How can we efficiently compress a model while maintaining its performance? Knowledge Distillation (KD) is one of the widely known methods for model compression. In essence, KD trains a smaller student model based on a larger teacher model and tries to retain the teacher model's level of performance as much as possible. However, existing KD methods suffer from the following limitations. First, since the student model is smaller in absolute size, it inherently lacks model capacity. Second, the absence of an initial guide for the student model makes it difficult for the student to imitate the teacher model to its fullest. Conventional KD methods yield low performance due to these limitations. In this paper, we propose Pea-KD (Parameter-efficient and accurate Knowledge Distillation), a novel approach to KD. Pea-KD consists of two main parts: Shuffled Parameter Sharing (SPS) and Pretraining with Teacher's Predictions (PTP). Using this combination, we are capable of alleviating the KD's limitations. SPS is a new parameter sharing method that increases the student model capacity. PTP is a KD-specialized initialization method, which can act as a good initial guide for the student. When combined, this method yields a significant increase in student model's performance. Experiments conducted on BERT with different datasets and tasks show that the proposed approach improves the student model's performance by 4.4\% on average in four GLUE tasks, outperforming existing KD baselines by significant margins.
Generative Adversarial Networks (GANs) have become a powerful approach for generative image modeling. However, GANs are notorious for their training instability, especially on large-scale, complex datasets. While the recent work of BigGAN has significantly improved the quality of image generation on ImageNet, it requires a huge model, making it hard to deploy on resource-constrained devices. To reduce the model size, we propose a black-box knowledge distillation framework for compressing GANs, which highlights a stable and efficient training process. Given BigGAN as the teacher network, we manage to train a much smaller student network to mimic its functionality, achieving competitive performance on Inception and FID scores with the generator having $16\times$ fewer parameters.
Existing language model compression methods mostly use a simple L2 loss to distill knowledge in the intermediate representations of a large BERT model to a smaller one. Although widely used, this objective by design assumes that all the dimensions of hidden representations are independent, failing to capture important structural knowledge in the intermediate layers of the teacher network. To achieve better distillation efficacy, we propose Contrastive Distillation on Intermediate Representations (CoDIR), a principled knowledge distillation framework where the student is trained to distill knowledge through intermediate layers of the teacher via a contrastive objective. By learning to distinguish positive sample from a large set of negative samples, CoDIR facilitates the student's exploitation of rich information in teacher's hidden layers. CoDIR can be readily applied to compress large-scale language models in both pre-training and finetuning stages, and achieves superb performance on the GLUE benchmark, outperforming state-of-the-art compression methods.
Transformer-based pre-training models like BERT have achieved remarkable performance in many natural language processing tasks.However, these models are both computation and memory expensive, hindering their deployment to resource-constrained devices. In this work, we propose TernaryBERT, which ternarizes the weights in a fine-tuned BERT model. Specifically, we use both approximation-based and loss-aware ternarization methods and empirically investigate the ternarization granularity of different parts of BERT. Moreover, to reduce the accuracy degradation caused by the lower capacity of low bits, we leverage the knowledge distillation technique in the training process. Experiments on the GLUE benchmark and SQuAD show that our proposed TernaryBERT outperforms the other BERT quantization methods, and even achieves comparable performance as the full-precision model while being 14.9x smaller.
Deep reinforcement learning has recently seen huge success across multiple areas in the robotics domain. Owing to the limitations of gathering real-world data, i.e., sample inefficiency and the cost of collecting it, simulation environments are utilized for training the different agents. This not only aids in providing a potentially infinite data source, but also alleviates safety concerns with real robots. Nonetheless, the gap between the simulated and real worlds degrades the performance of the policies once the models are transferred into real robots. Multiple research efforts are therefore now being directed towards closing this sim-to-real gap and accomplish more efficient policy transfer. Recent years have seen the emergence of multiple methods applicable to different domains, but there is a lack, to the best of our knowledge, of a comprehensive review summarizing and putting into context the different methods. In this survey paper, we cover the fundamental background behind sim-to-real transfer in deep reinforcement learning and overview the main methods being utilized at the moment: domain randomization, domain adaptation, imitation learning, meta-learning and knowledge distillation. We categorize some of the most relevant recent works, and outline the main application scenarios. Finally, we discuss the main opportunities and challenges of the different approaches and point to the most promising directions.
This paper is an attempt to bridge the gap between deep learning and grammatical inference. Indeed, it provides an algorithm to extract a (stochastic) formal language from any recurrent neural network trained for language modelling. In detail, the algorithm uses the already trained network as an oracle -- and thus does not require the access to the inner representation of the black-box -- and applies a spectral approach to infer a weighted automaton. As weighted automata compute linear functions, they are computationally more efficient than neural networks and thus the nature of the approach is the one of knowledge distillation. We detail experiments on 62 data sets (both synthetic and from real-world applications) that allow an in-depth study of the abilities of the proposed algorithm. The results show the WA we extract are good approximations of the RNN, validating the approach. Moreover, we show how the process provides interesting insights toward the behavior of RNN learned on data, enlarging the scope of this work to the one of explainability of deep learning models.
Adder Neural Networks (ANNs) which only contain additions bring us a new way of developing deep neural networks with low energy consumption. Unfortunately, there is an accuracy drop when replacing all convolution filters by adder filters. The main reason here is the optimization difficulty of ANNs using $\ell_1$-norm, in which the estimation of gradient in back propagation is inaccurate. In this paper, we present a novel method for further improving the performance of ANNs without increasing the trainable parameters via a progressive kernel based knowledge distillation (PKKD) method. A convolutional neural network (CNN) with the same architecture is simultaneously initialized and trained as a teacher network, features and weights of ANN and CNN will be transformed to a new space to eliminate the accuracy drop. The similarity is conducted in a higher-dimensional space to disentangle the difference of their distributions using a kernel based method. Finally, the desired ANN is learned based on the information from both the ground-truth and teacher, progressively. The effectiveness of the proposed method for learning ANN with higher performance is then well-verified on several benchmarks. For instance, the ANN-50 trained using the proposed PKKD method obtains a 76.8\% top-1 accuracy on ImageNet dataset, which is 0.6\% higher than that of the ResNet-50.
We introduce \texttt{N-LTP}, an open-source neural language technology platform supporting six fundamental Chinese NLP tasks: {lexical analysis} (Chinese word segmentation, part-of-speech tagging, and named entity recognition), {syntactic parsing} (dependency parsing), and {semantic parsing} (semantic dependency parsing and semantic role labeling). Unlike the existing state-of-the-art toolkits, such as \texttt{Stanza}, that adopt an independent model for each task, \texttt{N-LTP} adopts the multi-task framework by using a shared pre-trained model, which has the advantage of capturing the shared knowledge across relevant Chinese tasks. In addition, a knowledge distillation method \cite{DBLP:journals/corr/abs-1907-04829} where the single-task model teaches the multi-task model is further introduced to encourage the multi-task model to surpass its single-task teacher. Finally, we provide a collection of easy-to-use APIs and a visualization tool to make users to use and view the processing results more easily and directly. To the best of our knowledge, this is the first toolkit to support six Chinese NLP fundamental tasks. Source code, documentation, and pre-trained models are available at \url{https://github.com/HIT-SCIR/ltp}.
A common dilemma in 3D object detection for autonomous driving is that high-quality, dense point clouds are only available during training, but not testing. We use knowledge distillation to bridge the gap between a model trained on high-quality inputs at training time and another tested on low-quality inputs at inference time. In particular, we design a two-stage training pipeline for point cloud object detection. First, we train an object detection model on dense point clouds, which are generated from multiple frames using extra information only available at training time. Then, we train the model's identical counterpart on sparse single-frame point clouds with consistency regularization on features from both models. We show that this procedure improves performance on low-quality data during testing, without additional overhead.
With the boom of edge intelligence, its vulnerability to adversarial attacks becomes an urgent problem. The so-called adversarial example can fool a deep learning model on the edge node to misclassify. Due to the property of transferability, the adversary can easily make a black-box attack using a local substitute model. Nevertheless, the limitation of resource of edge nodes cannot afford a complicated defense mechanism as doing on the cloud data center. To overcome the challenge, we propose a dynamic defense mechanism, namely EI-MTD. It first obtains robust member models with small size through differential knowledge distillation from a complicated teacher model on the cloud data center. Then, a dynamic scheduling policy based on a Bayesian Stackelberg game is applied to the choice of a target model for service. This dynamic defense can prohibit the adversary from selecting an optimal substitute model for black-box attacks. Our experimental result shows that this dynamic scheduling can effectively protect edge intelligence against adversarial attacks under the black-box setting.
Knowledge distillation has been proven to be effective in model acceleration and compression. It allows a small network to learn to generalize in the same way as a large network. Recent successes in pre-training suggest the effectiveness of transferring model parameters. Inspired by this, we investigate methods of model acceleration and compression in another line of research. We propose Weight Distillation to transfer the knowledge in the large network parameters through a parameter generator. Our experiments on WMT16 En-Ro, NIST12 Zh-En, and WMT14 En-De machine translation tasks show that weight distillation can train a small network that is 1.88~2.94x faster than the large network but with competitive performance. With the same sized small network, weight distillation can outperform knowledge distillation by 0.51~1.82 BLEU points.
In recent years, many explanation methods have been proposed to explain individual classifications of deep neural networks. However, how to leverage the created explanations to improve the learning process has been less explored. As the privileged information, the explanations of a model can be used to guide the learning process of the model itself. In the community, another intensively investigated privileged information used to guide the training of a model is the knowledge from a powerful teacher model. The goal of this work is to leverage the self-explanation to improve the learning process by borrowing ideas from knowledge distillation. We start by investigating the effective components of the knowledge transferred from the teacher network to the student network. Our investigation reveals that both the responses in non-ground-truth classes and class-similarity information in teacher's outputs contribute to the success of the knowledge distillation. Motivated by the conclusion, we propose an implementation of introspective learning by distilling knowledge from online self-explanations. The models trained with the introspective learning procedure outperform the ones trained with the standard learning procedure, as well as the ones trained with different regularization methods. When compared to the models learned from peer networks or teacher networks, our models also show competitive performance and requires neither peers nor teachers.
In forensic applications, it is very common that only small naturalistic datasets consisting of short utterances in complex or unknown acoustic environments are available. In this study, we propose a pipeline solution to improve speaker verification on a small actual forensic field dataset. By leveraging large-scale out-of-domain datasets, a knowledge distillation based objective function is proposed for teacher-student learning, which is applied for short utterance forensic speaker verification. The objective function collectively considers speaker classification loss, Kullback-Leibler divergence, and similarity of embeddings. In order to advance the trained deep speaker embedding network to be robust for a small target dataset, we introduce a novel strategy to fine-tune the pre-trained student model towards a forensic target domain by utilizing the model as a finetuning start point and a reference in regularization. The proposed approaches are evaluated on the 1st48-UTD forensic corpus, a newly established naturalistic dataset of actual homicide investigations consisting of short utterances recorded in uncontrolled conditions. We show that the proposed objective function can efficiently improve the performance of teacher-student learning on short utterances and that our fine-tuning strategy outperforms the commonly used weight decay method by providing an explicit inductive bias towards the pre-trained model.
With the success of deep neural networks, knowledge distillation which guides the learning of a small student network from a large teacher network is being actively studied for model compression and transfer learning. However, few studies have been performed to resolve the poor learning issue of the student network when the student and teacher model sizes significantly differ. In this paper, we propose a densely guided knowledge distillation using multiple teacher assistants that gradually decreases the model size to efficiently bridge the large gap between the teacher and student networks. To stimulate more efficient learning of the student network, we guide each teacher assistant to every other smaller teacher assistants iteratively. Specifically, when teaching a smaller teacher assistant at the next step, the existing larger teacher assistants from the previous step are used as well as the teacher network. Moreover, we design stochastic teaching where, for each mini-batch, a teacher or teacher assistants are randomly dropped. This acts as a regularizer to improve the efficiency of teaching of the student network. Thus, the student can always learn salient distilled knowledge from the multiple sources. We verified the effectiveness of the proposed method for a classification task using CIFAR-10, CIFAR-100, and ImageNet. We also achieved significant performance improvements with various backbone architectures such as ResNet, WideResNet, and VGG.
Deep Metric Learning (DML) provides a crucial tool for visual similarity and zero-shot applications by learning generalizing embedding spaces, although recent work in DML has shown strong performance saturation across training objectives. However, generalization capacity is known to scale with the embedding space dimensionality. Unfortunately, high dimensional embeddings also create higher retrieval cost for downstream applications. To remedy this, we propose \emph{Simultaneous Similarity-based Self-distillation (S2SD). S2SD extends DML with knowledge distillation from auxiliary, high-dimensional embedding and feature spaces to leverage complementary context during training while retaining test-time cost and with negligible changes to the training time. Experiments and ablations across different objectives and standard benchmarks show S2SD offers notable improvements of up to 7% in Recall@1, while also setting a new state-of-the-art. Code available at https://github.com/MLforHealth/S2SD.
Pre-trained large-scale language models have increasingly demonstrated high accuracy on many natural language processing (NLP) tasks. However, the limited weight storage and computational speed on hardware platforms have impeded the popularity of pre-trained models, especially in the era of edge computing. In this work, we propose an efficient transformer-based large-scale language representation using hardware-friendly block structure pruning. We incorporate the reweighted group Lasso into block-structured pruning for optimization. Besides the significantly reduced weight storage and computation, the proposed approach achieves high compression rates. Experimental results on different models (BERT, RoBERTa, and DistilBERT) on the General Language Understanding Evaluation (GLUE) benchmark tasks show that we achieve up to 5.0x with zero or minor accuracy degradation on certain task(s). Our proposed method is also orthogonal to existing compact pre-trained language models such as DistilBERT using knowledge distillation, since a further 1.79x average compression rate can be achieved on top of DistilBERT with zero or minor accuracy degradation. It is suitable to deploy the final compressed model on resource-constrained edge devices.
We introduce a simple yet effective distillation framework that is able to boost the vanilla ResNet-50 to 80%+ Top-1 accuracy on ImageNet without tricks. We construct such a framework through analyzing the problems in the existing classification system and simplify the base method ensemble knowledge distillation via discriminators by: (1) adopting the similarity loss and discriminator only on the final outputs and (2) using the average of softmax probabilities from all teacher ensembles as the stronger supervision. Intriguingly, three novel perspectives are presented for distillation: (1) weight decay can be weakened or even completely removed since the soft label also has a regularization effect; (2) using a good initialization for students is critical; and (3) one-hot/hard label is not necessary in the distillation process if the weights are well initialized. We show that such a straight-forward framework can achieve state-of-the-art results without involving any commonly-used techniques, such as architecture modification; outside training data beyond ImageNet; autoaug/randaug; cosine learning rate; mixup/cutmix training; label smoothing; etc. Our method obtains 80.67% top-1 accuracy on ImageNet using a single crop-size of 224x224 with vanilla ResNet-50, outperforming the previous state-of-the-arts by a significant margin under the same network structure. Our result can be regarded as a strong baseline using knowledge distillation, and to our best knowledge, this is also the first method that is able to boost vanilla ResNet-50 to surpass 80% on ImageNet without architecture modification or additional training data. On smaller ResNet-18, our distillation framework consistently improves from 69.76% to 73.19%, which shows tremendous practical values in real-world applications. Our code and models are available at: https://github.com/szq0214/MEAL-V2.
Syntax has been shown useful for various NLP tasks, while existing work mostly encodes singleton syntactic tree using one hierarchical neural network. In this paper, we investigate a simple and effective method, Knowledge Distillation, to integrate heterogeneous structure knowledge into a unified sequential LSTM encoder. Experimental results on four typical syntax-dependent tasks show that our method outperforms tree encoders by effectively integrating rich heterogeneous structure syntax, meanwhile reducing error propagation, and also outperforms ensemble methods, in terms of both the efficiency and accuracy.
Despite the effectiveness of utilizing the BERT model for document ranking, the high computational cost of such approaches limits their uses. To this end, this paper first empirically investigates the effectiveness of two knowledge distillation models on the document ranking task. In addition, on top of the recently proposed TinyBERT model, two simplifications are proposed. Evaluations on two different and widely-used benchmarks demonstrate that Simplified TinyBERT with the proposed simplifications not only boosts TinyBERT, but also significantly outperforms BERT-Base when providing 15$\times$ speedup.
In this paper we apply self-knowledge distillation to text summarization which we argue can alleviate problems with maximum-likelihood training on single reference and noisy datasets. Instead of relying on one-hot annotation labels, our student summarization model is trained with guidance from a teacher which generates smoothed labels to help regularize training. Furthermore, to better model uncertainty during training, we introduce multiple noise signals for both teacher and student models. We demonstrate experimentally on three benchmarks that our framework boosts the performance of both pretrained and non-pretrained summarizers achieving state-of-the-art results.