In this paper, we propose Stochastic Knowledge Distillation (SKD) to obtain compact BERT-style language model dubbed SKDBERT. In each iteration, SKD samples a teacher model from a pre-defined teacher ensemble, which consists of multiple teacher models with multi-level capacities, to transfer knowledge into student model in an one-to-one manner. Sampling distribution plays an important role in SKD. We heuristically present three types of sampling distributions to assign appropriate probabilities for multi-level teacher models. SKD has two advantages: 1) it can preserve the diversities of multi-level teacher models via stochastically sampling single teacher model in each iteration, and 2) it can also improve the efficacy of knowledge distillation via multi-level teacher models when large capacity gap exists between the teacher model and the student model. Experimental results on GLUE benchmark show that SKDBERT reduces the size of a BERT$_{\rm BASE}$ model by 40% while retaining 99.5% performances of language understanding and being 100% faster.
translated by 谷歌翻译
Language model pre-training, such as BERT, has significantly improved the performances of many natural language processing tasks. However, pre-trained language models are usually computationally expensive, so it is difficult to efficiently execute them on resourcerestricted devices. To accelerate inference and reduce model size while maintaining accuracy, we first propose a novel Transformer distillation method that is specially designed for knowledge distillation (KD) of the Transformer-based models. By leveraging this new KD method, the plenty of knowledge encoded in a large "teacher" BERT can be effectively transferred to a small "student" Tiny-BERT. Then, we introduce a new two-stage learning framework for TinyBERT, which performs Transformer distillation at both the pretraining and task-specific learning stages. This framework ensures that TinyBERT can capture the general-domain as well as the task-specific knowledge in BERT. TinyBERT 41 with 4 layers is empirically effective and achieves more than 96.8% the performance of its teacher BERT BASE on GLUE benchmark, while being 7.5x smaller and 9.4x faster on inference. TinyBERT 4 is also significantly better than 4-layer state-of-the-art baselines on BERT distillation, with only ∼28% parameters and ∼31% inference time of them. Moreover, TinyBERT 6 with 6 layers performs on-par with its teacher BERT BASE .
translated by 谷歌翻译
Knowledge Distillation (KD) has been extensively used for natural language understanding (NLU) tasks to improve a small model's (a student) generalization by transferring the knowledge from a larger model (a teacher). Although KD methods achieve state-of-the-art performance in numerous settings, they suffer from several problems limiting their performance. It is shown in the literature that the capacity gap between the teacher and the student networks can make KD ineffective. Additionally, existing KD techniques do not mitigate the noise in the teacher's output: modeling the noisy behaviour of the teacher can distract the student from learning more useful features. We propose a new KD method that addresses these problems and facilitates the training compared to previous techniques. Inspired by continuation optimization, we design a training procedure that optimizes the highly non-convex KD objective by starting with the smoothed version of this objective and making it more complex as the training proceeds. Our method (Continuation-KD) achieves state-of-the-art performance across various compact architectures on NLU (GLUE benchmark) and computer vision tasks (CIFAR-10 and CIFAR-100).
translated by 谷歌翻译
已经证明了对比学习适合学习句子嵌入,可以显着提高语义文本相似性(STS)任务。最近,大型对比学习模型,例如句子T5倾向于学到更强大的句子嵌入。虽然有效,但由于计算资源或时间成本限制,这种大型型号很难在线服务。为了解决这个问题,通常采用知识蒸馏(KD),这可以将大型“教师”模型压缩成一个小的“学生”模型,但通常会遭受一些性能损失。在这里,我们提出了一个增强的KD框架,称为蒸馏 - 对比度(迪斯科)。所提出的迪斯科框架首先利用KD将大句子嵌入模型的能力转移到大型未标记数据的小学生模型,然后在标记的训练数据上具有对比学习的学生模型。对于迪斯科舞厅的KD进程,我们进一步提出了对比的知识蒸馏(CKD),以增强教师模型培训,KD和学生模型的一致性,这可能会提高迅速学习的表现。 7 STS基准测试的广泛实验表明,使用所提出的迪斯科和CKD培训的学生模型很少或甚至没有性能损失,并且始终如一地优于相同参数大小的相应对应物。令人惊讶的是,我们的110米学生模型甚至可以优于最新的最新(SOTA)模型,即句子T5(11B),只有1%的参数。
translated by 谷歌翻译
先前的研究证明,跨语性知识蒸馏可以显着提高预训练模型的跨语义相似性匹配任务的性能。但是,在此操作中,学生模型必须大。否则,其性能将急剧下降,从而使部署到内存限制设备的不切实际。为了解决这个问题,我们深入研究了跨语言知识蒸馏,并提出了一个多阶段蒸馏框架,用于构建一个小型但高性能的跨语性模型。在我们的框架中,合并了对比度学习,瓶颈和参数复发策略,以防止在压缩过程中损害性能。实验结果表明,我们的方法可以压缩XLM-R和Minilm的大小超过50 \%,而性能仅降低约1%。
translated by 谷歌翻译
由于从大规模预先训练的语言模型的转移学习在自然语言处理中普遍存在,在计算受限环境中运行这些模型仍然是一个具有挑战性的问题。已经提出了包括知识蒸馏,网络量化或网络修剪的几种解决方案;然而,这些方法主要关注英语,从而在考虑低资源语言时扩大差距。在这项工作中,我们为罗马尼亚语推出了三种轻型和快速版本的罗马尼亚语言:Distil-Bert-Base-Ro,Distil-Robert-Base和DistilMulti-Bert-Bas-Ro。前两种模型因单独蒸馏在文献中提供的两个基础版本的罗马尼亚伯爵的知识,而最后一个是通过蒸馏它们的集合来获得的。为了我们的知识,这是第一次尝试创建公开可用的罗马尼亚蒸馏BERT模型,这是在五个任务上进行彻底评估的:语音标记,名为实体识别,情感分析,语义文本相似性和方言识别。这些基准测试的实验结果证明,我们的三种蒸馏模型在与老师的准确性方面保持最大的表现,而GPU的两倍于GPU和〜35 \%较小。此外,我们进一步测试了我们的学生和他们的老师之间的相似性,通过测量其标签和概率忠诚度以及回归忠诚度 - 在这项工作中引入的新指标。
translated by 谷歌翻译
鉴于将语言模型转移到NLP任务的成功,我们询问全BERT模型是否始终是最好的,并且它存在一个简单但有效的方法,可以在没有的最先进的深神经网络中找到获胜的票复杂的计算。我们构建了一系列基于BERT的模型,具有不同的大小,并对8个二进制分类任务进行比较。结果表明,真正存在的较小的子网比完整模型更好。然后我们提供进一步的研究,并提出一种简单的方法在微调之前适当地收缩斜率。一些扩展实验表明,我们的方法可以省略甚至没有准确性损失的时间和存储开销。
translated by 谷歌翻译
最初引入了知识蒸馏,以利用来自单一教师模型的额外监督为学生模型培训。为了提高学生表现,最近的一些变体试图利用多个教师利用不同的知识来源。然而,现有研究主要通过对多种教师预测的平均或将它们与其他无标签策略相结合,将知识集成在多种来源中,可能在可能存在低质量的教师预测存在中误导学生。为了解决这个问题,我们提出了信心感知的多教师知识蒸馏(CA-MKD),该知识蒸馏(CA-MKD)在地面真理标签的帮助下,适用于每个教师预测的样本明智的可靠性,与那些接近单热的教师预测标签分配了大量的重量。此外,CA-MKD包含中间层,以进一步提高学生表现。广泛的实验表明,我们的CA-MKD始终如一地优于各种教师学生架构的所有最先进的方法。
translated by 谷歌翻译
用于预培训语言模型的自我监督学习的核心包括预训练任务设计以及适当的数据增强。语言模型中的大多数数据增强都是独立于上下文的。最近在电子中提出了一个开创性的增强,并通过引入辅助生成网络(发电机)来实现最先进的性能,以产生用于培训主要辨别网络(鉴别者)的上下文化数据增强。然而,这种设计引入了发电机的额外计算成本,并且需要调整发电机和鉴别器之间的相对能力。在本文中,我们提出了一种自增强策略(SAS),其中单个网络用于审视以后的时期的培训常规预训练和上下文化数据增强。基本上,该策略消除了单独的发电机,并使用单个网络共同执行具有MLM(屏蔽语言建模)和RTD(替换令牌检测)头的两个预训练任务。它避免了寻找适当大小的发电机的挑战,这对于在电子中证明的性能至关重要,以及其随后的变体模型至关重要。此外,SAS是一项常规策略,可以与最近或将来的许多新技术无缝地结合,例如杜伯塔省的解除关注机制。我们的实验表明,SAS能够在具有相似或更少的计算成本中优于胶水任务中的电磁和其他最先进的模型。
translated by 谷歌翻译
随着AI芯片(例如GPU,TPU和NPU)的改进以及物联网(IOT)的快速发展,一些强大的深神经网络(DNN)通常由数百万甚至数亿个参数组成,这些参数是可能不适合直接部署在低计算和低容量单元(例如边缘设备)上。最近,知识蒸馏(KD)被认为是模型压缩的有效方法之一,以减少模型参数。 KD的主要概念是从大型模型(即教师模型)的特征图中提取有用的信息,以引用成功训练一个小型模型(即学生模型),该模型大小比老师小得多。尽管已经提出了许多基于KD的方法来利用教师模型中中间层的特征图中的信息,但是,它们中的大多数并未考虑教师模型和学生模型之间的特征图的相似性,这可能让学生模型学习无用的信息。受到注意机制的启发,我们提出了一种新颖的KD方法,称为代表教师钥匙(RTK),该方法不仅考虑了特征地图的相似性,而且还会过滤掉无用的信息以提高目标学生模型的性能。在实验中,我们使用多个骨干网络(例如Resnet和wideresnet)和数据集(例如CIFAR10,CIFAR100,SVHN和CINIC10)验证了我们提出的方法。结果表明,我们提出的RTK可以有效地提高基于注意的KD方法的分类精度。
translated by 谷歌翻译
将最新的变压器模型蒸馏成轻量级的学生模型是降低推理时计算成本的有效方法。学生模型通常是紧凑的变压器,参数较少,而昂贵的操作(例如自我发项)持续存在。因此,对于实时或大量用例,提高的推理速度仍然不令人满意。在本文中,我们旨在通过将教师模型提炼成更大,更稀疏的学生模型来进一步推动推理速度的极限 - 更大的是它们扩展到数十亿个参数;稀疏,大多数模型参数是N-gram嵌入。我们对六个单词文本分类任务的实验表明,这些学生模型平均保留了罗伯塔大师教师表现的97%,同时推理时GPU和CPU的加速速度最高为600倍。进一步的调查表明,我们的管道也有助于句子对分类任务和域泛化设置。
translated by 谷歌翻译
近年来,知识蒸馏有显着改善,可以为更好的效率产生紧凑的学生模型,同时保留教师模型的模型效果。以前的研究发现:由于能力不匹配,更准确的教师对更好的教师无需。在本文中,我们旨在通过模型校准的角度分析现象。我们发现较大的教师模型可能过于过度自信,因此学生模型无法有效地模仿。虽然,在教师模型的简单模型校准之后,教师模型的大小与学生模型的性能具有正相关。
translated by 谷歌翻译
知识蒸馏是一种通过减少差异来将有关陈述信息从教师转移到学生的方法。这种方法的一个挑战是减少学生表现的灵活性,从而导致对教师知识的学习不准确。为了解决BERT转移,我们研究了指定为三种类型的表示结构的蒸馏:功能内,局部局部互感,全局功能间结构。要转移它们,我们基于中心内核对齐方式介绍了\ textit {特征结构蒸馏}方法,该方法为相似的特征结构分配了一致的价值,并揭示了更有信息的关系。特别是,针对全局结构实现了一种带有聚类的内存调节方法。在对胶合数据集的语言理解的九项任务的实验中,与最新的蒸馏方法相比,提出的方法有效地传递了三种类型的结构并提高性能。实际上,这些方法的代码可在https://github.com/maroo-sky/fsd中获得
translated by 谷歌翻译
在过去的几年中,基于变压器的预训练的语言模型在行业和学术界都取得了惊人的成功。但是,较大的模型尺寸和高运行时间延迟是在实践中应用它们的严重障碍,尤其是在手机和物联网(IoT)设备上。为了压缩该模型,最近有大量文献围绕知识蒸馏(KD)的主题长大。然而,KD在基于变压器的模型中的工作方式仍不清楚。我们取消了KD的组件,并提出了一个统一的KD框架。通过框架,花费了23,000多个GPU小时的系统和广泛的实验,从知识类型的角度,匹配策略,宽度深度折衷,初始化,型号大小等。在培训前语言模型中,对先前最新的(SOTA)的相对显着改善。最后,我们为基于变压器模型的KD提供了最佳实践指南。
translated by 谷歌翻译
Knowledge distillation is often used to transfer knowledge from a strong teacher model to a relatively weak student model. Traditional knowledge distillation methods include response-based methods and feature-based methods. Response-based methods are used the most widely but suffer from lower upper limit of model performance, while feature-based methods have constraints on the vocabularies and tokenizers. In this paper, we propose a tokenizer-free method liberal feature-based distillation (LEAD). LEAD aligns the distribution between teacher model and student model, which is effective, extendable, portable and has no requirements on vocabularies, tokenizer, or model architecture. Extensive experiments show the effectiveness of LEAD on several widely-used benchmarks, including MS MARCO Passage, TREC Passage 19, TREC Passage 20, MS MARCO Document, TREC Document 19 and TREC Document 20.
translated by 谷歌翻译
Knowledge Distillation (KD) aims to distill the knowledge of a cumbersome teacher model into a lightweight student model. Its success is generally attributed to the privileged information on similarities among categories provided by the teacher model, and in this sense, only strong teacher models are deployed to teach weaker students in practice. In this work, we challenge this common belief by following experimental observations: 1) beyond the acknowledgment that the teacher can improve the student, the student can also enhance the teacher significantly by reversing the KD procedure; 2) a poorly-trained teacher with much lower accuracy than the student can still improve the latter significantly. To explain these observations, we provide a theoretical analysis of the relationships between KD and label smoothing regularization. We prove that 1) KD is a type of learned label smoothing regularization and 2) label smoothing regularization provides a virtual teacher model for KD. From these results, we argue that the success of KD is not fully due to the similarity information between categories from teachers, but also to the regularization of soft targets, which is equally or even more important.Based on these analyses, we further propose a novel Teacher-free Knowledge Distillation (Tf-KD) framework, where a student model learns from itself or manuallydesigned regularization distribution. The Tf-KD achieves comparable performance with normal KD from a superior teacher, which is well applied when a stronger teacher model is unavailable. Meanwhile, Tf-KD is generic and can be directly deployed for training deep neural networks. Without any extra computation cost, Tf-KD achieves up to 0.65% improvement on ImageNet over well-established baseline models, which is superior to label smoothing regularization.
translated by 谷歌翻译
Knowledge distillation (KD) has been widely used for model compression and knowledge transfer. Typically, a big teacher model trained on sufficient data transfers knowledge to a small student model. However, despite the success of KD, little effort has been made to study whether KD leaks the training data of the teacher model. In this paper, we experimentally reveal that KD suffers from the risk of privacy leakage. To alleviate this issue, we propose a novel knowledge distillation method, swing distillation, which can effectively protect the private information of the teacher model from flowing to the student model. In our framework, the temperature coefficient is dynamically and adaptively adjusted according to the degree of private information contained in the data, rather than a predefined constant hyperparameter. It assigns different temperatures to tokens according to the likelihood that a token in a position contains private information. In addition, we inject noise into soft targets provided to the student model, in order to avoid unshielded knowledge transfer. Experiments on multiple datasets and tasks demonstrate that the proposed swing distillation can significantly reduce (by over 80% in terms of canary exposure) the risk of privacy leakage in comparison to KD with competitive or better performance. Furthermore, swing distillation is robust against the increasing privacy budget.
translated by 谷歌翻译
我们从任务特定的BERT基教师模型执行知识蒸馏(KD)基准到各种学生模型:Bilstm,CNN,Bert-Tiny,Bert-Mini和Bert-small。我们的实验涉及在两个任务中分组的12个数据集:印度尼西亚语言中的文本分类和序列标记。我们还比较蒸馏的各个方面,包括使用Word Embeddings和未标记的数据增强的使用。我们的实验表明,尽管基于变压器的模型的普及程度不断上升,但是使用Bilstm和CNN学生模型,与修剪的BERT模型相比,使用Bilstm和CNN学生模型提供了性能和计算资源(CPU,RAM和存储)之间的最佳权衡。我们进一步提出了一些快速胜利,通过涉及涉及丢失功能,Word Embeddings和未标记的数据准备的简单选择的高效KD培训机制来生产小型NLP模型。
translated by 谷歌翻译
大型的语言模型(PRELMS)正在彻底改变所有基准的自然语言处理。但是,它们的巨大尺寸对于小型实验室或移动设备上的部署而言是过分的。修剪和蒸馏等方法可减少模型尺寸,但通常保留相同的模型体系结构。相反,我们探索了蒸馏预告片中的更有效的架构,单词的持续乘法(CMOW),该构造将每个单词嵌入为矩阵,并使用矩阵乘法来编码序列。我们扩展了CMOW体系结构及其CMOW/CBOW-HYBRID变体,具有双向组件,以提供更具表现力的功能,在预绘制期间进行一般(任务无义的)蒸馏的单次表示,并提供了两种序列编码方案,可促进下游任务。句子对,例如句子相似性和自然语言推断。我们的基于矩阵的双向CMOW/CBOW-HYBRID模型在问题相似性和识别文本范围内的Distilbert具有竞争力,但仅使用参数数量的一半,并且在推理速度方面快三倍。除了情感分析任务SST-2和语言可接受性任务COLA外,我们匹配或超过ELMO的ELMO分数。但是,与以前的跨架结构蒸馏方法相比,我们证明了检测语言可接受性的分数增加了一倍。这表明基于基质的嵌入可用于将大型预赛提炼成竞争模型,并激励朝这个方向进行进一步的研究。
translated by 谷歌翻译
本文研究了从预先训练的模型,尤其是蒙面自动编码器中提取知识的潜力。我们的方法很简单:除了优化掩盖输入的像素重建损失外,我们还将教师模型的中间特征图与学生模型的中间特征图之间的距离最小化。此设计导致一个计算高效的知识蒸馏框架,给定1)仅使用一个少量可见的补丁子集,2)(笨拙的)教师模型仅需要部分执行,\ ie,\ ie,在前几个中,向前传播输入层,用于获得中间特征图。与直接蒸馏微型模型相比,提炼预训练的模型显着改善了下游性能。例如,通过将知识从MAE预先训练的VIT-L提炼为VIT-B,我们的方法可实现84.0%的Imagenet Top-1精度,表现优于直接将微型VIT-L蒸馏的基线,降低1.2%。更有趣的是,我们的方法即使具有极高的掩盖率也可以从教师模型中进行鲁棒性蒸馏:例如,在蒸馏过程中仅可见十个斑块,我们的VIT-B具有竞争力的前1个Imagenet精度为83.6%,在95%的掩盖率中,只有十个斑块。 ;令人惊讶的是,它仍然可以通过仅四个可见斑(98%的掩盖率)积极训练来确保82.4%的Top-1 Imagenet精度。代码和模型可在https://github.com/ucsc-vlaa/dmae上公开获得。
translated by 谷歌翻译