Weakly Supervised Semantic Segmentation (WSSS) has received increasing attentions recently, which leverages only limited supervision to train the segmentation networks. Meanwhile, Transformer utilizes multi-head attention mechanism to overcome the limited receptive fields of convolution neural networks (CNN). As a result, a large amount of WSSS work has been built upon Transformer architecture. However, most of them overlook over-smoothing issue (i.e. the patch tokens in deep layers tend to be similar, indicating that the learned information is nearly identical), which will affect the accuracy of Class Activation Map (CAM) in representing the pixel level semantic information. In this paper, an Attention-based Layer Fusion scheme with Token Masking is proposed to address the above issue. The core modules include a Weighted Layer Attention Aggregation (WLA) module and a Random-Masked Class Token Refinement (RMTR) module. The former integrates the attention information of different layers in an adaptive way, which refines CAM to obtain a refined object contour. The latter introduces random masks to aggregate higher-level semantic information. Extensive experiments have been conducted on public datasets, in which our network manifests superior performances than other state-of-the-arts methods.

With the unprecedented development of deep learning, autonomous vehicles (AVs) have achieved tremendous progress nowadays. However, AV supported by DNN models is vulnerable to data poisoning attacks, hindering the large-scale application of autonomous driving. For example, by injecting carefully designed poisons into the training dataset of the DNN model in the traffic sign recognition system, the attacker can mislead the system to make targeted misclassification or cause a reduction in model classification accuracy. In this article, we conduct a thorough investigation of the state-of-the-art data poisoning attacks and defenses against AVs. According to whether the attacker needs to manipulate the data labeling process, we divide the state-of-the-art attack approaches into two categories, i.e., dirty-label attacks and clean-label attacks. We also differentiate the existing defense methods into two categories based on whether to modify the training data or the models, i.e., data-based defenses and model-based defenses. In addition to a detailed review of attacks and defenses in each category, we also give a qualitative comparison of the existing attacks and defenses. Besides, we provide a quantitative comparison of the existing attack and defense methods through experiments. Last but not least, we pinpoint several future directions for data poisoning attacks and defenses in AVs, providing possible ways for further research.

The blockchain with a distributed network structure can provide a reliable and secure environment for Internet of Things (IoT) transactions, which also suffers from low throughput, high computation overhead, and large transaction fee. Payment channel networks (PCNs) are developed to address the scalability issue of blockchain. A key enabler of PCNs is the path-finding services. Most of existing routing algorithms target at finding a single feasible path, which may lead to failure of large payments. Moreover, previous solutions did not consider the transaction fee of the chosen path, which is extremely important to cost-sensitive users. In this work, we design a new multipath routing algorithm for PCNs that aims at minimizing transaction fees. Together considering the determination of the optimal number of paths, the optimal path routes and the optimal allocation leads to difficulty of the problem. In addition, the transaction fees along the path are closely related to the amount of payment, and the capacity of a payment channel limits the payment that can be carried. To address these challenges, we propose MILPA-PCN, a cost-effective multipath routing framework for PCNs. We develop a genetic algorithm-based routing determination algorithm with carefully designed genetic operations. We evaluate MILPA-PCN based on the real trace of the lightning network, and verify that MILPA-PCN can reduce the transaction fee by 33.56% and improve the payment success rate by 14.45%.

Leveraged by a large-scale diffusion of sensing networks and scanning devices in modern cities, huge volumes of geo-referenced urban data are collected every day. Such amount of information is analyzed to discover data-driven models, which can be exploited to tackle the major issues that cities face, including air pollution, virus diffusion, human mobility, traffic flows. In particular, the detection of city hotspots is becoming a valuable organization technique for framing detailed knowledge of a metropolitan area, providing high-level summaries for spatial datasets, which are valuable for planners, scientists, and policymakers. However, while classic density-based clustering algorithms show to be suitable to discover hotspots characterized by homogeneous density, their application on multi-density data can produce inaccurate results. For such a reason, since metropolitan cities are heavily characterized by variable densities, multi-density clustering seems to be more appropriate to discover city hotspots. This paper presents a study about how density-based clustering algorithms are suitable for discovering urban hotspots in a city, by showing a comparative analysis of single-density and multi-density clustering on both state-of-the-art data and real-world data. The experimental evaluation shows that, in an urban scenario, multi-density clustering achieves higher quality hotspots than a single-density approach.

Building efficient deep neural network models has become a hot-spot in recent years for deep learning research. Many works on network compression try to quantize a neural network with low bitwidth weights and activations. However, most of the existing network quantization methods set a fixed bitwidth for the whole network, which leads to large performance drop under high compression rate. In this paper we introduce an adaptive layerwise quantization method which quantizes the network with different bitwidth assigned to different layers. By using entropy of weights and activations as an importance indicator for each layer, we keep most of the layers under a high compression rate while a few most important layers receive more bit assignment. Experiments on CI-FAR10 and ImageNet2012 datasets demonstrate that our layerwise quantization could achieve smaller model size and less computation cost than the comparison fixed bitwidth methods with comparable accuracy, or higher accuracy with similar model size and computational complexity.

Modern deep learning models usually suffer high complexity in model size and computation when transplanted to resource constrained platforms. To this end, many works are dedicated to compressing deep neural networks. Adding group LASSO regularization is one of the most effective model compression methods since it generates structured sparse networks. We investigate the deep neural networks trained by group LASSO constraint and observe that even with strong sparsity regularization imposed, there still exists substantial filter correlation among the convolution filters, which is undesired for a compact neural network. We propose to suppress such correlation with a new kind of constraint called decorrelation regularization, which explicitly forces the network to learn a set of less correlated filters. The experiments on CIFAR10/100 and ILSVRC2012 datasets show that when combined our decorrelation regularization with group LASSO, the correlation between filters could be effectively weakened, which increases the sparsity of the resulting model and leads to better compressing performance.

Internet-based services monitor and detect anomalies on KPIs (Key Performance Indicators, say CPU utilization, number of queries per second, response latency) of their applications and systems in order to keep their services reliable. This paper identifies a common, important, yet little-studied problem of KPI anomaly detection: rapid deployment of anomaly detection models for large number of emerging KPI streams, without manual algorithm selection, parameter tuning, or new anomaly labeling for any newly emerging KPI streams. We propose the first framework ADS (Anomaly Detection through Self-training) that tackles the above problem, via clustering and semi-supervised learning. Our extensive experiments using real-world data show that, with the labels of only the 5 cluster centroids of 70 historical KPI streams, ADS achieves an averaged best F-score of 0.92 on 81 new KPI streams, almost the same as a state-of-art supervised approach, and greatly outperforming a state-of-art unsupervised approach by 61.40% on average.

With the globalization and integration of world financial markets, the success in information system has become a critical issue for financial investment institutes. Information system support has started to encompass the whole range of operational and decision-making activities in investment and financial industry. In this study, we take the challenge by integrating artificial intelligent techniques into the framework of the financial management information system. Specifically, we address the effectiveness of neural networks based trading strategy decision support system and discuss how it could be integrated into the information system to improve uncovering accurate trading signals and maximize trading profits. Our results shed light on the potential value of neural networks in financial management information system, by discovering patterns and trading signals in noisy and dynamic financial data and by integrating with other decision support systems in making a more optimized trading strategy.

This paper presents a study of using artificial neural networks in predicting stock index increments. The data of five major stock indices, DAX, DJIA, FTSE-100, HSI and NASDAQ, are applied to test our network model. Unlike other financial forecasting models, our model directly uses the component stocks of the index as inputs for the prediction. For the neural network training, a trust region dogleg path algorithm is applied. For comparison purposes, other neural network training algorithms are also considered, in particular, optimization techniques with line searches are applied for solving the same class of problems. Computational results from five different financial markets show that the trust region based neural network model obtained better results compared with the results obtained by other neural network models. In particular, we show that our model is able to forecast the sign of the index increments with an average success rate above 60% in all the five stock markets. Furthermore, the best prediction result in our application reaches the accuracy rate of 74%.

A description of a teaching support system that has been put into use in CUMT is given. The system is actually an Internet based program that provides a variety of multimedia courses for CUMT's students both in and out of the campus. Besides that, it also provides a friendly environment for better communication between students and teachers. By using it, many teaching and learning tasks like consultations, discussions and seminars, can be fulfilled. In addition, the program has a useful function that allows a large number of experienced teachers in different departments to work in a co-coordinated manner. This working method, where distributed teachers work as collaborators, makes the instant replacement of information possible. With it's remarkable multimedia potential, the system lets students enjoy desired discussions with their preferred teachers as much as they like. The modification plan and the future of the system are also given.

The paper presents a novel multimedia courseware evaluation system. It is based on Internet-supported cooperative work. After we built the logical model based upon the workflow of the evaluation and management system, we developed a cooperative system for the evaluation of multimedia courseware. An expert can work from a distance to use the evaluation system on the Internet; it would reduce an expert's working hours and enhance management's working efficiency to a great extent.