Self-Distillation Amplifies Regularization in Hilbert Space
DeepMind；Google Research（谷歌研究）；EECS Dept., University of California at Berkeley（加州大学伯克利分校电气工程与计算机科学系）；University of California at Berkeley（加州大学伯克利分校）；Google（谷歌）
Hossein Mobahi, Mehrdad Farajtabar, Peter L. Bartlett
Knowledge distillation introduced in the deep learning context is a method to transfer knowledge from one architecture to another. In particular, when the architectures are identical, this is called self-distillation. The idea is to feed in predictions of the trained model as new target values for retraining (and iterate this loop possibly a few times). It has been empirically observed that the self-distilled model often achieves higher accuracy on held out data. Why this happens, however, has been a mystery: the self-distillation dynamics does not receive any new information about the task and solely evolves by looping over training. To the best of our knowledge, there is no rigorous understanding of this phenomenon. This work provides the first theoretical analysis of self-distillation. We focus on fitting a nonlinear function to training data, where the model space is Hilbert space and fitting is subject to $\ell_2$ regularization in this function space. We show that self-distillation iterations modify regularization by progressively limiting the number of basis functions that can be used to represent the solution. This implies (as we also verify empirically) that while a few rounds of self-distillation may reduce over-fitting, further rounds may lead to under-fitting and thus worse performance.
Knowledge distillation is classically a procedure where a neural network is trained on the output of another network along with the original targets in order to transfer knowledge between the architectures. The special case of self-distillation, where the network architectures are identical, has been observed to improve generalization accuracy. In this paper, we consider an iterative variant of self-distillation in a kernel regression setting, in which successive steps incorporate both model outputs and the ground-truth targets. This allows us to provide the first theoretical results on the im…
South China University of Technology（华南理工大学）；Northwestern Polytechnical University（西北工业大学）
Inspired by the strong correlation between the Label Smoothing Regularization(LSR) and Knowledge distillation(KD), we propose an algorithm LsrKD for training boost by extending the LSR method to the KD regime and applying a softer temperature. Then we improve the LsrKD by a Teacher Correction(TC) method, which manually sets a constant larger proportion for the right class in the uniform distribution teacher. To further improve the performance of LsrKD, we develop a self-distillation method named Memory-replay Knowledge Distillation (MrKD) that provides a knowledgeable teacher to replace the un…
Stanford University（斯坦福大学）；Ecole Polytechnique Fédérale de Lausanne（洛桑联邦理工学院）
We study the use of knowledge distillation to compress the U-net architecture. We show that, while standard distillation is not sufficient to reliably train a compressed U-net, introducing other regularization methods, such as batch normalization and class re-weighting, in knowledge distillation significantly improves the training process. This allows us to compress a U-net by over 1000x, i.e., to 0.1% of its original number of parameters, at a negligible decrease in performance.
LG CNS AI Research（LG CNS人工智能研究院）；Seoul National University of Science and Technology（首尔科学技术大学）
The generalization capability of deep neural networks has been substantially improved by applying a wide spectrum of regularization methods, e.g., restricting function space, injecting randomness during training, augmenting data, etc. In this work, we propose a simple yet effective regularization method named progressive self-knowledge distillation (PS-KD), which progressively distills a model's own knowledge to soften hard targets (i.e., one-hot vectors) during training. Hence, it can be interpreted within a framework of knowledge distillation as a student becomes a teacher itself. Specifical…
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-triv…
Ping An Technology (Shenzhen) Co., Ltd.（平安科技（深圳）有限公司）
Knowledge distillation refers to a technique of transferring the knowledge from a large learned model or an ensemble of learned models to a small model. This method relies on access to the original training set, which might not always be available. A possible solution is a data-free adversarial distillation framework, which deploys a generative network to transfer the teacher model's knowledge to the student model. However, the data generation efficiency is low in the data-free adversarial distillation. We add an activation regularizer and a virtual interpolation method to improve the data gen…
AI-KD: Adversarial learning and Implicit regularization for self-Knowledge Distillation
Korea Advanced Institute of Science and Technology (KAIST)（韩国科学技术院（KAIST））；Samsung Electro-Mechanics（三星电机）；German Research Center for Artificial Intelligence (DFKI)（德国人工智能研究中心）；RPTU Kaiserslautern-Landau（凯泽斯劳滕-兰道工业大学）
National University of Singapore（新加坡国立大学）；Huawei Noah’s Ark Lab（华为诺亚方舟实验室）
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 poo…
Horizon Robotics（地平线机器人）；University at Buffalo（布法罗大学）；Tsinghua University（清华大学）
Knowledge distillation is an effective approach to leverage a well-trained network or an ensemble of them, named as the teacher, to guide the training of a student network. The outputs from the teacher network are used as soft labels for supervising the training of a new network. Recent studies \citep{muller2019does,yuan2020revisiting} revealed an intriguing property of the soft labels that making labels soft serves as a good regularization to the student network. From the perspective of statistical learning, regularization aims to reduce the variance, however how bias and variance change is n…
Courant Institute of Mathematical Sciences（柯朗数学科学研究所）；New York University（纽约大学）；University of Texas at Austin（德克萨斯大学奥斯汀分校）；Center of Data Science（数据科学中心）
Knowledge distillation is widely used as a means of improving the performance of a relatively simple student model using the predictions from a complex teacher model. Several works have shown that distillation significantly boosts the student's overall performance; however, are these gains uniform across all data subgroups? In this paper, we show that distillation can harm performance on certain subgroups, e.g., classes with few associated samples. We trace this behaviour to errors made by the teacher distribution being transferred to and amplified by the student model. To mitigate this proble…
We address the problem of detecting speech directed to a device that does not contain a specific wake-word. Specifically, we focus on audio coming from a touch-based invocation. Mitigating virtual assistants (VAs) activation due to accidental button presses is critical for user experience. While the majority of approaches to false trigger mitigation (FTM) are designed to detect the presence of a target keyword, inferring user intent in absence of keyword is difficult. This also poses a challenge when creating the training/evaluation data for such systems due to inherent ambiguity in the user's…
Tsinghua University（清华大学）；Chinese University of Hong Kong（香港中文大学）；PengCheng Laboratory（鹏城实验室）
Recently, a variety of regularization techniques have been widely applied in deep neural networks, such as dropout, batch normalization, data augmentation, and so on. These methods mainly focus on the regularization of weight parameters to prevent overfitting effectively. In addition, label regularization techniques such as label smoothing and label disturbance have also been proposed with the motivation of adding a stochastic perturbation to labels. In this paper, we propose a novel adaptive label regularization method, which enables the neural network to learn from the erroneous experience a…
Knowledge Distillation (KD) is a model-agnostic technique to improve model quality while having a fixed capacity budget. It is a commonly used technique for model compression, where a larger capacity teacher model with better quality is used to train a more compact student model with better inference efficiency. Through distillation, one hopes to benefit from student's compactness, without sacrificing too much on model quality. Despite the large success of knowledge distillation, better understanding of how it benefits student model's training dynamics remains under-explored. In this paper, we…
Korea Advanced Institute of Science and Technology（韩国科学技术院）；University of California, Berkeley（加州大学伯克利分校）
Deep neural networks with millions of parameters may suffer from poor generalization due to overfitting. To mitigate the issue, we propose a new regularization method that penalizes the predictive distribution between similar samples. In particular, we distill the predictive distribution between different samples of the same label during training. This results in regularizing the dark knowledge (i.e., the knowledge on wrong predictions) of a single network (i.e., a self-knowledge distillation) by forcing it to produce more meaningful and consistent predictions in a class-wise manner. Consequen…
Momentum Posterior Regularization for Multi-hop Dense Retrieval
State Key Laboratory for Novel Software Technology, Nanjing University（南京大学现代软件技术国家重点实验室）；School of Computer Science, Nanjing University（南京大学计算机科学学院）；School of Artificial Intelligence, Nanjing University（南京大学人工智能学院）
To put a state-of-the-art neural network to practical use, it is necessary to design a model that has a good trade-off between the resource consumption and performance on the test set. Many researchers and engineers are developing methods that enable training or designing a model more efficiently. Developing an efficient model includes several strategies such as network architecture search, pruning, quantization, knowledge distillation, utilizing cheap convolution, regularization, and also includes any craft that leads to a better performance-resource trade-off. When combining these technologi…
Lawrence Livermore National Laboratory（劳伦斯利弗莫尔国家实验室）；IBM Research（IBM研究院）
Weakly supervised instance labeling using only image-level labels, in lieu of expensive fine-grained pixel annotations, is crucial in several applications including medical image analysis. In contrast to conventional instance segmentation scenarios in computer vision, the problems that we consider are characterized by a small number of training images and non-local patterns that lead to the diagnosis. In this paper, we explore the use of multiple instance learning (MIL) to design an instance label generator under this weakly supervised setting. Motivated by the observation that an MIL model ca…
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 …
Knowledge Distillation and Dataset Distillation of Large Language Models: Emerging Trends, Challenges, and Future Directions
University of Georgia（佐治亚大学）；The University of Texas at Arlington（德克萨斯大学阿灵顿分校）；Harvard University（哈佛大学）；Carnegie Mellon University（卡内基梅隆大学）；Vanderbilt University（范德堡大学）；Mayo Clinic Arizona（亚利桑那梅奥诊所）；Augusta University（奥古斯塔大学）
New Jersey Institute of Technology（新泽西理工学院）；Microsoft Research（微软研究院）；Rice University（莱斯大学）
Recent work has focused on compressing pre-trained language models (PLMs) like BERT where the major focus has been to improve the in-distribution performance for downstream tasks. However, very few of these studies have analyzed the impact of compression on the generalizability and robustness of compressed models for out-of-distribution (OOD) data. Towards this end, we study two popular model compression techniques including knowledge distillation and pruning and show that the compressed models are significantly less robust than their PLM counterparts on OOD test sets although they obtain simi…
Ajou University（亚洲大学）；Hyundai Motor Company（现代汽车公司）；Naver AI Lab（Naver人工智能实验室）
Point-cloud based 3D object detectors recently have achieved remarkable progress. However, most studies are limited to the development of network architectures for improving only their accuracy without consideration of the computational efficiency. In this paper, we first propose an autoencoder-style framework comprising channel-wise compression and decompression via interchange transfer-based knowledge distillation. To learn the map-view feature of a teacher network, the features from teacher and student networks are independently passed through the shared autoencoder; here, we use a compress…
Seoul National University（首尔国立大学）；Qualcomm AI Research（高通人工智能研究院）；Qualcomm Technologies, Inc.（高通技术公司）
Quantization and Knowledge distillation (KD) methods are widely used to reduce memory and power consumption of deep neural networks (DNNs), especially for resource-constrained edge devices. Although their combination is quite promising to meet these requirements, it may not work as desired. It is mainly because the regularization effect of KD further diminishes the already reduced representation power of a quantized model. To address this short-coming, we propose Quantization-aware Knowledge Distillation (QKD) wherein quantization and KD are care-fully coordinated in three phases. First, Self-…
Overconfidence has been shown to impair generalization and calibration of a neural network. Previous studies remedy this issue by adding a regularization term to a loss function, preventing a model from making a peaked distribution. Label smoothing smoothes target labels with a pre-defined prior label distribution; as a result, a model is learned to maximize the likelihood of predicting the soft label. Nonetheless, the amount of smoothing is the same in all samples and remains fixed in training. In other words, label smoothing does not reflect the change in probability distribution mapped by a…
Although deep Reinforcement Learning (RL) has proven successful in a wide range of tasks, one challenge it faces is interpretability when applied to real-world problems. Saliency maps are frequently used to provide interpretability for deep neural networks. However, in the RL domain, existing saliency map approaches are either computationally expensive and thus cannot satisfy the real-time requirement of real-world scenarios or cannot produce interpretable saliency maps for RL policies. In this work, we propose an approach of Distillation with selective Input Gradient Regularization (DIGR) whi…
It has been recently demonstrated that multi-generational self-distillation can improve generalization. Despite this intriguing observation, reasons for the enhancement remain poorly understood. In this paper, we first demonstrate experimentally that the improved performance of multi-generational self-distillation is in part associated with the increasing diversity in teacher predictions. With this in mind, we offer a new interpretation for teacher-student training as amortized MAP estimation, such that teacher predictions enable instance-specific regularization. Our framework allows us to the…
SAE-FD: Sparse Autoencoder Feature Distillation for Continual Learning of Large Language Models
The Hong Kong University of Science and Technology (Guangzhou)（香港科技大学（广州））；Nanjing University of Aeronautics and Astronautics（南京航空航天大学）；The 63rd Research Institute, National University of Defense Technology, Nanjing（国防科技大学第六十三研究所，南京）
Methods for improving deep neural network training times and model generalizability consist of various data augmentation, regularization, and optimization approaches, which tend to be sensitive to hyperparameter settings and make reproducibility more challenging. This work jointly considers two recent training strategies that address model generalizability: sharpness-aware minimization, and self-distillation, and proposes the novel training strategy of Sharpness-Aware Distilled Teachers (SADT). The experimental section of this work shows that SADT consistently outperforms previously published …
The University of Tokyo（东京大学）；National Institute of Informatics（国立信息学研究所）；Wuhan University（武汉大学）
Most deep metric learning (DML) methods employ a strategy that forces all positive samples to be close in the embedding space while keeping them away from negative ones. However, such a strategy ignores the internal relationships of positive (negative) samples and often leads to overfitting, especially in the presence of hard samples and mislabeled samples. In this work, we propose a simple yet effective regularization, namely Listwise Self-Distillation (LSD), which progressively distills a model's own knowledge to adaptively assign a more appropriate distance target to each sample pair in a b…
Zhejiang University（浙江大学）；Zhejiang University of Technology（浙江工业大学）
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 automat…
Department of Computer Science, University of California, Los Angeles, CA, USA（加州大学洛杉矶分校计算机科学系）；Princeton AI Laboratory, Princeton University, Princeton, NJ, USA（普林斯顿大学普林斯顿AI实验室）
IBM Research Almaden（IBM阿尔马登研究中心）；Lawrence Livermore National Labs（劳伦斯利弗莫尔国家实验室）
Automated diagnostic assistants in healthcare necessitate accurate AI models that can be trained with limited labeled data, can cope with severe class imbalances and can support simultaneous prediction of multiple disease conditions. To this end, we present a deep learning framework that utilizes a number of key components to enable robust modeling in such challenging scenarios. Using an important use-case in chest X-ray classification, we provide several key insights on the effective use of data augmentation, self-training via distillation and confidence tempering for small data learning in m…
TRACER: Persistent Regularization for Robust Multimodal Finetuning
School of Computing and Information Systems (CIS), Faculty of Engineering and IT (FEIT), University of Melbourne, Australia（墨尔本大学计算机科学与信息系统学院（CIS）、工程与信息技术学院（FEIT））
