PyTorchKR Papers

최신 논문

총 4,526편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2110.13471

Response-based Distillation for Incremental Object Detection

Traditional object detection are ill-equipped for incremental learning. However, fine-tuning directly on a well-trained detection model with only new data will leads to catastrophic forgetting. Knowledge distillation is a straightforward way to mitigate catastrophic forgetting. In Incremental Object Detection (IOD), previous work mainly focuses on feature-level knowledge distillation, but the different response of detector has not been fully explored yet. In this paper, we propose a fully response-based incremental distillation method focusing on learning response from detection bounding boxes and classification predictions. Firstly, our method transferring category knowledge while equipping student model with the ability to retain localization knowledge during incremental learning. In addition, we further evaluate the qualities of all locations and provides valuable response by adaptive pseudo-label selection (APS) strategies. Finally, we elucidate that knowledge from different responses should be assigned with different importance during incremental distillation. Extensive experiments conducted on MS COCO demonstrate significant advantages of our method, which substantially narrow the performance gap towards full training.

arXiv:2110.12072

How and When Adversarial Robustness Transfers in Knowledge Distillation?

Knowledge distillation (KD) has been widely used in teacher-student training, with applications to model compression in resource-constrained deep learning. Current works mainly focus on preserving the accuracy of the teacher model. However, other important model properties, such as adversarial robustness, can be lost during distillation. This paper studies how and when the adversarial robustness can be transferred from a teacher model to a student model in KD. We show that standard KD training fails to preserve adversarial robustness, and we propose KD with input gradient alignment (KDIGA) for remedy. Under certain assumptions, we prove that the student model using our proposed KDIGA can achieve at least the same certified robustness as the teacher model. Our experiments of KD contain a diverse set of teacher and student models with varying network architectures and sizes evaluated on ImageNet and CIFAR-10 datasets, including residual neural networks (ResNets) and vision transformers (ViTs). Our comprehensive analysis shows several novel insights that (1) With KDIGA, students can preserve or even exceed the adversarial robustness of the teacher model, even when their models have fundamentally different architectures; (2) KDIGA enables robustness to transfer to pre-trained students, such as KD from an adversarially trained ResNet to a pre-trained ViT, without loss of clean accuracy; and (3) Our derived local linearity bounds for characterizing adversarial robustness in KD are consistent with the empirical results.

arXiv:2110.12844

Reconstructing Pruned Filters using Cheap Spatial Transformations

We present an efficient alternative to the convolutional layer using cheap spatial transformations. This construction exploits an inherent spatial redundancy of the learned convolutional filters to enable a much greater parameter efficiency, while maintaining the top-end accuracy of their dense counter-parts. Training these networks is modelled as a generalised pruning problem, whereby the pruned filters are replaced with cheap transformations from the set of non-pruned filters. We provide an efficient implementation of the proposed layer, followed by two natural extensions to avoid excessive feature compression and to improve the expressivity of the transformed features. We show that these networks can achieve comparable or improved performance to state-of-the-art pruning models across both the CIFAR-10 and ImageNet-1K datasets.

arXiv:2110.12724공식 코드 · Papers with Code

Instance-Conditional Knowledge Distillation for Object Detection

Knowledge distillation has shown great success in classification, however, it is still challenging for detection. In a typical image for detection, representations from different locations may have different contributions to detection targets, making the distillation hard to balance. In this paper, we propose a conditional distillation framework to distill the desired knowledge, namely knowledge that is beneficial in terms of both classification and localization for every instance. The framework introduces a learnable conditional decoding module, which retrieves information given each target instance as query. Specifically, we encode the condition information as query and use the teacher's representations as key. The attention between query and key is used to measure the contribution of different features, guided by a localization-recognition-sensitive auxiliary task. Extensive experiments demonstrate the efficacy of our method: we observe impressive improvements under various settings. Notably, we boost RetinaNet with ResNet-50 backbone from 37.4 to 40.7 mAP (+3.3) under 1x schedule, that even surpasses the teacher (40.4 mAP) with ResNet-101 backbone under 3x schedule. Code has been released on https://github.com/megvii-research/ICD.

arXiv:2110.12606

MUSE: Feature Self-Distillation with Mutual Information and Self-Information

We present a novel information-theoretic approach to introduce dependency among features of a deep convolutional neural network (CNN). The core idea of our proposed method, called MUSE, is to combine MUtual information and SElf-information to jointly improve the expressivity of all features extracted from different layers in a CNN. We present two variants of the realization of MUSE -- Additive Information and Multiplicative Information. Importantly, we argue and empirically demonstrate that MUSE, compared to other feature discrepancy functions, is a more functional proxy to introduce dependency and effectively improve the expressivity of all features in the knowledge distillation framework. MUSE achieves superior performance over a variety of popular architectures and feature discrepancy functions for self-distillation and online distillation, and performs competitively with the state-of-the-art methods for offline distillation. MUSE is also demonstrably versatile that enables it to be easily extended to CNN-based models on tasks other than image classification such as object detection.

arXiv:2110.12516

X-Distill: Improving Self-Supervised Monocular Depth via Cross-Task Distillation

In this paper, we propose a novel method, X-Distill, to improve the self-supervised training of monocular depth via cross-task knowledge distillation from semantic segmentation to depth estimation. More specifically, during training, we utilize a pretrained semantic segmentation teacher network and transfer its semantic knowledge to the depth network. In order to enable such knowledge distillation across two different visual tasks, we introduce a small, trainable network that translates the predicted depth map to a semantic segmentation map, which can then be supervised by the teacher network. In this way, this small network enables the backpropagation from the semantic segmentation teacher's supervision to the depth network during training. In addition, since the commonly used object classes in semantic segmentation are not directly transferable to depth, we study the visual and geometric characteristics of the objects and design a new way of grouping them that can be shared by both tasks. It is noteworthy that our approach only modifies the training process and does not incur additional computation during inference. We extensively evaluate the efficacy of our proposed approach on the standard KITTI benchmark and compare it with the latest state of the art. We further test the generalizability of our approach on Make3D. Overall, the results show that our approach significantly improves the depth estimation accuracy and outperforms the state of the art.

arXiv:2110.11650공식 코드 · Papers with Code

Pixel-by-Pixel Cross-Domain Alignment for Few-Shot Semantic Segmentation

In this paper we consider the task of semantic segmentation in autonomous driving applications. Specifically, we consider the cross-domain few-shot setting where training can use only few real-world annotated images and many annotated synthetic images. In this context, aligning the domains is made more challenging by the pixel-wise class imbalance that is intrinsic in the segmentation and that leads to ignoring the underrepresented classes and overfitting the well represented ones. We address this problem with a novel framework called Pixel-By-Pixel Cross-Domain Alignment (PixDA). We propose a novel pixel-by-pixel domain adversarial loss following three criteria: (i) align the source and the target domain for each pixel, (ii) avoid negative transfer on the correctly represented pixels, and (iii) regularize the training of infrequent classes to avoid overfitting. The pixel-wise adversarial training is assisted by a novel sample selection procedure, that handles the imbalance between source and target data, and a knowledge distillation strategy, that avoids overfitting towards the few target images. We demonstrate on standard synthetic-to-real benchmarks that PixDA outperforms previous state-of-the-art methods in (1-5)-shot settings.

arXiv:2110.11545

Pseudo Supervised Monocular Depth Estimation with Teacher-Student Network

Despite recent improvement of supervised monocular depth estimation, the lack of high quality pixel-wise ground truth annotations has become a major hurdle for further progress. In this work, we propose a new unsupervised depth estimation method based on pseudo supervision mechanism by training a teacher-student network with knowledge distillation. It strategically integrates the advantages of supervised and unsupervised monocular depth estimation, as well as unsupervised binocular depth estimation. Specifically, the teacher network takes advantage of the effectiveness of binocular depth estimation to produce accurate disparity maps, which are then used as the pseudo ground truth to train the student network for monocular depth estimation. This effectively converts the problem of unsupervised learning to supervised learning. Our extensive experimental results demonstrate that the proposed method outperforms the state-of-the-art on the KITTI benchmark.

arXiv:2110.11023

Augmenting Knowledge Distillation With Peer-To-Peer Mutual Learning For Model Compression

Knowledge distillation (KD) is an effective model compression technique where a compact student network is taught to mimic the behavior of a complex and highly trained teacher network. In contrast, Mutual Learning (ML) provides an alternative strategy where multiple simple student networks benefit from sharing knowledge, even in the absence of a powerful but static teacher network. Motivated by these findings, we propose a single-teacher, multi-student framework that leverages both KD and ML to achieve better performance. Furthermore, an online distillation strategy is utilized to train the teacher and students simultaneously. To evaluate the performance of the proposed approach, extensive experiments were conducted using three different versions of teacher-student networks on benchmark biomedical classification (MSI vs. MSS) and object detection (Polyp Detection) tasks. Ensemble of student networks trained in the proposed manner achieved better results than the ensemble of students trained using KD or ML individually, establishing the benefit of augmenting knowledge transfer from teacher to students with peer-to-peer learning between students.

arXiv:2110.10741

Class Incremental Online Streaming Learning

A wide variety of methods have been developed to enable lifelong learning in conventional deep neural networks. However, to succeed, these methods require a `batch' of samples to be available and visited multiple times during training. While this works well in a static setting, these methods continue to suffer in a more realistic situation where data arrives in \emph{online streaming manner}. We empirically demonstrate that the performance of current approaches degrades if the input is obtained as a stream of data with the following restrictions: $(i)$ each instance comes one at a time and can be seen only once, and $(ii)$ the input data violates the i.i.d assumption, i.e., there can be a class-based correlation. We propose a novel approach (CIOSL) for the class-incremental learning in an \emph{online streaming setting} to address these challenges. The proposed approach leverages implicit and explicit dual weight regularization and experience replay. The implicit regularization is leveraged via the knowledge distillation, while the explicit regularization incorporates a novel approach for parameter regularization by learning the joint distribution of the buffer replay and the current sample. Also, we propose an efficient online memory replay and replacement buffer strategy that significantly boosts the model's performance. Extensive experiments and ablation on challenging datasets show the efficacy of the proposed method.

arXiv:2110.10429

Knowledge distillation from language model to acoustic model: a hierarchical multi-task learning approach

The remarkable performance of the pre-trained language model (LM) using self-supervised learning has led to a major paradigm shift in the study of natural language processing. In line with these changes, leveraging the performance of speech recognition systems with massive deep learning-based LMs is a major topic of speech recognition research. Among the various methods of applying LMs to speech recognition systems, in this paper, we focus on a cross-modal knowledge distillation method that transfers knowledge between two types of deep neural networks with different modalities. We propose an acoustic model structure with multiple auxiliary output layers for cross-modal distillation and demonstrate that the proposed method effectively compensates for the shortcomings of the existing label-interpolation-based distillation method. In addition, we extend the proposed method to a hierarchical distillation method using LMs trained in different units (senones, monophones, and subwords) and reveal the effectiveness of the hierarchical distillation method through an ablation study.

arXiv:2110.09910공식 코드 · Papers with Code

FedHe: Heterogeneous Models and Communication-Efficient Federated Learning

Federated learning (FL) is able to manage edge devices to cooperatively train a model while maintaining the training data local and private. One common assumption in FL is that all edge devices share the same machine learning model in training, for example, identical neural network architecture. However, the computation and store capability of different devices may not be the same. Moreover, reducing communication overheads can improve the training efficiency though it is still a challenging problem in FL. In this paper, we propose a novel FL method, called FedHe, inspired by knowledge distillation, which can train heterogeneous models and support asynchronous training processes with significantly reduced communication overheads. Our analysis and experimental results demonstrate that the performance of our proposed method is better than the state-of-the-art algorithms in terms of communication overheads and model accuracy.

arXiv:2110.09674공식 코드 · Papers with Code

Adaptive Distillation: Aggregating Knowledge from Multiple Paths for Efficient Distillation

Knowledge Distillation is becoming one of the primary trends among neural network compression algorithms to improve the generalization performance of a smaller student model with guidance from a larger teacher model. This momentous rise in applications of knowledge distillation is accompanied by the introduction of numerous algorithms for distilling the knowledge such as soft targets and hint layers. Despite this advancement in different techniques for distilling the knowledge, the aggregation of different paths for distillation has not been studied comprehensively. This is of particular significance, not only because different paths have different importance, but also due to the fact that some paths might have negative effects on the generalization performance of the student model. Hence, we need to adaptively adjust the importance of each path to maximize the impact of distillation on the student model. In this paper, we explore different approaches for aggregating these different paths and introduce our proposed adaptive approach based on multitask learning methods. We empirically demonstrate the effectiveness of the proposed approach over other baselines on the applications of knowledge distillation in classification, semantic segmentation, and object detection tasks.

arXiv:2110.08532

Pro-KD: Progressive Distillation by Following the Footsteps of the Teacher

With ever growing scale of neural models, knowledge distillation (KD) attracts more attention as a prominent tool for neural model compression. However, there are counter intuitive observations in the literature showing some challenging limitations of KD. A case in point is that the best performing checkpoint of the teacher might not necessarily be the best teacher for training the student in KD. Therefore, one important question would be how to find the best checkpoint of the teacher for distillation? Searching through the checkpoints of the teacher would be a very tedious and computationally expensive process, which we refer to as the \textit{checkpoint-search problem}. Moreover, another observation is that larger teachers might not necessarily be better teachers in KD which is referred to as the \textit{capacity-gap} problem. To address these challenging problems, in this work, we introduce our progressive knowledge distillation (Pro-KD) technique which defines a smoother training path for the student by following the training footprints of the teacher instead of solely relying on distilling from a single mature fully-trained teacher. We demonstrate that our technique is quite effective in mitigating the capacity-gap problem and the checkpoint search problem. We evaluate our technique using a comprehensive set of experiments on different tasks such as image classification (CIFAR-10 and CIFAR-100), natural language understanding tasks of the GLUE benchmark, and question answering (SQuAD 1.1 and 2.0) using BERT-based models and consistently got superior results over state-of-the-art techniques.

arXiv:2110.08460

A Short Study on Compressing Decoder-Based Language Models

Pre-trained Language Models (PLMs) have been successful for a wide range of natural language processing (NLP) tasks. The state-of-the-art of PLMs, however, are extremely large to be used on edge devices. As a result, the topic of model compression has attracted increasing attention in the NLP community. Most of the existing works focus on compressing encoder-based models (tiny-BERT, distilBERT, distilRoBERTa, etc), however, to the best of our knowledge, the compression of decoder-based models (such as GPT-2) has not been investigated much. Our paper aims to fill this gap. Specifically, we explore two directions: 1) we employ current state-of-the-art knowledge distillation techniques to improve fine-tuning of DistilGPT-2. 2) we pre-train a compressed GPT-2 model using layer truncation and compare it against the distillation-based method (DistilGPT2). The training time of our compressed model is significantly less than DistilGPT-2, but it can achieve better performance when fine-tuned on downstream tasks. We also demonstrate the impact of data cleaning on model performance.

arXiv:2110.08419

Robustness Challenges in Model Distillation and Pruning for Natural Language Understanding

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 similar performance on in-distribution development sets for a task. Further analysis indicates that the compressed models overfit on the shortcut samples and generalize poorly on the hard ones. We further leverage this observation to develop a regularization strategy for robust model compression based on sample uncertainty. Experimental results on several natural language understanding tasks demonstrate that our bias mitigation framework improves the OOD generalization of the compressed models, while not sacrificing the in-distribution task performance.

arXiv:2110.08270

From Multimodal to Unimodal Attention in Transformers using Knowledge Distillation

Multimodal Deep Learning has garnered much interest, and transformers have triggered novel approaches, thanks to the cross-attention mechanism. Here we propose an approach to deal with two key existing challenges: the high computational resource demanded and the issue of missing modalities. We introduce for the first time the concept of knowledge distillation in transformers to use only one modality at inference time. We report a full study analyzing multiple student-teacher configurations, levels at which distillation is applied, and different methodologies. With the best configuration, we improved the state-of-the-art accuracy by 3%, we reduced the number of parameters by 2.5 times and the inference time by 22%. Such performance-computation tradeoff can be exploited in many applications and we aim at opening a new research area where the deployment of complex models with limited resources is demanded.

arXiv:2110.08727공식 코드 · Papers with Code

Graph-less Neural Networks: Teaching Old MLPs New Tricks via Distillation

Graph Neural Networks (GNNs) are popular for graph machine learning and have shown great results on wide node classification tasks. Yet, they are less popular for practical deployments in the industry owing to their scalability challenges incurred by data dependency. Namely, GNN inference depends on neighbor nodes multiple hops away from the target, and fetching them burdens latency-constrained applications. Existing inference acceleration methods like pruning and quantization can speed up GNNs by reducing Multiplication-and-ACcumulation (MAC) operations, but the improvements are limited given the data dependency is not resolved. Conversely, multi-layer perceptrons (MLPs) have no graph dependency and infer much faster than GNNs, even though they are less accurate than GNNs for node classification in general. Motivated by these complementary strengths and weaknesses, we bring GNNs and MLPs together via knowledge distillation (KD). Our work shows that the performance of MLPs can be improved by large margins with GNN KD. We call the distilled MLPs Graph-less Neural Networks (GLNNs) as they have no inference graph dependency. We show that GLNNs with competitive accuracy infer faster than GNNs by 146X-273X and faster than other acceleration methods by 14X-27X. Under a production setting involving both transductive and inductive predictions across 7 datasets, GLNN accuracies improve over stand-alone MLPs by 12.36% on average and match GNNs on 6/7 datasets. Comprehensive analysis shows when and why GLNNs can achieve competitive accuracies to GNNs and suggests GLNN as a handy choice for latency-constrained applications.

arXiv:2110.08551공식 코드 · Papers with Code

HRKD: Hierarchical Relational Knowledge Distillation for Cross-domain Language Model Compression

On many natural language processing tasks, large pre-trained language models (PLMs) have shown overwhelming performances compared with traditional neural network methods. Nevertheless, their huge model size and low inference speed have hindered the deployment on resource-limited devices in practice. In this paper, we target to compress PLMs with knowledge distillation, and propose a hierarchical relational knowledge distillation (HRKD) method to capture both hierarchical and domain relational information. Specifically, to enhance the model capability and transferability, we leverage the idea of meta-learning and set up domain-relational graphs to capture the relational information across different domains. And to dynamically select the most representative prototypes for each domain, we propose a hierarchical compare-aggregate mechanism to capture hierarchical relationships. Extensive experiments on public multi-domain datasets demonstrate the superior performance of our HRKD method as well as its strong few-shot learning ability. For reproducibility, we release the code at https://github.com/cheneydon/hrkd.

arXiv:2110.08190

Sparse Progressive Distillation: Resolving Overfitting under Pretrain-and-Finetune Paradigm

Conventional wisdom in pruning Transformer-based language models is that pruning reduces the model expressiveness and thus is more likely to underfit rather than overfit. However, under the trending pretrain-and-finetune paradigm, we postulate a counter-traditional hypothesis, that is: pruning increases the risk of overfitting when performed at the fine-tuning phase. In this paper, we aim to address the overfitting problem and improve pruning performance via progressive knowledge distillation with error-bound properties. We show for the first time that reducing the risk of overfitting can help the effectiveness of pruning under the pretrain-and-finetune paradigm. Ablation studies and experiments on the GLUE benchmark show that our method outperforms the leading competitors across different tasks.