To train robust deep neural networks (DNNs), we systematically study several target modification approaches, which include output regularisation, self and non-self label correction (LC). Two key issues are discovered: (1) Self LC is the most appealing as it exploits its own knowledge and requires no extra models. However, how to automatically decide the trust degree of a learner as training goes is not well answered in the literature? (2) Some methods penalise while the others reward low-entropy predictions, prompting us to ask which one is better? To resolve the first issue, taking two well-accepted propositions--deep neural networks learn meaningful patterns before fitting noise [3] and minimum entropy regularisation principle [10]--we propose a novel end-to-end method named ProSelfLC, which is designed according to learning time and entropy. Specifically, given a data point, we progressively increase trust in its predicted label distribution versus its annotated one if a model has been trained for enough time and the prediction is of low entropy (high confidence). For the second issue, according to ProSelfLC, we empirically prove that it is better to redefine a meaningful low-entropy status and optimise the learner toward it. This serves as a defence of entropy minimisation. We demonstrate the effectiveness of ProSelfLC through extensive experiments in both clean and noisy settings. The source code is available at https://github.com/XinshaoAmosWang/ProSelfLC-CVPR2021. Keywords: entropy minimisation, maximum entropy, confidence penalty, self knowledge distillation, label correction, label noise, semi-supervised learning, output regularisation
Deep neural networks with millions of parameters may suffer from poor generalization due to overfitting. To mitigate the issue, we propose a new regularization method that penalizes the predictive distribution between similar samples. In particular, we distill the predictive distribution between different samples of the same label during training. This results in regularizing the dark knowledge (i.e., the knowledge on wrong predictions) of a single network (i.e., a self-knowledge distillation) by forcing it to produce more meaningful and consistent predictions in a class-wise manner. Consequently, it mitigates overconfident predictions and reduces intra-class variations. Our experimental results on various image classification tasks demonstrate that the simple yet powerful method can significantly improve not only the generalization ability but also the calibration performance of modern convolutional neural networks.
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.
Since deep learning became a key player in natural language processing (NLP), many deep learning models have been showing remarkable performances in a variety of NLP tasks, and in some cases, they are even outperforming humans. Such high performance can be explained by efficient knowledge representation of deep learning models. While many methods have been proposed to learn more efficient representation, knowledge distillation from pretrained deep networks suggest that we can use more information from the soft target probability to train other neural networks. In this paper, we propose a new knowledge distillation method self-knowledge distillation, based on the soft target probabilities of the training model itself, where multimode information is distilled from the word embedding space right below the softmax layer. Due to the time complexity, our method approximates the soft target probabilities. In experiments, we applied the proposed method to two different and fundamental NLP tasks: language model and neural machine translation. The experiment results show that our proposed method improves performance on the tasks.
분포 외(OOD) 탐지에서 보조 이상치(outlier) 데이터를 활용한 미세조정(fine-tuning)은 종종 분류(classification) 정확도를 희생시키면서 탐지 성능을 향상시킵니다. 이러한 절충(trade-off)은 미세조정 과정에서 기존의 분포 내(ID) 분포가 손실(loss)되는 데서 기인합니다. 보다 실용적이고 효과적인 패러다임을 구축하기 위해, 본 연구에서는 모델(model) 리마인더, 데이터 샘플링, 표현 학습이라는 세 가지 핵심 요인을 최적화합니다. 구체적으로 본 논문에서는 (1) 정확도 저하를 완화하는 자체 지식 증류(SKD), (2) 최소한의 데이터로 탐지 효율성을 높이는 준-어려운 이상치 샘플링(SOS), (3) ID와 OOD 간의 분리성을 촉진하는 이상치 인식 지도 대조 학습(OSCL)을 제안합니다. 이러한 요인들을 최적화함으로써 누적된 성능 향상을 달성하여 OOD 탐지 성능과 분류 정확도를 모두 끌어올립니다. 제안하는 프레임워크는 다양한 벤치마크 전반에서 기존 방법들을 능가하며, 특히 롱테일(long-tailed) 시나리오에서 우수한 성능을 보여 실제 환경의 OOD 탐지를 위한 견고한 기준선(baseline)을 제공합니다.
Micro-expressions are subtle facial movements that occur spontaneously when people try to conceal real emotions. Micro-expression recognition is crucial in many fields, including criminal analysis and psychotherapy. However, micro-expression recognition is challenging since micro-expressions have low intensity and public datasets are small in size. To this end, a three-stream temporal-shift attention network based on self-knowledge distillation is proposed in this paper. Firstly, to address the low intensity of muscle movements, we utilize learning-based motion magnification modules to enhance the intensity of muscle movements. Secondly, we employ efficient channel attention modules in the local-spatial stream to make the network focus on facial regions that are highly relevant to micro-expressions. In addition, temporal shift modules are used in the dynamic-temporal stream, which enables temporal modeling with no additional parameters by mixing motion information from two different temporal domains. Furthermore, we introduce self-knowledge distillation into the micro-expression recognition task by introducing auxiliary classifiers and using the deepest section of the network for supervision, encouraging all blocks to fully explore the features of the training set. Finally, extensive experiments are conducted on five publicly available micro-expression datasets. The experimental results demonstrate that our network outperforms other existing methods and achieves new state-of-the-art performance. Our code is available at https://github.com/GuanghaoZhu663/SKD-TSTSAN.