Improvement of generalization capability of two-dimensional salt segmentation via iterative semi-supervised learning

Accurate interpretations of subsurface salts are vital to oil and gas exploration. Manually interpreting them from seismic depth images, however, is labor-intensive. Consequently, use of deep learning tools such as a convolutional neural network for automatic salt interpretation recently became popular. Because of poor generalization capabilities, interpreting salt boundaries using these tools is difficult when labeled data are available from one geological region and we like to make predictions for other nearby regions with varied geological features. At the same time, due to vast amount of the data involved and the associated computational complexities needed for training, such generalization is necessary for solving practical salt interpretation problems. In this work, we propose a semi-supervised training, which allows the predicted model to iteratively improve as more and more information is distilled from the unlabeled data into the model. In addition, by performing mixup between labeled and unlabeled data during training, we encourage the predicted models to linearly behave across training samples; thereby improving the generalization capability of the method. For each iteration, we use the model obtained from previous iteration to generate pseudo labels for the unlabeled data. This automated consecutive data distillation allows our model prediction to improve with iteration, without any need for human intervention. To demonstrate the effectiveness and efficiency, we apply the method on two-dimensional images extracted from a real three-dimensional seismic data volume. By comparing our predictions and fully supervised baseline predictions with those that were manually interpreted and we consider as “ground truth”, we find than the prediction quality our new method surpasses the baseline prediction. We therefore conclude that our new method is a viable tool for automated salt delineation from seismic depth images.

Unsupervised Domain Expansion for Visual Categorization

Expanding visual categorization into a novel domain without the need of extra annotation has been a long-term interest for multimedia intelligence. Previously, this challenge has been approached by unsupervised domain adaptation (UDA). Given labeled data from a source domain and unlabeled data from a target domain, UDA seeks for a deep representation that is both discriminative and domain-invariant. While UDA focuses on the target domain, we argue that the performance on both source and target domains matters, as in practice which domain a test example comes from is unknown. In this article, we extend UDA by proposing a new task called unsupervised domain expansion (UDE), which aims to adapt a deep model for the target domain with its unlabeled data, meanwhile maintaining the model’s performance on the source domain. We propose Knowledge Distillation Domain Expansion (KDDE) as a general method for the UDE task. Its domain-adaptation module can be instantiated with any existing model. We develop a knowledge distillation-based learning mechanism, enabling KDDE to optimize a single objective wherein the source and target domains are equally treated. Extensive experiments on two major benchmarks, i.e., Office-Home and DomainNet, show that KDDE compares favorably against four competitive baselines, i.e., DDC, DANN, DAAN, and CDAN, for both UDA and UDE tasks. Our study also reveals that the current UDA models improve their performance on the target domain at the cost of noticeable performance loss on the source domain.

Text Classification Model Enhanced by Unlabeled Data for LaTeX Formula

Generic language models pretrained on large unspecific domains are currently the foundation of NLP. Labeled data are limited in most model training due to the cost of manual annotation, especially in domains including massive Proper Nouns such as mathematics and biology, where it affects the accuracy and robustness of model prediction. However, directly applying a generic language model on a specific domain does not work well. This paper introduces a BERT-based text classification model enhanced by unlabeled data (UL-BERT) in the LaTeX formula domain. A two-stage Pretraining model based on BERT(TP-BERT) is pretrained by unlabeled data in the LaTeX formula domain. A double-prediction pseudo-labeling (DPP) method is introduced to obtain high confidence pseudo-labels for unlabeled data by self-training. Moreover, a multi-rounds teacher–student model training approach is proposed for UL-BERT model training with few labeled data and more unlabeled data with pseudo-labels. Experiments on the classification of the LaTex formula domain show that the classification accuracies have been significantly improved by UL-BERT where the F1 score has been mostly enhanced by 2.76%, and lower resources are needed in model training. It is concluded that our method may be applicable to other specific domains with enormous unlabeled data and limited labelled data.

Toward Automatically Labeling Situations in Soccer

We study the automatic annotation of situations in soccer games. At first sight, this translates nicely into a standard supervised learning problem. However, in a fully supervised setting, predictive accuracies are supposed to correlate positively with the amount of labeled situations: more labeled training data simply promise better performance. Unfortunately, non-trivially annotated situations in soccer games are scarce, expensive and almost always require human experts; a fully supervised approach appears infeasible. Hence, we split the problem into two parts and learn (i) a meaningful feature representation using variational autoencoders on unlabeled data at large scales and (ii) a large-margin classifier acting in this feature space but utilize only a few (manually) annotated examples of the situation of interest. We propose four different architectures of the variational autoencoder and empirically study the detection of corner kicks, crosses and counterattacks. We observe high predictive accuracies above 90% AUC irrespectively of the task.

NIMG-40. MRI-BASED ESTIMATION OF THE ABUNDANCE OF IMMUNOHISTOCHEMISTRY MARKERS IN GBM BRAIN USING DEEP LEARNING

Glioblastoma (GBM) is a devastating primary brain tumor known for its heterogeneity with a median survival of 15 months. Clinical imaging remains the primary modality to assess brain tumor response, but it is nearly impossible to distinguish between tumor growth and treatment response. Ki67 is a marker of active cell proliferation that shows inter- and intra-patient heterogeneity and should change under many therapies. In this work, we assessed the utility of a semi-supervised deep learning approach for regionally predicting high-vs-low Ki67 in GBM patients based on MRI. We used both labeled and unlabeled datasets to train the model. Labeled data included 114 MRI-localized biopsies from 43 unique GBM patients with available immunohistochemistry Ki67 labels. Unlabeled data included nine repeat routine pretreatment paired scans of newly-diagnosed GBM patients acquired within three days. Data augmentation techniques were utilized to enhance the size of our data and increase generalizability. Data was split between training, validation, and testing sets using 65-15-20 percent ratios. Model inputs were 16x16x3 patches around biopsies on T1Gd and T2 MRIs for labeled data, and around randomly selected patches inside the T2 abnormal region for unlabeled data. The network was a 4-conv layered VGG-inspired architecture. Training objective was accurate prediction of Ki67 in labeled patches and consistency in predictions across repeat unlabeled patches. We measured final model accuracy on held-out test samples. Our promising preliminary results suggest potential for deep learning in deconvolving the spatial heterogeneity of proliferative GBM subpopulations. If successful, this model can provide a non-invasive readout of cell proliferation and reveal the effectiveness of a given cytotoxic therapy dynamically during the patient's routine follow up. Further, the spatial resolution of our approach provides insights into the intra-tumoral heterogeneity of response which can be related to heterogeneity in localization of therapies (e.g. radiation therapy, drug dose heterogeneity).

Adversarial Semi-supervised Learning for Corporate Credit Ratings

Corporate credit rating is an analysis of credit risks withina corporation, which plays a vital role during the management of financial risk. Traditionally, the rating assessment process based on the historical profile of corporation is usually expensive and complicated, which often takes months. Therefore, most of the corporations, duetothelack in money and time, can’t get their own credit level. However, we believe that although these corporations haven’t their credit rating levels (unlabeled data), this big data contains useful knowledgeto improve credit system. In this work, its major challenge lies in how to effectively learn the knowledge from unlabeled data and help improve the performance of the credit rating system. Specifically, we consider the problem of adversarial semi-supervised learning (ASSL) for corporate credit rating which has been rarely researched before. A novel framework adversarial semi-supervised learning for corporate credit rating (ASSL4CCR) which includes two phases is proposed to address these problems. In the first phase, we train a normal rating system via a machine-learning algorithm to give unlabeled data pseudo rating level. Then in the second phase, adversarial semi-supervised learning is applied uniting labeled data and pseudo-labeleddatato build the final model. To demonstrate the effectiveness of the proposed ASSL4CCR, we conduct extensive experiments on the Chinese public-listed corporate rating dataset, which proves that ASSL4CCR outperforms the state-of-the-art methods consistently.