Recent breakthroughs in semi-supervised semantic segmentation have been developed through contrastive learning. In prevalent pixel-wise contrastive learning solutions, the model maps pixels to deterministic representations and regularizes them in the latent space. However, there exist inaccurate pseudo-labels which map the ambiguous representations of pixels to the wrong classes due to the limited cognitive ability of the model. In this paper, we define pixel-wise representations from a new perspective of probability theory and propose a Probabilistic Representation Contrastive Learning (PRCL) framework that improves representation quality by taking its probability into consideration. Through modelling the mapping from pixels to representations as the probability via multivariate Gaussian distributions, we can tune the contribution of the ambiguous representations to tolerate the risk of inaccurate pseudo-labels. Furthermore, we define prototypes in the form of distributions, which indicates the confidence of a class, while the point prototype cannot. More- over, we propose to regularize the distribution variance to enhance the reliability of representations. Taking advantage of these benefits, high-quality feature representations can be derived in the latent space, thereby the performance of se- mantic segmentation can be further improved. We conduct sufficient experiment to evaluate PRCL on Pascal VOC and CityScapes to demonstrate its superiority. The code is available at https://github.com/Haoyu-Xie/PRCL.

Semi-Supervised Semantic Segmentation (S4) aims to train a segmentation model with limited labeled images and a substantial volume of unlabeled images. To improve the robustness of representations, powerful methods introduce a pixel-wise contrastive learning approach in latent space (i.e., representation space) that aggregates the representations to their prototypes in a fully supervised manner. However, previous contrastive-based S4 methods merely rely on the supervision from the model’s output (logits) in logit space during unlabeled training. In contrast, we utilize the outputs in both logit space and representation space to obtain supervision in a collaborative way. The supervision from two spaces plays two roles: 1) reduces the risk of over-fitting to incorrect semantic information in logits with the help of representations; 2) enhances the knowledge exchange between the two spaces. Furthermore, unlike previous approaches, we use the similarity between representations and prototypes as a new indicator to tilt training those under-performing representations and achieve a more efficient contrastive learning process. Results on two public benchmarks demonstrate the competitive performance of our method compared with state-of-the-art methods.

Big data can help with providing valuable perceptions into business activities and disclosing the potential benefits. Advances in machine learning and deep learning technologies make it easier to achieve significant performance in a wide range of domains from city planning and marketing analysis to credit evaluation and identity theft detection. However, it still requires great efforts in selecting efficient learning algorithms and precise model parameters that are deemed confidential in the light of experience. Also worth noting is that there is a fundamental gap between impracticable business requirements and the available value of data reflected. The data holder or data service provider may not have a clear understanding of data interference. The solution to these two problems depends on the capability of predicting the data utility in advance, which raises a fundamental question: to what degree is the data utility predictable? In this work, we present a primary analytical framework for information-theoretic bounds of data utility and utilize the current state-of-the-art and representative algorithms to obtain the achievable lower bounds on a real-world data set. The gap between theoretical upper bounds and achievable lower bounds indicates that the achievable lower bounds can still be optimized for performance.

The rapid development of internet finance has caused increasing concern in online payment fraud due to its great threat. It is typical to employ rule systems or machine learning-based techniques to detect frauds. For the most significant features of such fraudulent transactions are exhibited in a sequential form, the sliding time window is a widely-recognized effective tool for this problem. With a sliding time window, features about the transaction characteristics can be extracted, and the latent patterns hidden in transaction records can be captured. However, the adaptive setting of sliding time window is really a big challenge, since the transaction patterns in real-life application scenarios are often too elusive to be captured. As a matter of fact, the practical setting usually needs to be updated and refined with manual intervention regularly. This is time-consuming indeed. In this article, we pursue an adaptive learning approach to detect fraudulent online payment transactions with automatic sliding time windows. Accordingly, we make efforts on optimizing the setting of windows and improving the adaptability. We design an intelligent window, called learning automatic window (LAW). It utilizes the learning automata to learn the proper parameters of time windows and adjust them dynamically and regularly according to the variation and oscillation of fraudulent transaction patterns. By the experiments over a real-world dataset of the online payment service from a commercial bank, we validate the gain of LAW in terms of detection effectiveness and robustness. To the best of our knowledge, this is the first work to make a sliding time window for fraud detection capable of learning its proper size in changing situations.