
Cross-lingual Machine Reading Comprehension (CLMRC) remains a challenging problem due to the lack of large-scale annotated datasets in low-source languages, such as Arabic, Hindi, and Vietnamese. Many previous approaches use translation data by translating from a rich-source language, such as English, to low-source languages as auxiliary supervision. However, how to effectively leverage translation data and reduce the impact of noise introduced by translation remains onerous. In this paper, we tackle this challenge and enhance the cross-lingual transferring performance by a novel augmentation approach named Language Branch Machine Reading Comprehension (LBMRC). A language branch is a group of passages in one single language paired with questions in all target languages. We train multiple machine reading comprehension (MRC) models proficient in individual language based on LBMRC. Then, we devise a multilingual distillation approach to amalgamate knowledge from multiple language branch models to a single model for all target languages. Combining the LBMRC and multilingual distillation can be more robust to the data noises, therefore, improving the model's cross-lingual ability. Meanwhile, the produced single multilingual model is applicable to all target languages, which saves the cost of training, inference, and maintenance for multiple models. Extensive experiments on two CLMRC benchmarks clearly show the effectiveness of our proposed method.
Transformer-based models are the state-of-the-art for Natural Language Understanding (NLU) applications. Models are getting bigger and better on various tasks. However, Transformer models remain computationally challenging since they are not efficient at inference-time compared to traditional approaches. In this paper, we present FastFormers, a set of recipes to achieve efficient inference-time performance for Transformer-based models on various NLU tasks. We show how carefully utilizing knowledge distillation, structured pruning and numerical optimization can lead to drastic improvements on inference efficiency. We provide effective recipes that can guide practitioners to choose the best settings for various NLU tasks and pretrained models. Applying the proposed recipes to the SuperGLUE benchmark, we achieve from 9.8x up to 233.9x speed-up compared to out-of-the-box models on CPU. On GPU, we also achieve up to 12.4x speed-up with the presented methods. We show that FastFormers can drastically reduce cost of serving 100 million requests from 4,223 USD to just 18 USD on an Azure F16s_v2 instance. This translates to a sustainable runtime by reducing energy consumption 6.9x - 125.8x according to the metrics used in the SustaiNLP 2020 shared task.
Automatically generating financial report from a piece of news is quite a challenging task. Apparently, the difficulty of this task lies in the lack of sufficient background knowledge to effectively generate long financial report. To address this issue, this paper proposes the conditional variational autoencoders (CVAE) based approach which distills external knowledge from a corpus of news-report data. Particularly, we choose Bi-GRU as the encoder and decoder component of CVAE, and learn the latent variable distribution from input news. A higher level latent variable distribution is learnt from a corpus set of news-report data, respectively extr acted for each input news, to provide background knowledge to previously learnt latent variable distribution. Then, a teacher-student network is employed to distill knowledge to refine theoutput of the decoder component. To evaluate the model performance of the proposed approach, extensive experiments are preformed on a public dataset and two widely adopted evaluation criteria, i.e., BLEU and ROUGE, are chosen in the experiment. The promising experimental results demonstrate that the proposed approach is superior to the rest compared methods.
We present an approach to ranking with dense representations that applies knowledge distillation to improve the recently proposed late-interaction ColBERT model. Specifically, we distill the knowledge from ColBERT's expressive MaxSim operator for computing relevance scores into a simple dot product, thus enabling single-step ANN search. Our key insight is that during distillation, tight coupling between the teacher model and the student model enables more flexible distillation strategies and yields better learned representations. We empirically show that our approach improves query latency and greatly reduces the onerous storage requirements of ColBERT, while only making modest sacrifices in terms of effectiveness. By combining our dense representations with sparse representations derived from document expansion, we are able to approach the effectiveness of a standard cross-encoder reranker using BERT that is orders of magnitude slower.
A pre-trained language model, BERT, has brought significant performance improvements across a range of natural language processing tasks. Since the model is trained on a large corpus of diverse topics, it shows robust performance for domain shift problems in which data distributions at training (source data) and testing (target data) differ while sharing similarities. Despite its great improvements compared to previous models, it still suffers from performance degradation due to domain shifts. To mitigate such problems, we propose a simple but effective unsupervised domain adaptation method, adversarial adaptation with distillation (AAD), which combines the adversarial discriminative domain adaptation (ADDA) framework with knowledge distillation. We evaluate our approach in the task of cross-domain sentiment classification on 30 domain pairs, advancing the state-of-the-art performance for unsupervised domain adaptation in text sentiment classification.
Spoken question answering (SQA) is a challenging task that requires the machine to fully understand the complex spoken documents. Automatic speech recognition (ASR) plays a significant role in the development of QA systems. However, the recent work shows that ASR systems generate highly noisy transcripts, which critically limit the capability of machine comprehension on the SQA task. To address the issue, we present a novel distillation framework. Specifically, we devise a training strategy to perform knowledge distillation (KD) from spoken documents and written counterparts. Our work makes a step towards distilling knowledge from the language model as a supervision signal to lead to better student accuracy by reducing the misalignment between automatic and manual transcriptions. Experiments demonstrate that our approach outperforms several state-of-the-art language models on the Spoken-SQuAD dataset.
Spoken conversational question answering (SCQA) requires machines to model complex dialogue flow given the speech utterances and text corpora. Different from traditional text question answering (QA) tasks, SCQA involves audio signal processing, passage comprehension, and contextual understanding. However, ASR systems introduce unexpected noisy signals to the transcriptions, which result in performance degradation on SCQA. To overcome the problem, we propose CADNet, a novel contextualized attention-based distillation approach, which applies both cross-attention and self-attention to obtain ASR-robust contextualized embedding representations of the passage and dialogue history for performance improvements. We also introduce the spoken conventional knowledge distillation framework to distill the ASR-robust knowledge from the estimated probabilities of the teacher model to the student. We conduct extensive experiments on the Spoken-CoQA dataset and demonstrate that our approach achieves remarkable performance in this task.
Since the wide employment of deep learning frameworks in video salient object detection, the accuracy of the recent approaches has made stunning progress. These approaches mainly adopt the sequential modules, based on optical flow or recurrent neural network (RNN), to learn robust spatiotemporal features. These modules are effective but significantly increase the computational burden of the corresponding deep models. In this paper, to simplify the network and maintain the accuracy, we present a lightweight network tailored for video salient object detection through the spatiotemporal knowledge distillation. Specifically, in the spatial aspect, we combine a saliency guidance feature embedding structure and spatial knowledge distillation to refine the spatial features. In the temporal aspect, we propose a temporal knowledge distillation strategy, which allows the network to learn the robust temporal features through the infer-frame feature encoding and distilling information from adjacent frames. The experiments on widely used video datasets (e.g., DAVIS, DAVSOD, SegTrack-V2) prove that our approach achieves competitive performance. Furthermore, without the employment of the complex sequential modules, the proposed network can obtain high efficiency with 0.01s per frame.
Federated learning (FL), which utilizes communication between the server (core) and local devices (edges) to indirectly learn from more data, is an emerging field in deep learning research. Recently, Knowledge Distillation-based FL methods with notable performance and high applicability have been suggested. In this paper, we choose knowledge distillation-based FL method as our baseline and tackle a challenging problem that ensues from using these methods. Especially, we focus on the problem incurred in the server model that tries to mimic different datasets, each of which is unique to an individual edge device. We dub the problem 'edge bias', which occurs when multiple teacher models trained on different datasets are used individually to distill knowledge. We introduce this nuisance that occurs in certain scenarios of FL, and to alleviate it, we propose a simple yet effective distillation scheme named 'buffered distillation'. In addition, we also experimentally show that this scheme is effective in mitigating the straggler problem caused by delayed edges.
Knowledge distillation is a strategy of training a student network with guide of the soft output from a teacher network. It has been a successful method of model compression and knowledge transfer. However, currently knowledge distillation lacks a convincing theoretical understanding. On the other hand, recent finding on neural tangent kernel enables us to approximate a wide neural network with a linear model of the network's random features. In this paper, we theoretically analyze the knowledge distillation of a wide neural network. First we provide a transfer risk bound for the linearized model of the network. Then we propose a metric of the task's training difficulty, called data inefficiency. Based on this metric, we show that for a perfect teacher, a high ratio of teacher's soft labels can be beneficial. Finally, for the case of imperfect teacher, we find that hard labels can correct teacher's wrong prediction, which explains the practice of mixing hard and soft labels.
Non-autoregressive models generate target words in a parallel way, which achieve a faster decoding speed but at the sacrifice of translation accuracy. To remedy a flawed translation by non-autoregressive models, a promising approach is to train a conditional masked translation model (CMTM), and refine the generated results within several iterations. Unfortunately, such approach hardly considers the \textit{sequential dependency} among target words, which inevitably results in a translation degradation. Hence, instead of solely training a Transformer-based CMTM, we propose a Self-Review Mechanism to infuse sequential information into it. Concretely, we insert a left-to-right mask to the same decoder of CMTM, and then induce it to autoregressively review whether each generated word from CMTM is supposed to be replaced or kept. The experimental results (WMT14 En$\leftrightarrow$De and WMT16 En$\leftrightarrow$Ro) demonstrate that our model uses dramatically less training computations than the typical CMTM, as well as outperforms several state-of-the-art non-autoregressive models by over 1 BLEU. Through knowledge distillation, our model even surpasses a typical left-to-right Transformer model, while significantly speeding up decoding.
Link prediction based on knowledge graph embeddings (KGE) aims to predict new triples to automatically construct knowledge graphs (KGs). However, recent KGE models achieve performance improvements by excessively increasing the embedding dimensions, which may cause enormous training costs and require more storage space. In this paper, instead of training high-dimensional models, we propose MulDE, a novel knowledge distillation framework, which includes multiple low-dimensional hyperbolic KGE models as teachers and two student components, namely Junior and Senior. Under a novel iterative distillation strategy, the Junior component, a low-dimensional KGE model, asks teachers actively based on its preliminary prediction results, and the Senior component integrates teachers' knowledge adaptively to train the Junior component based on two mechanisms: relation-specific scaling and contrast attention. The experimental results show that MulDE can effectively improve the performance and training speed of low-dimensional KGE models. The distilled 32-dimensional model is competitive compared to the state-of-the-art high-dimensional methods on several widely-used datasets.
Large-scale pre-trained models have attracted extensive attention in the research community and shown promising results on various tasks of natural language processing. However, these pre-trained models are memory and computation intensive, hindering their deployment into industrial online systems like Ad Relevance. Meanwhile, how to design an effective yet efficient model architecture is another challenging problem in online Ad Relevance. Recently, AutoML shed new lights on architecture design, but how to integrate it with pre-trained language models remains unsettled. In this paper, we propose AutoADR (Automatic model design for AD Relevance) -- a novel end-to-end framework to address this challenge, and share our experience to ship these cutting-edge techniques into online Ad Relevance system at Microsoft Bing. Specifically, AutoADR leverages a one-shot neural architecture search algorithm to find a tailored network architecture for Ad Relevance. The search process is simultaneously guided by knowledge distillation from a large pre-trained teacher model (e.g. BERT), while taking the online serving constraints (e.g. memory and latency) into consideration. We add the model designed by AutoADR as a sub-model into the production Ad Relevance model. This additional sub-model improves the Precision-Recall AUC (PR AUC) on top of the original Ad Relevance model by 2.65X of the normalized shipping bar. More importantly, adding this automatically designed sub-model leads to a statistically significant 4.6% Bad-Ad ratio reduction in online A/B testing. This model has been shipped into Microsoft Bing Ad Relevance Production model.
Knowledge distillation aims at obtaining a compact and effective model by learning the mapping function from a much larger one. Due to the limited capacity of the student, the student would underfit the teacher. Therefore, student performance would unexpectedly drop when distilling from an oversized teacher, termed the capacity gap problem. We investigate this problem by study the gap of confidence between teacher and student. We find that the magnitude of confidence is not necessary for knowledge distillation and could harm the student performance if the student are forced to learn confidence. We propose Spherical Knowledge Distillation to eliminate this gap explicitly, which eases the underfitting problem. We find this novel knowledge representation can improve compact models with much larger teachers and is robust to temperature. We conducted experiments on both CIFAR100 and ImageNet, and achieve significant improvement. Specifically, we train ResNet18 to 73.0 accuracy, which is a substantial improvement over previous SOTA and is on par with resnet34 almost twice the student size. The implementation has been shared at https://github.com/forjiuzhou/Spherical-Knowledge-Distillation.
The Transformer self-attention network has recently shown promising performance as an alternative to recurrent neural networks in end-to-end (E2E) automatic speech recognition (ASR) systems. However, Transformer has a drawback in that the entire input sequence is required to compute self-attention. We have proposed a block processing method for the Transformer encoder by introducing a context-aware inheritance mechanism. An additional context embedding vector handed over from the previously processed block helps encode not only local acoustic information but also global linguistic, channel, and speaker attributes. In this paper, we extend block processing towards an entire streaming E2E ASR system without additional training, by introducing a blockwise synchronous decoding process inspired by a neural transducer into the Transformer decoder. We further apply a knowledge distillation technique with which training of the streaming Transformer is guided by the ordinary batch Transformer model. Evaluations of the HKUST and AISHELL-1 Mandarin tasks and LibriSpeech English task show that our proposed streaming Transformer outperforms conventional online approaches including monotonic chunkwise attention (MoChA). We also confirm that the knowledge distillation technique improves the accuracy further. Our streaming ASR models achieve comparable/superior performance to the batch models and other streaming-based transformer methods in all the tasks.
Streaming automatic speech recognition (ASR) aims to emit each hypothesized word as quickly and accurately as possible, while full-context ASR waits for the completion of a full speech utterance before emitting completed hypotheses. In this work, we propose a unified framework, Dual-mode ASR, to train a single end-to-end ASR model with shared weights for both streaming and full-context speech recognition. We show that the latency and accuracy of streaming ASR significantly benefit from weight sharing and joint training of full-context ASR, especially with inplace knowledge distillation during the training. The Dual-mode ASR framework can be applied to recent state-of-the-art convolution-based and transformer-based ASR networks. We present extensive experiments with two state-of-the-art ASR networks, ContextNet and Conformer, on two datasets, a widely used public dataset LibriSpeech and a large-scale dataset MultiDomain. Experiments and ablation studies demonstrate that Dual-mode ASR not only simplifies the workflow of training and deploying streaming and full-context ASR models, but also significantly improves both emission latency and recognition accuracy of streaming ASR. With Dual-mode ASR, we achieve new state-of-the-art streaming ASR results on both LibriSpeech and MultiDomain in terms of accuracy and latency.
Knowledge distillation (KD) is an effective technique to transfer knowledge from one neural network (teacher) to another (student), thus improving the performance of the student. To make the student better mimic the behavior of the teacher, the existing work focuses on designing different criteria to align their logits or representations. Different from these efforts, we address knowledge distillation from a novel data perspective. We argue that transferring knowledge at sparse training data points cannot enable the student to well capture the local shape of the teacher function. To address this issue, we propose locally linear region knowledge distillation ($\rm L^2$RKD) which transfers the knowledge in local, linear regions from a teacher to a student. This is achieved by enforcing the student to mimic the outputs of the teacher function in local, linear regions. To the end, the student is able to better capture the local shape of the teacher function and thus achieves a better performance. Despite its simplicity, extensive experiments demonstrate that $\rm L^2$RKD is superior to the original KD in many aspects as it outperforms KD and the other state-of-the-art approaches by a large margin, shows robustness and superiority under few-shot settings, and is more compatible with the existing distillation approaches to further improve their performances significantly.