Structured belief states are crucial for user goal tracking and database query in task-oriented dialog systems. However, training belief trackers often requires expensive turn-level annotations of every user utterance. In this paper we aim at alleviating the reliance on belief state labels in building end-to-end dialog systems, by leveraging unlabeled dialog data towards semi-supervised learning. We propose a probabilistic dialog model, called the LAtent BElief State (LABES) model, where belief states are represented as discrete latent variables and jointly modeled with system responses given user inputs. Such latent variable modeling enables us to develop semi-supervised learning under the principled variational learning framework. Furthermore, we introduce LABES-S2S, which is a copyaugmented Seq2Seq model instantiation of LABES 1 . In supervised experiments, LABES-S2S obtains strong results on three benchmark datasets of different scales. In utilizing unlabeled dialog data, semi-supervised LABES-S2S significantly outperforms both supervisedonly and semi-supervised baselines. Remarkably, we can reduce the annotation demands to 50% without performance loss on MultiWOZ.
In agriculture, pest always causes the major damage in fields and results in significant crop yield losses. Currently, manual pest classification and counting are very time-consuming and many subjective factors can affect the population counting accuracy. In addition, the existing pest localization and recognition methods based on Convolutional Neural Network (CNN) are not satisfactory for practical pest prevention in fields because of pests' different scales and attitudes. In order to address these problems, an effective data augmentation strategy for CNN-based method is proposed in this paper. In training phase, we adopt data augmentation through rotating images by various degrees followed by cropping into different grids. In this way, we could obtain a large number of extra multi-scale examples that could be adopted to train a multi-scale pest detection model. In terms of test phase, we utilize the test time augmentation (TTA) strategy that separately inferences input images with various resolutions using the trained multi-scale model. Finally, we fuse these detection results from different image scales by non-maximum suppression (NMS) for the final result. Experimental results on wheat sawfly, wheat aphid, wheat mite and rice planthopper in our domain specific dataset, show that our proposed data augmentation strategy achieves the pest detection performance of 81.4% mean Average Precision (mAP), which improves 11.63%, 7.93% ,4.73% compared to three stateof-the-art approaches.
Hu Min)As the one of most powerful and natural signals of expressing emotion states [1], facial emotions account for the 55% role of emotional information [2]. Due to the influence of many factors, such as different subjects, races, illumination, complex background and so on, facial emotion analysis is a indubitable challenging task. Most of the previous researches [3,4,5] were based on data in laboratory-controlled environment which can avoid many factors mentioned above with the limitation of the number of datasets and algorithm. Compared to facial emotion recognition(FER) in laboratory-controlled environment, FER in-the-wild is closer to the business application. Nowadays, several works about FER in-the-wild has gradually
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.