Knowledge distillation (KD) which transfers the knowledge from a large teacher model to a small student model, has been widely used to compress the BERT model recently. Besides the supervision in the output in the original KD, recent works show that layer-level supervision is crucial to the performance of the student BERT model. However, previous works designed the layer mapping strategy heuristically (e.g., uniform or last-layer), which can lead to inferior performance. In this paper, we propose to use the genetic algorithm (GA) to search for the optimal layer mapping automatically. To accelerate the search process, we further propose a proxy setting where a small portion of the training corpus are sampled for distillation, and three representative tasks are chosen for evaluation. After obtaining the optimal layer mapping, we perform the task-agnostic BERT distillation with it on the whole corpus to build a compact student model, which can be directly fine-tuned on downstream tasks. Comprehensive experiments on the evaluation benchmarks demonstrate that 1) layer mapping strategy has a significant effect on task-agnostic BERT distillation and different layer mappings can result in quite different performances; 2) the optimal layer mapping strategy from the proposed search process consistently outperforms the other heuristic ones; 3) with the optimal layer mapping, our student model achieves state-of-the-art performance on the GLUE tasks.
In natural language processing (NLP) tasks, slow inference speed and huge footprints in GPU usage remain the bottleneck of applying pre-trained deep models in production. As a popular method for model compression, knowledge distillation transfers knowledge from one or multiple large (teacher) models to a small (student) model. When multiple teacher models are available in distillation, the state-of-the-art methods assign a fixed weight to a teacher model in the whole distillation. Furthermore, most of the existing methods allocate an equal weight to every teacher model. In this paper, we observe that, due to the complexity of training examples and the differences in student model capability, learning differentially from teacher models can lead to better performance of student models distilled. We systematically develop a reinforced method to dynamically assign weights to teacher models for different training instances and optimize the performance of student model. Our extensive experimental results on several NLP tasks clearly verify the feasibility and effectiveness of our approach.
Knowledge distillation is one of the most popular and effective techniques for knowledge transfer, model compression and semi-supervised learning. Most existing distillation approaches require the access to original or augmented training samples. But this can be problematic in practice due to privacy, proprietary and availability concerns. Recent work has put forward some methods to tackle this problem, but they are either highly time-consuming or unable to scale to large datasets. To this end, we propose a new method to train a generative image model by leveraging the intrinsic normalization layers' statistics of the trained teacher network. This enables us to build an ensemble of generators without training data that can efficiently produce substitute inputs for subsequent distillation. The proposed method pushes forward the data-free distillation performance on CIFAR-10 and CIFAR-100 to 95.02% and 77.02% respectively. Furthermore, we are able to scale it to ImageNet dataset, which to the best of our knowledge, has never been done using generative models in a data-free setting.
Direct speech translation (ST) has shown to be a complex task requiring knowledge transfer from its sub-tasks: automatic speech recognition (ASR) and machine translation (MT). For MT, one of the most promising techniques to transfer knowledge is knowledge distillation. In this paper, we compare the different solutions to distill knowledge in a sequence-to-sequence task like ST. Moreover, we analyze eventual drawbacks of this approach and how to alleviate them maintaining the benefits in terms of translation quality.
Knowledge distillation has demonstrated encouraging performances in deep model compression. Most existing approaches, however, require massive labeled data to accomplish the knowledge transfer, making the model compression a cumbersome and costly process. In this paper, we investigate the practical few-shot knowledge distillation scenario, where we assume only a few samples without human annotations are available for each category. To this end, we introduce a principled dual-stage distillation scheme tailored for few-shot data. In the first step, we graft the student blocks one by one onto the teacher, and learn the parameters of the grafted block intertwined with those of the other teacher blocks. In the second step, the trained student blocks are progressively connected and then together grafted onto the teacher network, allowing the learned student blocks to adapt themselves to each other and eventually replace the teacher network. Experiments demonstrate that our approach, with only a few unlabeled samples, achieves gratifying results on CIFAR10, CIFAR100, and ILSVRC-2012. On CIFAR10 and CIFAR100, our performances are even on par with those of knowledge distillation schemes that utilize the full datasets. The source code is available at https://github.com/zju-vipa/NetGraft.
The task of information retrieval is an important component of many natural language processing systems, such as open domain question answering. While traditional methods were based on hand-crafted features, continuous representations based on neural networks recently obtained competitive results. A challenge of using such methods is to obtain supervised data to train the retriever model, corresponding to pairs of query and support documents. In this paper, we propose a technique to learn retriever models for downstream tasks, inspired by knowledge distillation, and which does not require annotated pairs of query and documents. Our approach leverages attention scores of a reader model, used to solve the task based on retrieved documents, to obtain synthetic labels for the retriever. We evaluate our method on question answering, obtaining state-of-the-art results.
Recent recommender systems have started to employ knowledge distillation, which is a model compression technique distilling knowledge from a cumbersome model (teacher) to a compact model (student), to reduce inference latency while maintaining performance. The state-of-the-art methods have only focused on making the student model accurately imitate the predictions of the teacher model. They have a limitation in that the prediction results incompletely reveal the teacher's knowledge. In this paper, we propose a novel knowledge distillation framework for recommender system, called DE-RRD, which enables the student model to learn from the latent knowledge encoded in the teacher model as well as from the teacher's predictions. Concretely, DE-RRD consists of two methods: 1) Distillation Experts (DE) that directly transfers the latent knowledge from the teacher model. DE exploits "experts" and a novel expert selection strategy for effectively distilling the vast teacher's knowledge to the student with limited capacity. 2) Relaxed Ranking Distillation (RRD) that transfers the knowledge revealed from the teacher's prediction with consideration of the relaxed ranking orders among items. Our extensive experiments show that DE-RRD outperforms the state-of-the-art competitors and achieves comparable or even better performance to that of the teacher model with faster inference time.
Despite the recent success on image classification, self-training has only achieved limited gains on structured prediction tasks such as neural machine translation (NMT). This is mainly due to the compositionality of the target space, where the far-away prediction hypotheses lead to the notorious reinforced mistake problem. In this paper, we revisit the utilization of multiple diverse models and present a simple yet effective approach named Reciprocal-Supervised Learning (RSL). RSL first exploits individual models to generate pseudo parallel data, and then cooperatively trains each model on the combined synthetic corpus. RSL leverages the fact that different parameterized models have different inductive biases, and better predictions can be made by jointly exploiting the agreement among each other. Unlike the previous knowledge distillation methods built upon a much stronger teacher, RSL is capable of boosting the accuracy of one model by introducing other comparable or even weaker models. RSL can also be viewed as a more efficient alternative to ensemble. Extensive experiments demonstrate the superior performance of RSL on several benchmarks with significant margins.
Deep neural networks (DNNs) have been extremely successful in solving many challenging AI tasks in natural language processing, speech recognition, and computer vision nowadays. However, DNNs are typically computation intensive, memory demanding, and power hungry, which significantly limits their usage on platforms with constrained resources. Therefore, a variety of compression techniques (e.g. quantization, pruning, and knowledge distillation) have been proposed to reduce the size and power consumption of DNNs. Blockwise knowledge distillation is one of the compression techniques that can effectively reduce the size of a highly complex DNN. However, it is not widely adopted due to its long training time. In this paper, we propose a novel parallel blockwise distillation algorithm to accelerate the distillation process of sophisticated DNNs. Our algorithm leverages local information to conduct independent blockwise distillation, utilizes depthwise separable layers as the efficient replacement block architecture, and properly addresses limiting factors (e.g. dependency, synchronization, and load balancing) that affect parallelism. The experimental results running on an AMD server with four Geforce RTX 2080Ti GPUs show that our algorithm can achieve 3x speedup plus 19% energy savings on VGG distillation, and 3.5x speedup plus 29% energy savings on ResNet distillation, both with negligible accuracy loss. The speedup of ResNet distillation can be further improved to 3.87 when using four RTX6000 GPUs in a distributed cluster.
Knowledge distillation is a technique to enhance the generalization ability of a student model by exploiting outputs from a teacher model. Recently, feature-map based variants explore knowledge transfer between manually assigned teacher-student pairs in intermediate layers for further improvement. However, layer semantics may vary in different neural networks and semantic mismatch in manual layer associations will lead to performance degeneration due to negative regularization. To address this issue, we propose Semantic Calibration for cross-layer Knowledge Distillation (SemCKD), which automatically assigns proper target layers of the teacher model for each student layer with an attention mechanism. With a learned attention distribution, each student layer distills knowledge contained in multiple teacher layers rather than a specific intermediate layer for appropriate cross-layer supervision. We further provide theoretical analysis of the association weights and conduct extensive experiments to demonstrate the effectiveness of our approach. Code is avaliable at \url{https://github.com/DefangChen/SemCKD}.
Model compression methods are important to allow for easier deployment of deep learning models in compute, memory and energy-constrained environments such as mobile phones. Knowledge distillation is a class of model compression algorithm where knowledge from a large teacher network is transferred to a smaller student network thereby improving the student's performance. In this paper, we show how optimal transport-based loss functions can be used for training a student network which encourages learning student network parameters that help bring the distribution of student features closer to that of the teacher features. We present image classification results on CIFAR-100, SVHN and ImageNet and show that the proposed optimal transport loss functions perform comparably to or better than other loss functions.
Several methods of knowledge distillation have been developed for neural network compression. While they all use the KL divergence loss to align the soft outputs of the student model more closely with that of the teacher, the various methods differ in how the intermediate features of the student are encouraged to match those of the teacher. In this paper, we propose a simple-to-implement method using auxiliary classifiers at intermediate layers for matching features, which we refer to as multi-head knowledge distillation (MHKD). We add loss terms for training the student that measure the dissimilarity between student and teacher outputs of the auxiliary classifiers. At the same time, the proposed method also provides a natural way to measure differences at the intermediate layers even though the dimensions of the internal teacher and student features may be different. Through several experiments in image classification on multiple datasets we show that the proposed method outperforms prior relevant approaches presented in the literature.
Knowledge distillation (KD) is a general neural network training approach that uses a teacher model to guide the student model. Existing works mainly study KD from the network output side (e.g., trying to design a better KD loss function), while few have attempted to understand it from the input side. Especially, its interplay with data augmentation (DA) has not been well understood. In this paper, we ask: Why do some DA schemes (e.g., CutMix) inherently perform much better than others in KD? What makes a "good" DA in KD? Our investigation from a statistical perspective suggests that a good DA scheme should reduce the covariance of the teacher-student cross-entropy. A practical metric, the stddev of teacher's mean probability (T. stddev), is further presented and well justified empirically. Besides the theoretical understanding, we also introduce a new entropy-based data-mixing DA scheme, CutMixPick, to further enhance CutMix. Extensive empirical studies support our claims and demonstrate how we can harvest considerable performance gains simply by using a better DA scheme in knowledge distillation.
With the rise in edge-computing devices, there has been an increasing demand to deploy energy and resource-efficient models. A large body of research has been devoted to developing methods that can reduce the size of the model considerably without affecting the standard metrics such as top-1 accuracy. However, these pruning approaches tend to result in a significant mismatch in other metrics such as fairness across classes and explainability. To combat such misalignment, we propose a novel multi-part loss function inspired by the knowledge-distillation literature. Through extensive experiments, we demonstrate the effectiveness of our approach across different compression algorithms, architectures, tasks as well as datasets. In particular, we obtain up to $4.1\times$ reduction in the number of prediction mismatches between the compressed and reference models, and up to $5.7\times$ in cases where the reference model makes the correct prediction; all while making no changes to the compression algorithm, and minor modifications to the loss function. Furthermore, we demonstrate how inducing simple alignment between the predictions of the models naturally improves the alignment on other metrics including fairness and attributions. Our framework can thus serve as a simple plug-and-play component for compression algorithms in the future.
We propose a deep learning-based channel estimation, quantization, feedback, and precoding method for downlink multiuser multiple-input and multiple-output systems. In the proposed system, channel estimation and quantization for limited feedback are handled by a receiver deep neural network (DNN). Precoder selection is handled by a transmitter DNN. To emulate the traditional channel quantization, a binarization layer is adopted at each receiver DNN, and the binarization layer is also used to enable end-to-end learning. However, this can lead to inaccurate gradients, which can trap the receiver DNNs at a poor local minimum during training. To address this, we consider knowledge distillation, in which the existing DNNs are jointly trained with an auxiliary transmitter DNN. The use of an auxiliary DNN as a teacher network allows the receiver DNNs to additionally exploit lossless gradients, which is useful in avoiding a poor local minimum. For the same number of feedback bits, our DNN-based precoding scheme can achieve a higher downlink rate compared to conventional linear precoding with codebook-based limited feedback.
Semantic segmentation models have two fundamental weaknesses: i) they require large training sets with costly pixel-level annotations, and ii) they have a static output space, constrained to the classes of the training set. Toward addressing both problems, we introduce a new task, Incremental Few-Shot Segmentation (iFSS). The goal of iFSS is to extend a pretrained segmentation model with new classes from few annotated images and without access to old training data. To overcome the limitations of existing models iniFSS, we propose Prototype-based Incremental Few-Shot Segmentation (PIFS) that couples prototype learning and knowledge distillation. PIFS exploits prototypes to initialize the classifiers of new classes, fine-tuning the network to refine its features representation. We design a prototype-based distillation loss on the scores of both old and new class prototypes to avoid overfitting and forgetting, and batch-renormalization to cope with non-i.i.d.few-shot data. We create an extensive benchmark for iFSS showing that PIFS outperforms several few-shot and incremental learning methods in all scenarios.
Pre-trained language models have been applied to various NLP tasks with considerable performance gains. However, the large model sizes, together with the long inference time, limit the deployment of such models in real-time applications. One line of model compression approaches considers knowledge distillation to distill large teacher models into small student models. Most of these studies focus on single-domain only, which ignores the transferable knowledge from other domains. We notice that training a teacher with transferable knowledge digested across domains can achieve better generalization capability to help knowledge distillation. Hence we propose a Meta-Knowledge Distillation (Meta-KD) framework to build a meta-teacher model that captures transferable knowledge across domains and passes such knowledge to students. Specifically, we explicitly force the meta-teacher to capture transferable knowledge at both instance-level and feature-level from multiple domains, and then propose a meta-distillation algorithm to learn single-domain student models with guidance from the meta-teacher. Experiments on public multi-domain NLP tasks show the effectiveness and superiority of the proposed Meta-KD framework. Further, we also demonstrate the capability of Meta-KD in the settings where the training data is scarce.
In recent years the empirical success of transfer learning with neural networks has stimulated an increasing interest in obtaining a theoretical understanding of its core properties. Knowledge distillation where a smaller neural network is trained using the outputs of a larger neural network is a particularly interesting case of transfer learning. In the present work, we introduce a statistical physics framework that allows an analytic characterization of the properties of knowledge distillation (KD) in shallow neural networks. Focusing the analysis on a solvable model that exhibits a non-trivial generalization gap, we investigate the effectiveness of KD. We are able to show that, through KD, the regularization properties of the larger teacher model can be inherited by the smaller student and that the yielded generalization performance is closely linked to and limited by the optimality of the teacher. Finally, we analyze the double descent phenomenology that can arise in the considered KD setting.
Knowledge distillation (KD) has become an important technique for model compression and knowledge transfer. In this work, we first perform a comprehensive analysis of the knowledge transferred by different KD methods. We demonstrate that traditional KD methods, which minimize the KL divergence of softmax outputs between networks, are related to the knowledge alignment of an individual sample only. Meanwhile, recent contrastive learning-based KD methods mainly transfer relational knowledge between different samples, namely, knowledge correlation. While it is important to transfer the full knowledge from teacher to student, we introduce the Multi-level Knowledge Distillation (MLKD) by effectively considering both knowledge alignment and correlation. MLKD is task-agnostic and model-agnostic, and can easily transfer knowledge from supervised or self-supervised pretrained teachers. We show that MLKD can improve the reliability and transferability of learned representations. Experiments demonstrate that MLKD outperforms other state-of-the-art methods on a large number of experimental settings including different (a) pretraining strategies (b) network architectures (c) datasets (d) tasks.
In recent years, the growing size of neural networks has led to a vast amount of research concerning compression techniques to mitigate the drawbacks of such large sizes. Most of these research works can be categorized into three broad families : Knowledge Distillation, Pruning, and Quantization. While there has been steady research in this domain, adoption and commercial usage of the proposed techniques has not quite progressed at the rate. We present KD-Lib, an open-source PyTorch based library, which contains state-of-the-art modular implementations of algorithms from the three families on top of multiple abstraction layers. KD-Lib is model and algorithm-agnostic, with extended support for hyperparameter tuning using Optuna and Tensorboard for logging and monitoring. The library can be found at - https://github.com/SforAiDl/KD_Lib.