PyTorchKR Papers

최신 논문

총 4,526편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2104.09866공식 코드 · Papers with Code

Distill on the Go: Online knowledge distillation in self-supervised learning

Self-supervised learning solves pretext prediction tasks that do not require annotations to learn feature representations. For vision tasks, pretext tasks such as predicting rotation, solving jigsaw are solely created from the input data. Yet, predicting this known information helps in learning representations useful for downstream tasks. However, recent works have shown that wider and deeper models benefit more from self-supervised learning than smaller models. To address the issue of self-supervised pre-training of smaller models, we propose Distill-on-the-Go (DoGo), a self-supervised learning paradigm using single-stage online knowledge distillation to improve the representation quality of the smaller models. We employ deep mutual learning strategy in which two models collaboratively learn from each other to improve one another. Specifically, each model is trained using self-supervised learning along with distillation that aligns each model's softmax probabilities of similarity scores with that of the peer model. We conduct extensive experiments on multiple benchmark datasets, learning objectives, and architectures to demonstrate the potential of our proposed method. Our results show significant performance gain in the presence of noisy and limited labels and generalization to out-of-distribution data.

arXiv:2104.09191

Compact CNN Structure Learning by Knowledge Distillation

The concept of compressing deep Convolutional Neural Networks (CNNs) is essential to use limited computation, power, and memory resources on embedded devices. However, existing methods achieve this objective at the cost of a drop in inference accuracy in computer vision tasks. To address such a drawback, we propose a framework that leverages knowledge distillation along with customizable block-wise optimization to learn a lightweight CNN structure while preserving better control over the compression-performance tradeoff. Considering specific resource constraints, e.g., floating-point operations per inference (FLOPs) or model-parameters, our method results in a state of the art network compression while being capable of achieving better inference accuracy. In a comprehensive evaluation, we demonstrate that our method is effective, robust, and consistent with results over a variety of network architectures and datasets, at negligible training overhead. In particular, for the already compact network MobileNet_v2, our method offers up to 2x and 5.2x better model compression in terms of FLOPs and model-parameters, respectively, while getting 1.05% better model performance than the baseline network.

arXiv:2104.09044공식 코드 · Papers with Code

Distilling Knowledge via Knowledge Review

Knowledge distillation transfers knowledge from the teacher network to the student one, with the goal of greatly improving the performance of the student network. Previous methods mostly focus on proposing feature transformation and loss functions between the same level's features to improve the effectiveness. We differently study the factor of connection path cross levels between teacher and student networks, and reveal its great importance. For the first time in knowledge distillation, cross-stage connection paths are proposed. Our new review mechanism is effective and structurally simple. Our finally designed nested and compact framework requires negligible computation overhead, and outperforms other methods on a variety of tasks. We apply our method to classification, object detection, and instance segmentation tasks. All of them witness significant student network performance improvement. Code is available at https://github.com/Jia-Research-Lab/ReviewKD

arXiv:2104.08572

On Learning the Geodesic Path for Incremental Learning

Neural networks notoriously suffer from the problem of catastrophic forgetting, the phenomenon of forgetting the past knowledge when acquiring new knowledge. Overcoming catastrophic forgetting is of significant importance to emulate the process of "incremental learning", where the model is capable of learning from sequential experience in an efficient and robust way. State-of-the-art techniques for incremental learning make use of knowledge distillation towards preventing catastrophic forgetting. Therein, one updates the network while ensuring that the network's responses to previously seen concepts remain stable throughout updates. This in practice is done by minimizing the dissimilarity between current and previous responses of the network one way or another. Our work contributes a novel method to the arsenal of distillation techniques. In contrast to the previous state of the art, we propose to firstly construct low-dimensional manifolds for previous and current responses and minimize the dissimilarity between the responses along the geodesic connecting the manifolds. This induces a more formidable knowledge distillation with smooth properties which preserves the past knowledge more efficiently as observed by our comprehensive empirical study.

arXiv:2104.07905공식 코드 · Papers with Code

Ego-Exo: Transferring Visual Representations from Third-person to First-person Videos

We introduce an approach for pre-training egocentric video models using large-scale third-person video datasets. Learning from purely egocentric data is limited by low dataset scale and diversity, while using purely exocentric (third-person) data introduces a large domain mismatch. Our idea is to discover latent signals in third-person video that are predictive of key egocentric-specific properties. Incorporating these signals as knowledge distillation losses during pre-training results in models that benefit from both the scale and diversity of third-person video data, as well as representations that capture salient egocentric properties. Our experiments show that our Ego-Exo framework can be seamlessly integrated into standard video models; it outperforms all baselines when fine-tuned for egocentric activity recognition, achieving state-of-the-art results on Charades-Ego and EPIC-Kitchens-100.

논문 ID: 4486.34030

Disentangled Self-Supervision in Sequential Recommenders

To learn a sequential recommender, the existing methods typically adopt the sequence-to-item (seq2item) training strategy, which supervises a sequence model with a user’s next behavior as the label and the user’s past behaviors as the input. The seq2item strategy, however, is myopic and usually produces non-diverse recommendation lists. In this paper, we study the problem of mining extra signals for supervision by looking at the longer-term future. There exist two challenges: i) reconstructing a future sequence containing many behaviors is exponentially harder than reconstructing a single next behavior, which can lead to difficulty in convergence, and ii) the sequence of all future behaviors can involve many intentions, not all of which may be predictable from the sequence of earlier behaviors. To address these challenges, we propose a sequence-to-sequence (seq2seq) training strategy based on latent self-supervision and disentanglement. Specifically, we perform self-supervision in the latent space, i.e., reconstructing the representation of the future sequence as a whole, instead of reconstructing the items in the future sequence individually. We also disentangle the intentions behind any given sequence of behaviors and construct seq2seq training samples using only pairs of sub-sequences that involve a shared intention. Results on real-world benchmarks and synthetic data demonstrate the improvement brought by seq2seq training.

arXiv:2104.08154공식 코드 · Papers with Code

Counter-Interference Adapter for Multilingual Machine Translation

Developing a unified multilingual model has long been a pursuit for machine translation. However, existing approaches suffer from performance degradation -- a single multilingual model is inferior to separately trained bilingual ones on rich-resource languages. We conjecture that such a phenomenon is due to interference caused by joint training with multiple languages. To accommodate the issue, we propose CIAT, an adapted Transformer model with a small parameter overhead for multilingual machine translation. We evaluate CIAT on multiple benchmark datasets, including IWSLT, OPUS-100, and WMT. Experiments show that CIAT consistently outperforms strong multilingual baselines on 64 of total 66 language directions, 42 of which see above 0.5 BLEU improvement. Our code is available at \url{https://github.com/Yaoming95/CIAT}~.

arXiv:2104.07286

Continual Learning for Fake Audio Detection

Fake audio attack becomes a major threat to the speaker verification system. Although current detection approaches have achieved promising results on dataset-specific scenarios, they encounter difficulties on unseen spoofing data. Fine-tuning and retraining from scratch have been applied to incorporate new data. However, fine-tuning leads to performance degradation on previous data. Retraining takes a lot of time and computation resources. Besides, previous data are unavailable due to privacy in some situations. To solve the above problems, this paper proposes detecting fake without forgetting, a continual-learning-based method, to make the model learn new spoofing attacks incrementally. A knowledge distillation loss is introduced to loss function to preserve the memory of original model. Supposing the distribution of genuine voice is consistent among different scenarios, an extra embedding similarity loss is used as another constraint to further do a positive sample alignment. Experiments are conducted on the ASVspoof2019 dataset. The results show that our proposed method outperforms fine-tuning by the relative reduction of average equal error rate up to 81.62%.

arXiv:2104.07163

Annealing Knowledge Distillation

Significant memory and computational requirements of large deep neural networks restrict their application on edge devices. Knowledge distillation (KD) is a prominent model compression technique for deep neural networks in which the knowledge of a trained large teacher model is transferred to a smaller student model. The success of knowledge distillation is mainly attributed to its training objective function, which exploits the soft-target information (also known as "dark knowledge") besides the given regular hard labels in a training set. However, it is shown in the literature that the larger the gap between the teacher and the student networks, the more difficult is their training using knowledge distillation. To address this shortcoming, we propose an improved knowledge distillation method (called Annealing-KD) by feeding the rich information provided by the teacher's soft-targets incrementally and more efficiently. Our Annealing-KD technique is based on a gradual transition over annealed soft-targets generated by the teacher at different temperatures in an iterative process, and therefore, the student is trained to follow the annealed teacher output in a step-by-step manner. This paper includes theoretical and empirical evidence as well as practical experiments to support the effectiveness of our Annealing-KD method. We did a comprehensive set of experiments on different tasks such as image classification (CIFAR-10 and 100) and NLP language inference with BERT-based models on the GLUE benchmark and consistently got superior results.

arXiv:2104.07164

Unsupervised Continual Learning Via Pseudo Labels

Continual learning aims to learn new tasks incrementally using less computation and memory resources instead of retraining the model from scratch whenever new task arrives. However, existing approaches are designed in supervised fashion assuming all data from new tasks have been manually annotated, which are not practical for many real-life applications. In this work, we propose to use pseudo label instead of the ground truth to make continual learning feasible in unsupervised mode. The pseudo labels of new data are obtained by applying global clustering algorithm and we propose to use the model updated from last incremental step as the feature extractor. Due to the scarcity of existing work, we introduce a new benchmark experimental protocol for unsupervised continual learning of image classification task under class-incremental setting where no class label is provided for each incremental learning step. Our method is evaluated on the CIFAR-100 and ImageNet (ILSVRC) datasets by incorporating the pseudo label with various existing supervised approaches and show promising results in unsupervised scenario.

arXiv:2104.07253

Integration of Pre-trained Networks with Continuous Token Interface for End-to-End Spoken Language Understanding

Most End-to-End (E2E) SLU networks leverage the pre-trained ASR networks but still lack the capability to understand the semantics of utterances, crucial for the SLU task. To solve this, recently proposed studies use pre-trained NLU networks. However, it is not trivial to fully utilize both pre-trained networks; many solutions were proposed, such as Knowledge Distillation, cross-modal shared embedding, and network integration with Interface. We propose a simple and robust integration method for the E2E SLU network with novel Interface, Continuous Token Interface (CTI), the junctional representation of the ASR and NLU networks when both networks are pre-trained with the same vocabulary. Because the only difference is the noise level, we directly feed the ASR network's output to the NLU network. Thus, we can train our SLU network in an E2E manner without additional modules, such as Gumbel-Softmax. We evaluate our model using SLURP, a challenging SLU dataset and achieve state-of-the-art scores on both intent classification and slot filling tasks. We also verify the NLU network, pre-trained with Masked Language Model, can utilize a noisy textual representation of CTI. Moreover, we show our model can be trained with multi-task learning from heterogeneous data even after integration with CTI.

arXiv:2104.06457

Source and Target Bidirectional Knowledge Distillation for End-to-end Speech Translation

A conventional approach to improving the performance of end-to-end speech translation (E2E-ST) models is to leverage the source transcription via pre-training and joint training with automatic speech recognition (ASR) and neural machine translation (NMT) tasks. However, since the input modalities are different, it is difficult to leverage source language text successfully. In this work, we focus on sequence-level knowledge distillation (SeqKD) from external text-based NMT models. To leverage the full potential of the source language information, we propose backward SeqKD, SeqKD from a target-to-source backward NMT model. To this end, we train a bilingual E2E-ST model to predict paraphrased transcriptions as an auxiliary task with a single decoder. The paraphrases are generated from the translations in bitext via back-translation. We further propose bidirectional SeqKD in which SeqKD from both forward and backward NMT models is combined. Experimental evaluations on both autoregressive and non-autoregressive models show that SeqKD in each direction consistently improves the translation performance, and the effectiveness is complementary regardless of the model capacity.

arXiv:2104.06719

Sentence Embeddings by Ensemble Distillation

This paper contributes a new State Of The Art (SOTA) for Semantic Textual Similarity (STS). We compare and combine a number of recently proposed sentence embedding methods for STS, and propose a novel and simple ensemble knowledge distillation scheme that improves on previous approaches. Our experiments demonstrate that a model trained to learn the average embedding space from multiple ensemble students outperforms all the other individual models with high robustness. Utilizing our distillation method in combination with previous methods, we significantly improve on the SOTA unsupervised STS, and by proper hyperparameter tuning of previous methods we improve the supervised SOTA scores.

arXiv:2104.06476공식 코드 · Papers with Code

Incremental Multi-Target Domain Adaptation for Object Detection with Efficient Domain Transfer

Recent advances in unsupervised domain adaptation have significantly improved the recognition accuracy of CNNs by alleviating the domain shift between (labeled) source and (unlabeled) target data distributions. While the problem of single-target domain adaptation (STDA) for object detection has recently received much attention, multi-target domain adaptation (MTDA) remains largely unexplored, despite its practical relevance in several real-world applications, such as multi-camera video surveillance. Compared to the STDA problem that may involve large domain shifts between complex source and target distributions, MTDA faces additional challenges, most notably the computational requirements and catastrophic forgetting of previously-learned targets, which can depend on the order of target adaptations. STDA for detection can be applied to MTDA by adapting one model per target, or one common model with a mixture of data from target domains. However, these approaches are either costly or inaccurate. The only state-of-art MTDA method specialized for detection learns targets incrementally, one target at a time, and mitigates the loss of knowledge by using a duplicated detection model for knowledge distillation, which is computationally expensive and does not scale well to many domains. In this paper, we introduce an efficient approach for incremental learning that generalizes well to multiple target domains. Our MTDA approach is more suitable for real-world applications since it allows updating the detection model incrementally, without storing data from previous-learned target domains, nor retraining when a new target domain becomes available. Our proposed method, MTDA-DTM, achieved the highest level of detection accuracy compared against state-of-the-art approaches on several MTDA detection benchmarks and Wildtrack, a benchmark for multi-camera pedestrian detection.

arXiv:2104.06683공식 코드 · Papers with Code

The Curious Case of Hallucinations in Neural Machine Translation

In this work, we study hallucinations in Neural Machine Translation (NMT), which lie at an extreme end on the spectrum of NMT pathologies. Firstly, we connect the phenomenon of hallucinations under source perturbation to the Long-Tail theory of Feldman (2020), and present an empirically validated hypothesis that explains hallucinations under source perturbation. Secondly, we consider hallucinations under corpus-level noise (without any source perturbation) and demonstrate that two prominent types of natural hallucinations (detached and oscillatory outputs) could be generated and explained through specific corpus-level noise patterns. Finally, we elucidate the phenomenon of hallucination amplification in popular data-generation processes such as Backtranslation and sequence-level Knowledge Distillation.

arXiv:2104.06374공식 코드 · Papers with Code

Knowledge Distillation For Wireless Edge Learning

In this paper, we propose a framework for predicting frame errors in the collaborative spectrally congested wireless environments of the DARPA Spectrum Collaboration Challenge (SC2) via a recently collected dataset. We employ distributed deep edge learning that is shared among edge nodes and a central cloud. Using this close-to-practice dataset, we find that widely used federated learning approaches, specially those that are privacy preserving, are worse than local training for a wide range of settings. We hence utilize the synthetic minority oversampling technique to maintain privacy via avoiding the transfer of local data to the cloud, and utilize knowledge distillation with an aim to benefit from high cloud computing and storage capabilities. The proposed framework achieves overall better performance than both local and federated training approaches, while being robust against catastrophic failures as well as challenging channel conditions that result in high frame error rates.

arXiv:2104.05965

Dealing with Missing Modalities in the Visual Question Answer-Difference Prediction Task through Knowledge Distillation

In this work, we address the issues of missing modalities that have arisen from the Visual Question Answer-Difference prediction task and find a novel method to solve the task at hand. We address the missing modality-the ground truth answers-that are not present at test time and use a privileged knowledge distillation scheme to deal with the issue of the missing modality. In order to efficiently do so, we first introduce a model, the "Big" Teacher, that takes the image/question/answer triplet as its input and outperforms the baseline, then use a combination of models to distill knowledge to a target network (student) that only takes the image/question pair as its inputs. We experiment our models on the VizWiz and VQA-V2 Answer Difference datasets and show through extensive experimentation and ablation the performances of our method and a diverse possibility for future research.

arXiv:2104.05892공식 코드 · Papers with Code

CXR Segmentation by AdaIN-based Domain Adaptation and Knowledge Distillation

As segmentation labels are scarce, extensive researches have been conducted to train segmentation networks with domain adaptation, semi-supervised or self-supervised learning techniques to utilize abundant unlabeled dataset. However, these approaches appear different from each other, so it is not clear how these approaches can be combined for better performance. Inspired by recent multi-domain image translation approaches, here we propose a novel segmentation framework using adaptive instance normalization (AdaIN), so that a single generator is trained to perform both domain adaptation and semi-supervised segmentation tasks via knowledge distillation by simply changing task-specific AdaIN codes. Specifically, our framework is designed to deal with difficult situations in chest X-ray radiograph (CXR) segmentation, where labels are only available for normal data, but the trained model should be applied to both normal and abnormal data. The proposed network demonstrates great generalizability under domain shift and achieves the state-of-the-art performance for abnormal CXR segmentation.

arXiv:2104.04868

Data-Free Knowledge Distillation with Soft Targeted Transfer Set Synthesis

Knowledge distillation (KD) has proved to be an effective approach for deep neural network compression, which learns a compact network (student) by transferring the knowledge from a pre-trained, over-parameterized network (teacher). In traditional KD, the transferred knowledge is usually obtained by feeding training samples to the teacher network to obtain the class probabilities. However, the original training dataset is not always available due to storage costs or privacy issues. In this study, we propose a novel data-free KD approach by modeling the intermediate feature space of the teacher with a multivariate normal distribution and leveraging the soft targeted labels generated by the distribution to synthesize pseudo samples as the transfer set. Several student networks trained with these synthesized transfer sets present competitive performance compared to the networks trained with the original training set and other data-free KD approaches.

arXiv:2104.05382

Dual Discriminator Adversarial Distillation for Data-free Model Compression

Knowledge distillation has been widely used to produce portable and efficient neural networks which can be well applied on edge devices for computer vision tasks. However, almost all top-performing knowledge distillation methods need to access the original training data, which usually has a huge size and is often unavailable. To tackle this problem, we propose a novel data-free approach in this paper, named Dual Discriminator Adversarial Distillation (DDAD) to distill a neural network without any training data or meta-data. To be specific, we use a generator to create samples through dual discriminator adversarial distillation, which mimics the original training data. The generator not only uses the pre-trained teacher's intrinsic statistics in existing batch normalization layers but also obtains the maximum discrepancy from the student model. Then the generated samples are used to train the compact student network under the supervision of the teacher. The proposed method obtains an efficient student network which closely approximates its teacher network, despite using no original training data. Extensive experiments are conducted to to demonstrate the effectiveness of the proposed approach on CIFAR-10, CIFAR-100 and Caltech101 datasets for classification tasks. Moreover, we extend our method to semantic segmentation tasks on several public datasets such as CamVid and NYUv2. All experiments show that our method outperforms all baselines for data-free knowledge distillation.