PyTorchKR Papers

최신 논문

총 4,526편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2111.03664

Oracle Teacher: Leveraging Target Information for Better Knowledge Distillation of CTC Models

Knowledge distillation (KD), best known as an effective method for model compression, aims at transferring the knowledge of a bigger network (teacher) to a much smaller network (student). Conventional KD methods usually employ the teacher model trained in a supervised manner, where output labels are treated only as targets. Extending this supervised scheme further, we introduce a new type of teacher model for connectionist temporal classification (CTC)-based sequence models, namely Oracle Teacher, that leverages both the source inputs and the output labels as the teacher model's input. Since the Oracle Teacher learns a more accurate CTC alignment by referring to the target information, it can provide the student with more optimal guidance. One potential risk for the proposed approach is a trivial solution that the model's output directly copies the target input. Based on a many-to-one mapping property of the CTC algorithm, we present a training strategy that can effectively prevent the trivial solution and thus enables utilizing both source and target inputs for model training. Extensive experiments are conducted on two sequence learning tasks: speech recognition and scene text recognition. From the experimental results, we empirically show that the proposed model improves the students across these tasks while achieving a considerable speed-up in the teacher model's training time.

arXiv:2111.03505

Visualizing the Emergence of Intermediate Visual Patterns in DNNs

This paper proposes a method to visualize the discrimination power of intermediate-layer visual patterns encoded by a DNN. Specifically, we visualize (1) how the DNN gradually learns regional visual patterns in each intermediate layer during the training process, and (2) the effects of the DNN using non-discriminative patterns in low layers to construct disciminative patterns in middle/high layers through the forward propagation. Based on our visualization method, we can quantify knowledge points (i.e., the number of discriminative visual patterns) learned by the DNN to evaluate the representation capacity of the DNN. Furthermore, this method also provides new insights into signal-processing behaviors of existing deep-learning techniques, such as adversarial attacks and knowledge distillation.

arXiv:2111.03362

A methodology for training homomorphicencryption friendly neural networks

Privacy-preserving deep neural network (DNN) inference is a necessity in different regulated industries such as healthcare, finance and retail. Recently, homomorphic encryption (HE) has been used as a method to enable analytics while addressing privacy concerns. HE enables secure predictions over encrypted data. However, there are several challenges related to the use of HE, including DNN size limitations and the lack of support for some operation types. Most notably, the commonly used ReLU activation is not supported under some HE schemes. We propose a structured methodology to replace ReLU with a quadratic polynomial activation. To address the accuracy degradation issue, we use a pre-trained model that trains another HE-friendly model, using techniques such as trainable activation functions and knowledge distillation. We demonstrate our methodology on the AlexNet architecture, using the chest X-Ray and CT datasets for COVID-19 detection. Experiments using our approach reduced the gap between the F1 score and accuracy of the models trained with ReLU and the HE-friendly model to within a mere 0.32-5.3 percent degradation. We also demonstrate our methodology using the SqueezeNet architecture, for which we observed 7 percent accuracy and F1 improvements over training similar networks with other HE-friendly training methods.

arXiv:2111.03555

AUTOKD: Automatic Knowledge Distillation Into A Student Architecture Family

State-of-the-art results in deep learning have been improving steadily, in good part due to the use of larger models. However, widespread use is constrained by device hardware limitations, resulting in a substantial performance gap between state-of-the-art models and those that can be effectively deployed on small devices. While Knowledge Distillation (KD) theoretically enables small student models to emulate larger teacher models, in practice selecting a good student architecture requires considerable human expertise. Neural Architecture Search (NAS) appears as a natural solution to this problem but most approaches can be inefficient, as most of the computation is spent comparing architectures sampled from the same distribution, with negligible differences in performance. In this paper, we propose to instead search for a family of student architectures sharing the property of being good at learning from a given teacher. Our approach AutoKD, powered by Bayesian Optimization, explores a flexible graph-based search space, enabling us to automatically learn the optimal student architecture distribution and KD parameters, while being 20x more sample efficient compared to existing state-of-the-art. We evaluate our method on 3 datasets; on large images specifically, we reach the teacher performance while using 3x less memory and 10x less parameters. Finally, while AutoKD uses the traditional KD loss, it outperforms more advanced KD variants using hand-designed students.

arXiv:2111.03341

DVFL: A Vertical Federated Learning Method for Dynamic Data

Federated learning, which solves the problem of data island by connecting multiple computational devices into a decentralized system, has become a promising paradigm for privacy-preserving machine learning. This paper studies vertical federated learning (VFL), which tackles the scenarios where collaborating organizations share the same set of users but disjoint features. Contemporary VFL methods are mainly used in static scenarios where the active party and the passive party have all the data from the beginning and will not change. However, the data in real life often changes dynamically. To alleviate this problem, we propose a new vertical federation learning method, DVFL, which adapts to dynamic data distribution changes through knowledge distillation. In DVFL, most of the computations are held locally to improve data security and model efficiency. Our extensive experimental results show that DVFL can not only obtain results close to existing VFL methods in static scenes, but also adapt to changes in data distribution in dynamic scenarios.

arXiv:2111.01992

Leveraging Advantages of Interactive and Non-Interactive Models for Vector-Based Cross-Lingual Information Retrieval

Interactive and non-interactive model are the two de-facto standard frameworks in vector-based cross-lingual information retrieval (V-CLIR), which embed queries and documents in synchronous and asynchronous fashions, respectively. From the retrieval accuracy and computational efficiency perspectives, each model has its own superiority and shortcoming. In this paper, we propose a novel framework to leverage the advantages of these two paradigms. Concretely, we introduce semi-interactive mechanism, which builds our model upon non-interactive architecture but encodes each document together with its associated multilingual queries. Accordingly, cross-lingual features can be better learned like an interactive model. Besides, we further transfer knowledge from a well-trained interactive model to ours by reusing its word embeddings and adopting knowledge distillation. Our model is initialized from a multilingual pre-trained language model M-BERT, and evaluated on two open-resource CLIR datasets derived from Wikipedia and an in-house dataset collected from a real-world search engine. Extensive analyses reveal that our methods significantly boost the retrieval accuracy while maintaining the computational efficiency.

arXiv:2111.02331

LTD: Low Temperature Distillation for Robust Adversarial Training

Adversarial training has been widely used to enhance the robustness of neural network models against adversarial attacks. Despite the popularity of neural network models, a significant gap exists between the natural and robust accuracy of these models. In this paper, we identify one of the primary reasons for this gap is the common use of one-hot vectors as labels, which hinders the learning process for image recognition. Representing ambiguous images with one-hot vectors is imprecise and may lead the model to suboptimal solutions. To overcome this issue, we propose a novel method called Low Temperature Distillation (LTD) that generates soft labels using the modified knowledge distillation framework. Unlike previous approaches, LTD uses a relatively low temperature in the teacher model and fixed, but different temperatures for the teacher and student models. This modification boosts the model's robustness without encountering the gradient masking problem that has been addressed in defensive distillation. The experimental results demonstrate the effectiveness of the proposed LTD method combined with previous techniques, achieving robust accuracy rates of 58.19%, 31.13%, and 42.08% on CIFAR-10, CIFAR-100, and ImageNet data sets, respectively, without additional unlabeled data.

arXiv:2111.01363

Knowledge Cross-Distillation for Membership Privacy

A membership inference attack (MIA) poses privacy risks for the training data of a machine learning model. With an MIA, an attacker guesses if the target data are a member of the training dataset. The state-of-the-art defense against MIAs, distillation for membership privacy (DMP), requires not only private data for protection but a large amount of unlabeled public data. However, in certain privacy-sensitive domains, such as medicine and finance, the availability of public data is not guaranteed. Moreover, a trivial method for generating public data by using generative adversarial networks significantly decreases the model accuracy, as reported by the authors of DMP. To overcome this problem, we propose a novel defense against MIAs that uses knowledge distillation without requiring public data. Our experiments show that the privacy protection and accuracy of our defense are comparable to those of DMP for the benchmark tabular datasets used in MIA research, Purchase100 and Texas100, and our defense has a much better privacy-utility trade-off than those of the existing defenses that also do not use public data for the image dataset CIFAR10.

arXiv:2111.01684

Rethinking the Knowledge Distillation From the Perspective of Model Calibration

Recent years have witnessed dramatically improvements in the knowledge distillation, which can generate a compact student model for better efficiency while retaining the model effectiveness of the teacher model. Previous studies find that: more accurate teachers do not necessary make for better teachers due to the mismatch of abilities. In this paper, we aim to analysis the phenomenon from the perspective of model calibration. We found that the larger teacher model may be too over-confident, thus the student model cannot effectively imitate. While, after the simple model calibration of the teacher model, the size of the teacher model has a positive correlation with the performance of the student model.

arXiv:2111.00775공식 코드 · Papers with Code

PP-ShiTu: A Practical Lightweight Image Recognition System

In recent years, image recognition applications have developed rapidly. A large number of studies and techniques have emerged in different fields, such as face recognition, pedestrian and vehicle re-identification, landmark retrieval, and product recognition. In this paper, we propose a practical lightweight image recognition system, named PP-ShiTu, consisting of the following 3 modules, mainbody detection, feature extraction and vector search. We introduce popular strategies including metric learning, deep hash, knowledge distillation and model quantization to improve accuracy and inference speed. With strategies above, PP-ShiTu works well in different scenarios with a set of models trained on a mixed dataset. Experiments on different datasets and benchmarks show that the system is widely effective in different domains of image recognition. All the above mentioned models are open-sourced and the code is available in the GitHub repository PaddleClas on PaddlePaddle.

arXiv:2111.00674공식 코드 · Papers with Code

Distilling Object Detectors with Feature Richness

In recent years, large-scale deep models have achieved great success, but the huge computational complexity and massive storage requirements make it a great challenge to deploy them in resource-limited devices. As a model compression and acceleration method, knowledge distillation effectively improves the performance of small models by transferring the dark knowledge from the teacher detector. However, most of the existing distillation-based detection methods mainly imitating features near bounding boxes, which suffer from two limitations. First, they ignore the beneficial features outside the bounding boxes. Second, these methods imitate some features which are mistakenly regarded as the background by the teacher detector. To address the above issues, we propose a novel Feature-Richness Score (FRS) method to choose important features that improve generalized detectability during distilling. The proposed method effectively retrieves the important features outside the bounding boxes and removes the detrimental features within the bounding boxes. Extensive experiments show that our methods achieve excellent performance on both anchor-based and anchor-free detectors. For example, RetinaNet with ResNet-50 achieves 39.7% in mAP on the COCO2017 dataset, which even surpasses the ResNet-101 based teacher detector 38.9% by 0.8%. Our implementation is available at https://github.com/duzhixing/FRS.

arXiv:2111.00643공식 코드 · Papers with Code

Learning Distilled Collaboration Graph for Multi-Agent Perception

To promote better performance-bandwidth trade-off for multi-agent perception, we propose a novel distilled collaboration graph (DiscoGraph) to model trainable, pose-aware, and adaptive collaboration among agents. Our key novelties lie in two aspects. First, we propose a teacher-student framework to train DiscoGraph via knowledge distillation. The teacher model employs an early collaboration with holistic-view inputs; the student model is based on intermediate collaboration with single-view inputs. Our framework trains DiscoGraph by constraining post-collaboration feature maps in the student model to match the correspondences in the teacher model. Second, we propose a matrix-valued edge weight in DiscoGraph. In such a matrix, each element reflects the inter-agent attention at a specific spatial region, allowing an agent to adaptively highlight the informative regions. During inference, we only need to use the student model named as the distilled collaboration network (DiscoNet). Attributed to the teacher-student framework, multiple agents with the shared DiscoNet could collaboratively approach the performance of a hypothetical teacher model with a holistic view. Our approach is validated on V2X-Sim 1.0, a large-scale multi-agent perception dataset that we synthesized using CARLA and SUMO co-simulation. Our quantitative and qualitative experiments in multi-agent 3D object detection show that DiscoNet could not only achieve a better performance-bandwidth trade-off than the state-of-the-art collaborative perception methods, but also bring more straightforward design rationale. Our code is available on https://github.com/ai4ce/DiscoNet.

arXiv:2110.15946

Estimating and Maximizing Mutual Information for Knowledge Distillation

In this work, we propose Mutual Information Maximization Knowledge Distillation (MIMKD). Our method uses a contrastive objective to simultaneously estimate and maximize a lower bound on the mutual information of local and global feature representations between a teacher and a student network. We demonstrate through extensive experiments that this can be used to improve the performance of low capacity models by transferring knowledge from more performant but computationally expensive models. This can be used to produce better models that can be run on devices with low computational resources. Our method is flexible, we can distill knowledge from teachers with arbitrary network architectures to arbitrary student networks. Our empirical results show that MIMKD outperforms competing approaches across a wide range of student-teacher pairs with different capacities, with different architectures, and when student networks are with extremely low capacity. We are able to obtain 74.55% accuracy on CIFAR100 with a ShufflenetV2 from a baseline accuracy of 69.8% by distilling knowledge from ResNet-50. On Imagenet we improve a ResNet-18 network from 68.88% to 70.32% accuracy (1.44%+) using a ResNet-34 teacher network.

arXiv:2110.15766

NxMTransformer: Semi-Structured Sparsification for Natural Language Understanding via ADMM

Natural Language Processing (NLP) has recently achieved success by using huge pre-trained Transformer networks. However, these models often contain hundreds of millions or even billions of parameters, bringing challenges to online deployment due to latency constraints. Recently, hardware manufacturers have introduced dedicated hardware for NxM sparsity to provide the flexibility of unstructured pruning with the runtime efficiency of structured approaches. NxM sparsity permits arbitrarily selecting M parameters to retain from a contiguous group of N in the dense representation. However, due to the extremely high complexity of pre-trained models, the standard sparse fine-tuning techniques often fail to generalize well on downstream tasks, which have limited data resources. To address such an issue in a principled manner, we introduce a new learning framework, called NxMTransformer, to induce NxM semi-structured sparsity on pretrained language models for natural language understanding to obtain better performance. In particular, we propose to formulate the NxM sparsity as a constrained optimization problem and use Alternating Direction Method of Multipliers (ADMM) to optimize the downstream tasks while taking the underlying hardware constraints into consideration. ADMM decomposes the NxM sparsification problem into two sub-problems that can be solved sequentially, generating sparsified Transformer networks that achieve high accuracy while being able to effectively execute on newly released hardware. We apply our approach to a wide range of NLP tasks, and our proposed method is able to achieve 1.7 points higher accuracy in GLUE score than current practices. Moreover, we perform detailed analysis on our approach and shed light on how ADMM affects fine-tuning accuracy for downstream tasks. Finally, we illustrate how NxMTransformer achieves performance improvement with knowledge distillation.

arXiv:2110.15538

On Cross-Layer Alignment for Model Fusion of Heterogeneous Neural Networks

Layer-wise model fusion via optimal transport, named OTFusion, applies soft neuron association for unifying different pre-trained networks to save computational resources. While enjoying its success, OTFusion requires the input networks to have the same number of layers. To address this issue, we propose a novel model fusion framework, named CLAFusion, to fuse neural networks with a different number of layers, which we refer to as heterogeneous neural networks, via cross-layer alignment. The cross-layer alignment problem, which is an unbalanced assignment problem, can be solved efficiently using dynamic programming. Based on the cross-layer alignment, our framework balances the number of layers of neural networks before applying layer-wise model fusion. Our experiments indicate that CLAFusion, with an extra finetuning process, improves the accuracy of residual networks on the CIFAR10, CIFAR100, and Tiny-ImageNet datasets. Furthermore, we explore its practical usage for model compression and knowledge distillation when applying to the teacher-student setting.

arXiv:2110.15210

Towards Model Agnostic Federated Learning Using Knowledge Distillation

Is it possible to design an universal API for federated learning using which an ad-hoc group of data-holders (agents) collaborate with each other and perform federated learning? Such an API would necessarily need to be model-agnostic i.e. make no assumption about the model architecture being used by the agents, and also cannot rely on having representative public data at hand. Knowledge distillation (KD) is the obvious tool of choice to design such protocols. However, surprisingly, we show that most natural KD-based federated learning protocols have poor performance. To investigate this, we propose a new theoretical framework, Federated Kernel ridge regression, which can capture both model heterogeneity as well as data heterogeneity. Our analysis shows that the degradation is largely due to a fundamental limitation of knowledge distillation under data heterogeneity. We further validate our framework by analyzing and designing new protocols based on KD. Their performance on real world experiments using neural networks, though still unsatisfactory, closely matches our theoretical predictions.

arXiv:2110.15094공식 코드 · Papers with Code

Mosaicking to Distill: Knowledge Distillation from Out-of-Domain Data

Knowledge distillation~(KD) aims to craft a compact student model that imitates the behavior of a pre-trained teacher in a target domain. Prior KD approaches, despite their gratifying results, have largely relied on the premise that \emph{in-domain} data is available to carry out the knowledge transfer. Such an assumption, unfortunately, in many cases violates the practical setting, since the original training data or even the data domain is often unreachable due to privacy or copyright reasons. In this paper, we attempt to tackle an ambitious task, termed as \emph{out-of-domain} knowledge distillation~(OOD-KD), which allows us to conduct KD using only OOD data that can be readily obtained at a very low cost. Admittedly, OOD-KD is by nature a highly challenging task due to the agnostic domain gap. To this end, we introduce a handy yet surprisingly efficacious approach, dubbed as~\textit{MosaicKD}. The key insight behind MosaicKD lies in that, samples from various domains share common local patterns, even though their global semantic may vary significantly; these shared local patterns, in turn, can be re-assembled analogous to mosaic tiling, to approximate the in-domain data and to further alleviating the domain discrepancy. In MosaicKD, this is achieved through a four-player min-max game, in which a generator, a discriminator, a student network, are collectively trained in an adversarial manner, partially under the guidance of a pre-trained teacher. We validate MosaicKD over {classification and semantic segmentation tasks} across various benchmarks, and demonstrate that it yields results much superior to the state-of-the-art counterparts on OOD data. Our code is available at \url{https://github.com/zju-vipa/MosaicKD}.

arXiv:2110.14215

Beyond Classification: Knowledge Distillation using Multi-Object Impressions

Knowledge Distillation (KD) utilizes training data as a transfer set to transfer knowledge from a complex network (Teacher) to a smaller network (Student). Several works have recently identified many scenarios where the training data may not be available due to data privacy or sensitivity concerns and have proposed solutions under this restrictive constraint for the classification task. Unlike existing works, we, for the first time, solve a much more challenging problem, i.e., "KD for object detection with zero knowledge about the training data and its statistics". Our proposed approach prepares pseudo-targets and synthesizes corresponding samples (termed as "Multi-Object Impressions"), using only the pretrained Faster RCNN Teacher network. We use this pseudo-dataset as a transfer set to conduct zero-shot KD for object detection. We demonstrate the efficacy of our proposed method through several ablations and extensive experiments on benchmark datasets like KITTI, Pascal and COCO. Our approach with no training samples, achieves a respectable mAP of 64.2% and 55.5% on the student with same and half capacity while performing distillation from a Resnet-18 Teacher of 73.3% mAP on KITTI.

arXiv:2110.14553공식 코드 · Papers with Code

GenURL: A General Framework for Unsupervised Representation Learning

Unsupervised representation learning (URL), which learns compact embeddings of high-dimensional data without supervision, has made remarkable progress recently. However, the development of URLs for different requirements is independent, which limits the generalization of the algorithms, especially prohibitive as the number of tasks grows. For example, dimension reduction methods, t-SNE, and UMAP optimize pair-wise data relationships by preserving the global geometric structure, while self-supervised learning, SimCLR, and BYOL focus on mining the local statistics of instances under specific augmentations. To address this dilemma, we summarize and propose a unified similarity-based URL framework, GenURL, which can smoothly adapt to various URL tasks. In this paper, we regard URL tasks as different implicit constraints on the data geometric structure that help to seek optimal low-dimensional representations that boil down to data structural modeling (DSM) and low-dimensional transformation (LDT). Specifically, DMS provides a structure-based submodule to describe the global structures, and LDT learns compact low-dimensional embeddings with given pretext tasks. Moreover, an objective function, General Kullback-Leibler divergence (GKL), is proposed to connect DMS and LDT naturally. Comprehensive experiments demonstrate that GenURL achieves consistent state-of-the-art performance in self-supervised visual learning, unsupervised knowledge distillation (KD), graph embeddings (GE), and dimension reduction.

arXiv:2110.14131

Temporal Knowledge Distillation for On-device Audio Classification

Improving the performance of on-device audio classification models remains a challenge given the computational limits of the mobile environment. Many studies leverage knowledge distillation to boost predictive performance by transferring the knowledge from large models to on-device models. However, most lack a mechanism to distill the essence of the temporal information, which is crucial to audio classification tasks, or similar architecture is often required. In this paper, we propose a new knowledge distillation method designed to incorporate the temporal knowledge embedded in attention weights of large transformer-based models into on-device models. Our distillation method is applicable to various types of architectures, including the non-attention-based architectures such as CNNs or RNNs, while retaining the original network architecture during inference. Through extensive experiments on both an audio event detection dataset and a noisy keyword spotting dataset, we show that our proposed method improves the predictive performance across diverse on-device architectures.