PyTorchKR Papers

최신 논문

총 4,526편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:1910.12295공식 코드 · Papers with Code

MOD: A Deep Mixture Model with Online Knowledge Distillation for Large Scale Video Temporal Concept Localization

In this paper, we present and discuss a deep mixture model with online knowledge distillation (MOD) for large-scale video temporal concept localization, which is ranked 3rd in the 3rd YouTube-8M Video Understanding Challenge. Specifically, we find that by enabling knowledge sharing with online distillation, fintuning a mixture model on a smaller dataset can achieve better evaluation performance. Based on this observation, in our final solution, we trained and fintuned 12 NeXtVLAD models in parallel with a 2-layer online distillation structure. The experimental results show that the proposed distillation structure can effectively avoid overfitting and shows superior generalization performance. The code is publicly available at: https://github.com/linrongc/solution_youtube8m_v3

arXiv:1910.11450

An Empirical Study of Efficient ASR Rescoring with Transformers

Neural language models (LMs) have been proved to significantly outperform classical n-gram LMs for language modeling due to their superior abilities to model long-range dependencies in text and handle data sparsity problems. And recently, well configured deep Transformers have exhibited superior performance over shallow stack of recurrent neural network layers for language modeling. However, these state-of-the-art deep Transformer models were mostly engineered to be deep with high model capacity, which makes it computationally inefficient and challenging to be deployed into large-scale real-world applications. Therefore, it is important to develop Transformer LMs that have relatively small model sizes, while still retaining good performance of those much larger models. In this paper, we aim to conduct empirical study on training Transformers with small parameter sizes in the context of ASR rescoring. By combining techniques including subword units, adaptive softmax, large-scale model pre-training, and knowledge distillation, we show that we are able to successfully train small Transformer LMs with significant relative word error rate reductions (WERR) through n-best rescoring. In particular, our experiments on a video speech recognition dataset show that we are able to achieve WERRs ranging from 6.46% to 7.17% while only with 5.5% to 11.9% parameter sizes of the well-known large GPT model [1], whose WERR with rescoring on the same dataset is 7.58%.

arXiv:1910.11789

Secost: Sequential co-supervision for large scale weakly labeled audio event detection

Weakly supervised learning algorithms are critical for scaling audio event detection to several hundreds of sound categories. Such learning models should not only disambiguate sound events efficiently with minimal class-specific annotation but also be robust to label noise, which is more apparent with weak labels instead of strong annotations. In this work, we propose a new framework for designing learning models with weak supervision by bridging ideas from sequential learning and knowledge distillation. We refer to the proposed methodology as SeCoST (pronounced Sequest) -- Sequential Co-supervision for training generations of Students. SeCoST incrementally builds a cascade of student-teacher pairs via a novel knowledge transfer method. Our evaluations on Audioset (the largest weakly labeled dataset available) show that SeCoST achieves a mean average precision of 0.383 while outperforming prior state of the art by a considerable margin.

arXiv:1910.10986

Adversarial Feature Alignment: Avoid Catastrophic Forgetting in Incremental Task Lifelong Learning

Human beings are able to master a variety of knowledge and skills with ongoing learning. By contrast, dramatic performance degradation is observed when new tasks are added to an existing neural network model. This phenomenon, termed as \emph{Catastrophic Forgetting}, is one of the major roadblocks that prevent deep neural networks from achieving human-level artificial intelligence. Several research efforts, e.g. \emph{Lifelong} or \emph{Continual} learning algorithms, have been proposed to tackle this problem. However, they either suffer from an accumulating drop in performance as the task sequence grows longer, or require to store an excessive amount of model parameters for historical memory, or cannot obtain competitive performance on the new tasks. In this paper, we focus on the incremental multi-task image classification scenario. Inspired by the learning process of human students, where they usually decompose complex tasks into easier goals, we propose an adversarial feature alignment method to avoid catastrophic forgetting. In our design, both the low-level visual features and high-level semantic features serve as soft targets and guide the training process in multiple stages, which provide sufficient supervised information of the old tasks and help to reduce forgetting. Due to the knowledge distillation and regularization phenomenons, the proposed method gains even better performance than finetuning on the new tasks, which makes it stand out from other methods. Extensive experiments in several typical lifelong learning scenarios demonstrate that our method outperforms the state-of-the-art methods in both accuracies on new tasks and performance preservation on old tasks.

arXiv:1910.10699공식 코드 · Papers with Code

Contrastive Representation Distillation

Often we wish to transfer representational knowledge from one neural network to another. Examples include distilling a large network into a smaller one, transferring knowledge from one sensory modality to a second, or ensembling a collection of models into a single estimator. Knowledge distillation, the standard approach to these problems, minimizes the KL divergence between the probabilistic outputs of a teacher and student network. We demonstrate that this objective ignores important structural knowledge of the teacher network. This motivates an alternative objective by which we train a student to capture significantly more information in the teacher's representation of the data. We formulate this objective as contrastive learning. Experiments demonstrate that our resulting new objective outperforms knowledge distillation and other cutting-edge distillers on a variety of knowledge transfer tasks, including single model compression, ensemble distillation, and cross-modal transfer. Our method sets a new state-of-the-art in many transfer tasks, and sometimes even outperforms the teacher network when combined with knowledge distillation. Code: http://github.com/HobbitLong/RepDistiller.

arXiv:1910.08381

Model Compression with Two-stage Multi-teacher Knowledge Distillation for Web Question Answering System

Deep pre-training and fine-tuning models (such as BERT and OpenAI GPT) have demonstrated excellent results in question answering areas. However, due to the sheer amount of model parameters, the inference speed of these models is very slow. How to apply these complex models to real business scenarios becomes a challenging but practical problem. Previous model compression methods usually suffer from information loss during the model compression procedure, leading to inferior models compared with the original one. To tackle this challenge, we propose a Two-stage Multi-teacher Knowledge Distillation (TMKD for short) method for web Question Answering system. We first develop a general Q\&A distillation task for student model pre-training, and further fine-tune this pre-trained student model with multi-teacher knowledge distillation on downstream tasks (like Web Q\&A task, MNLI, SNLI, RTE tasks from GLUE), which effectively reduces the overfitting bias in individual teacher models, and transfers more general knowledge to the student model. The experiment results show that our method can significantly outperform the baseline methods and even achieve comparable results with the original teacher models, along with substantial speedup of model inference.

arXiv:1910.07331

A Generalized and Robust Method Towards Practical Gaze Estimation on Smart Phone

Gaze estimation for ordinary smart phone, e.g. estimating where the user is looking at on the phone screen, can be applied in various applications. However, the widely used appearance-based CNN methods still have two issues for practical adoption. First, due to the limited dataset, gaze estimation is very likely to suffer from over-fitting, leading to poor accuracy at run time. Second, the current methods are usually not robust, i.e. their prediction results having notable jitters even when the user is performing gaze fixation, which degrades user experience greatly. For the first issue, we propose a new tolerant and talented (TAT) training scheme, which is an iterative random knowledge distillation framework enhanced with cosine similarity pruning and aligned orthogonal initialization. The knowledge distillation is a tolerant teaching process providing diverse and informative supervision. The enhanced pruning and initialization is a talented learning process prompting the network to escape from the local minima and re-born from a better start. For the second issue, we define a new metric to measure the robustness of gaze estimator, and propose an adversarial training based Disturbance with Ordinal loss (DwO) method to improve it. The experimental results show that our TAT method achieves state-of-the-art performance on GazeCapture dataset, and that our DwO method improves the robustness while keeping comparable accuracy.

arXiv:1910.06720

Improving Word Embedding Factorization for Compression Using Distilled Nonlinear Neural Decomposition

Word-embeddings are vital components of Natural Language Processing (NLP) models and have been extensively explored. However, they consume a lot of memory which poses a challenge for edge deployment. Embedding matrices, typically, contain most of the parameters for language models and about a third for machine translation systems. In this paper, we propose Distilled Embedding, an (input/output) embedding compression method based on low-rank matrix decomposition and knowledge distillation. First, we initialize the weights of our decomposed matrices by learning to reconstruct the full pre-trained word-embedding and then fine-tune end-to-end, employing knowledge distillation on the factorized embedding. We conduct extensive experiments with various compression rates on machine translation and language modeling, using different data-sets with a shared word-embedding matrix for both embedding and vocabulary projection matrices. We show that the proposed technique is simple to replicate, with one fixed parameter controlling compression size, has higher BLEU score on translation and lower perplexity on language modeling compared to complex, difficult to tune state-of-the-art methods.

arXiv:1910.05057

Noise as a Resource for Learning in Knowledge Distillation

While noise is commonly considered a nuisance in computing systems, a number of studies in neuroscience have shown several benefits of noise in the nervous system from enabling the brain to carry out computations such as probabilistic inference as well as carrying additional information about the stimuli. Similarly, noise has been shown to improve the performance of deep neural networks. In this study, we further investigate the effect of adding noise in the knowledge distillation framework because of its resemblance to collaborative subnetworks in the brain regions. We empirically show that injecting constructive noise at different levels in the collaborative learning framework enables us to train the model effectively and distill desirable characteristics in the student model. In doing so, we propose three different methods that target the common challenges in deep neural networks: minimizing the performance gap between a compact model and large model (Fickle Teacher), training high performance compact adversarially robust models (Soft Randomization), and training models efficiently under label noise (Messy Collaboration). Our findings motivate further study in the role of noise as a resource for learning in a collaborative learning framework.

arXiv:1910.04985공식 코드 · Papers with Code

VarGFaceNet: An Efficient Variable Group Convolutional Neural Network for Lightweight Face Recognition

To improve the discriminative and generalization ability of lightweight network for face recognition, we propose an efficient variable group convolutional network called VarGFaceNet. Variable group convolution is introduced by VarGNet to solve the conflict between small computational cost and the unbalance of computational intensity inside a block. We employ variable group convolution to design our network which can support large scale face identification while reduce computational cost and parameters. Specifically, we use a head setting to reserve essential information at the start of the network and propose a particular embedding setting to reduce parameters of fully-connected layer for embedding. To enhance interpretation ability, we employ an equivalence of angular distillation loss to guide our lightweight network and we apply recursive knowledge distillation to relieve the discrepancy between the teacher model and the student model. The champion of deepglint-light track of LFR (2019) challenge demonstrates the effectiveness of our model and approach. Implementation of VarGFaceNet will be released at https://github.com/zma-c-137/VarGFaceNet soon.

arXiv:1910.04641

Cross-modal knowledge distillation for action recognition

In this work, we address the problem how a network for action recognition that has been trained on a modality like RGB videos can be adapted to recognize actions for another modality like sequences of 3D human poses. To this end, we extract the knowledge of the trained teacher network for the source modality and transfer it to a small ensemble of student networks for the target modality. For the cross-modal knowledge distillation, we do not require any annotated data. Instead we use pairs of sequences of both modalities as supervision, which are straightforward to acquire. In contrast to previous works for knowledge distillation that use a KL-loss, we show that the cross-entropy loss together with mutual learning of a small ensemble of student networks performs better. In fact, the proposed approach for cross-modal knowledge distillation nearly achieves the accuracy of a student network trained with full supervision.

arXiv:1910.03581

FedMD: Heterogenous Federated Learning via Model Distillation

Federated learning enables the creation of a powerful centralized model without compromising data privacy of multiple participants. While successful, it does not incorporate the case where each participant independently designs its own model. Due to intellectual property concerns and heterogeneous nature of tasks and data, this is a widespread requirement in applications of federated learning to areas such as health care and AI as a service. In this work, we use transfer learning and knowledge distillation to develop a universal framework that enables federated learning when each agent owns not only their private data, but also uniquely designed models. We test our framework on the MNIST/FEMNIST dataset and the CIFAR10/CIFAR100 dataset and observe fast improvement across all participating models. With 10 distinct participants, the final test accuracy of each model on average receives a 20% gain on top of what's possible without collaboration and is only a few percent lower than the performance each model would have obtained if all private datasets were pooled and made directly available for all participants.

arXiv:1910.03723

Knowledge Distillation from Internal Representations

Knowledge distillation is typically conducted by training a small model (the student) to mimic a large and cumbersome model (the teacher). The idea is to compress the knowledge from the teacher by using its output probabilities as soft-labels to optimize the student. However, when the teacher is considerably large, there is no guarantee that the internal knowledge of the teacher will be transferred into the student; even if the student closely matches the soft-labels, its internal representations may be considerably different. This internal mismatch can undermine the generalization capabilities originally intended to be transferred from the teacher to the student. In this paper, we propose to distill the internal representations of a large model such as BERT into a simplified version of it. We formulate two ways to distill such representations and various algorithms to conduct the distillation. We experiment with datasets from the GLUE benchmark and consistently show that adding knowledge distillation from internal representations is a more powerful method than only using soft-label distillation.

arXiv:1910.01769

Distilling BERT into Simple Neural Networks with Unlabeled Transfer Data

Recent advances in pre-training huge models on large amounts of text through self supervision have obtained state-of-the-art results in various natural language processing tasks. However, these huge and expensive models are difficult to use in practise for downstream tasks. Some recent efforts use knowledge distillation to compress these models. However, we see a gap between the performance of the smaller student models as compared to that of the large teacher. In this work, we leverage large amounts of in-domain unlabeled transfer data in addition to a limited amount of labeled training instances to bridge this gap for distilling BERT. We show that simple RNN based student models even with hard distillation can perform at par with the huge teachers given the transfer set. The student performance can be further improved with soft distillation and leveraging teacher intermediate representations. We show that our student models can compress the huge teacher by up to 26x while still matching or even marginally exceeding the teacher performance in low-resource settings with small amount of labeled data. Additionally, for the multilingual extension of this work with XtremeDistil (Mukherjee and Hassan Awadallah, 2020), we demonstrate massive distillation of multilingual BERT-like teacher models by upto 35x in terms of parameter compression and 51x in terms of latency speedup for batch inference while retaining 95% of its F1-score for NER over 41 languages.

arXiv:1910.01740

AntMan: Sparse Low-Rank Compression to Accelerate RNN inference

Wide adoption of complex RNN based models is hindered by their inference performance, cost and memory requirements. To address this issue, we develop AntMan, combining structured sparsity with low-rank decomposition synergistically, to reduce model computation, size and execution time of RNNs while attaining desired accuracy. AntMan extends knowledge distillation based training to learn the compressed models efficiently. Our evaluation shows that AntMan offers up to 100x computation reduction with less than 1pt accuracy drop for language and machine reading comprehension models. Our evaluation also shows that for a given accuracy target, AntMan produces 5x smaller models than the state-of-art. Lastly, we show that AntMan offers super-linear speed gains compared to theoretical speedup, demonstrating its practical value on commodity hardware.

arXiv:1910.01348

On the Efficacy of Knowledge Distillation

In this paper, we present a thorough evaluation of the efficacy of knowledge distillation and its dependence on student and teacher architectures. Starting with the observation that more accurate teachers often don't make good teachers, we attempt to tease apart the factors that affect knowledge distillation performance. We find crucially that larger models do not often make better teachers. We show that this is a consequence of mismatched capacity, and that small students are unable to mimic large teachers. We find typical ways of circumventing this (such as performing a sequence of knowledge distillation steps) to be ineffective. Finally, we show that this effect can be mitigated by stopping the teacher's training early. Our results generalize across datasets and models.

arXiv:1910.01108공식 코드 · Papers with Code

DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter

As Transfer Learning from large-scale pre-trained models becomes more prevalent in Natural Language Processing (NLP), operating these large models in on-the-edge and/or under constrained computational training or inference budgets remains challenging. In this work, we propose a method to pre-train a smaller general-purpose language representation model, called DistilBERT, which can then be fine-tuned with good performances on a wide range of tasks like its larger counterparts. While most prior work investigated the use of distillation for building task-specific models, we leverage knowledge distillation during the pre-training phase and show that it is possible to reduce the size of a BERT model by 40%, while retaining 97% of its language understanding capabilities and being 60% faster. To leverage the inductive biases learned by larger models during pre-training, we introduce a triple loss combining language modeling, distillation and cosine-distance losses. Our smaller, faster and lighter model is cheaper to pre-train and we demonstrate its capabilities for on-device computations in a proof-of-concept experiment and a comparative on-device study.

arXiv:1909.13063공식 코드 · Papers with Code

Training convolutional neural networks with cheap convolutions and online distillation

The large memory and computation consumption in convolutional neural networks (CNNs) has been one of the main barriers for deploying them on resource-limited systems. To this end, most cheap convolutions (e.g., group convolution, depth-wise convolution, and shift convolution) have recently been used for memory and computation reduction but with the specific architecture designing. Furthermore, it results in a low discriminability of the compressed networks by directly replacing the standard convolution with these cheap ones. In this paper, we propose to use knowledge distillation to improve the performance of the compact student networks with cheap convolutions. In our case, the teacher is a network with the standard convolution, while the student is a simple transformation of the teacher architecture without complicated redesigning. In particular, we propose a novel online distillation method, which online constructs the teacher network without pre-training and conducts mutual learning between the teacher and student network, to improve the performance of the student model. Extensive experiments demonstrate that the proposed approach achieves superior performance to simultaneously reduce memory and computation overhead of cutting-edge CNNs on different datasets, including CIFAR-10/100 and ImageNet ILSVRC 2012, compared to the state-of-the-art CNN compression and acceleration methods. The codes are publicly available at https://github.com/EthanZhangYC/OD-cheap-convolution.

arXiv:1909.11874공식 코드 · Papers with Code

Compact Trilinear Interaction for Visual Question Answering

In Visual Question Answering (VQA), answers have a great correlation with question meaning and visual contents. Thus, to selectively utilize image, question and answer information, we propose a novel trilinear interaction model which simultaneously learns high level associations between these three inputs. In addition, to overcome the interaction complexity, we introduce a multimodal tensor-based PARALIND decomposition which efficiently parameterizes trilinear interaction between the three inputs. Moreover, knowledge distillation is first time applied in Free-form Opened-ended VQA. It is not only for reducing the computational cost and required memory but also for transferring knowledge from trilinear interaction model to bilinear interaction model. The extensive experiments on benchmarking datasets TDIUC, VQA-2.0, and Visual7W show that the proposed compact trilinear interaction model achieves state-of-the-art results when using a single model on all three datasets.

arXiv:1909.11723공식 코드 · Papers with Code

Revisiting Knowledge Distillation via Label Smoothing Regularization

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 poorly-trained teacher with much lower accuracy than the student can still improve the latter significantly. To explain these observations, we provide a theoretical analysis of the relationships between KD and label smoothing regularization. We prove that 1) KD is a type of learned label smoothing regularization and 2) label smoothing regularization provides a virtual teacher model for KD. From these results, we argue that the success of KD is not fully due to the similarity information between categories from teachers, but also to the regularization of soft targets, which is equally or even more important. Based on these analyses, we further propose a novel Teacher-free Knowledge Distillation (Tf-KD) framework, where a student model learns from itself or manuallydesigned regularization distribution. The Tf-KD achieves comparable performance with normal KD from a superior teacher, which is well applied when a stronger teacher model is unavailable. Meanwhile, Tf-KD is generic and can be directly deployed for training deep neural networks. Without any extra computation cost, Tf-KD achieves up to 0.65\% improvement on ImageNet over well-established baseline models, which is superior to label smoothing regularization.