14:00-14:30

副院长宋胜利致辞并为获学院优秀学术论文学生颁奖

14:30-14:42

AP-10K: A Benchmark for Animal Pose Estimation in the Wild

喻航

14:42-14:54

A High-Speed SSVEP-based Continuous Speller Using CWT and Template Reconstruction

熊帮

14:54-15:06

一种基于FPGA的卷积神经网络二值化及其部署优化方法

林成民

15:06-15:18

HiSV: an accurate algorithm for structural vari-ation detection from Hi-C data

李俊萍

15:18-15:30

Charge Prediction by Constitutive Elements Matching of Crimes

赵杰

15:30-15:42

Graph Substructure Assembling Network with Soft Sequence and Context Attention

杨亚明

15:42-15:50

SkipNode: On Alleviating Over-smoothing for Deep Graph Convolutional Networks

陆维港

15:50-15:58

Deep Hierarchical Multimodal Metric Learning

丁阿强

15:58-16:06

Hierarchical Semantic Structure Preserving Hashing for Cross-Modal Retrieval

张彩平

16:06-16:14

Learning Universal Adversarial Perturbations from Local and Global Perspectives

王宇飞

16:14-16:20

Optimizing Task Location Privacy in Mobile Crowdsensing Systems

张文

16:20-16:28

Micro-Expression Recognition Based on MAML Meta-Learning Algorithm

董彩妤

16:28-16:36

A New Load Balance Scheme for Heterogeneous Entities in Cloud Network Convergence

刘家继

16:36-16:44

Modeling the Total Ionizing Dose Effect of CMOS Digital Integrated Circuit Based on the IBIS Model

梁博

16:44-16:52

Towards Machine Learning-Driven Effective Mashups Recommendation from Big Data in Mobile Networks and Internet-of-Things

王智莹

16:52-17:00

JEDERL：A task scheduling optimization algorithm for heterogeneous computing platforms

丁韵青

17:00-17:08

Adversarial Attack On Face Recognition In Physical World Based On Printable Stickers

周冰倩

17:08-17:16

Cross-modal Attention Shift Network for Multimodal Sentiment Analysis

刘帅

17:16-17:24

On Accurate Prediction of Cloud Workloads with Adaptive Pattern Mining

刘文婧

17:24-17:32

政务数据质量评估与提升的研究与实现

周晨羽

导 师：赵伟

报告题目：AP-10K: A Benchmark for Animal Pose Estimation in the Wild

报告摘要：

Accurate animal pose estimation is an essential step towards understanding animal behavior, and can potentially benefit many downstream applications, such as wildlife conservation. Previous works only focus on specific animals while ignoring the diversity of animal species, limiting the generalization ability. In this paper, we propose AP-10K, the first large-scale benchmark for general animal pose estimation, to facilitate the research in animal pose estimation. AP-10K consists of 10,015 images collected and filtered from 23 animal families and 54 species following the taxonomic rank and high-quality keypoint annotations labeled and checked manually. Based on AP-10K, we benchmark representative pose estimation models on the following three tracks: (1) supervised learning for animal pose estimation, (2) cross-domain transfer learning from human pose estimation to animal pose estimation, and (3) intra- and inter-family domain generalization for unseen animals. The experimental results provide sound empirical evidence on the superiority of learning from diverse animals species in terms of both accuracy and generalization ability. It opens new directions for facilitating future research in animal pose estimation. AP-10k is publicly available at https://github.com/AlexTheBad/AP10K.

导 师：万波

报告题目：

A High-Speed SSVEP-based Continuous Speller Using CWT and Template Reconstruction

报告摘要：

Steady-state visual evoked potential (SSVEP)-based brain-computer interface (BCI) spellers have achieved high information transfer rates (ITRs). To further enhance ITR, the spelling method is the key point to optimize. In this study, we propose a continuous spelling system using a character sequence as an input unit. We first extract SSVEP segments from the continuous EEG signals by continuous wavelet transform (CWT) analysis, the

n reconstruct templates for SSVEP segments recognition based on the superposition theory, in which the templates were obtained by the convolution of impulse response and periodic impulse. Both offline and online experiments were conducted with 11 participants by a 12-character speller. The continuous speller achieved the averaged ITRs of 196.25 ± 6.51 bits/min in the online experiment. The results demonstrated the feasibility and efficiency of the continuous spelling method in BCI spellers.

导 师：王泉

报告题目：

一种基于FPGA的卷积神经网络二值化及其部署优化方法

报告摘要：

资源受限的FPGA平台与神经网络的大规模计算存储需求产生矛盾，使得神经网络在FPGA的应用中面临部署难以及执行效率低的问题。针对此问题，本文提出了一种基于 FPGA的卷积神经网络二值化及其推理部署优化方法。首先，本文通过二值化压缩网络，然后调整网络的算子执行次序，使得卷积神经网络的池化输入由激活层调整为卷积层，解决二值化数据经过池化层丢失信息的问题；其次，本文通过重组网络计算模块，将模块的输出由卷积层调整为激活层，降低非二值化数据的搬移带来的开销；最后通过位运算代替乘加运算、流水线并行计算方法提高计算效率。相比于单精度LeNet，二值化LeNet网络在精度仅下降2.47%的情况下，乘法计算量下降了85%，在FPGA上的推理延迟降低了11.1%。实验结果表明，本文提出的方法可以在牺牲少量精度的情况下，有效降低卷积神经网络在FPGA上的计算和存储开销，提高模型推理效率。

导 师：高琳

报告题目：

HiSV: an accurate algorithm for structural vari-ation detection from Hi-C data

报告摘要：

Structural variations (SVs) play an essential role in the evolution of human genomes and are associated with cancer genetics and rare disease. High-throughput chromosome capture (Hi-C) is a new tool that breaks through the limitations of short-read sequencing techniques and can be used to detect SVs and predict their effects on gene expression. However, the identification of SVs from Hi-C data is still in the early stages with few methods and these methods lack the ability to detect all possible types of SV without a control sample. Here, we presented HiSV (Hi-C for structural variation detection), for structural variation detection from a single Hi-C sample. HiSV was compared to existing methods on simulation samples and cancer cell lines and achieved a higher level of sensitivity and accuracy. We further show that HiSV can be a powerful tool for identifying complex SVs on long-read-based validation data sets. Finally, by comparing and integrating the result of Hi-C and whole-genome sequencing (WGS), we demonstrated that Hi-C can supplement WGS to comprehensively detect SVs.

导 师：管子玉

报告题目：

Charge Prediction by Constitutive Elements Matching of Crimes

报告摘要：

Charge prediction is to automatically predict the final judgemental charges for legal cases. Most existing works incorporate legal knowledge either by calculating the correlations between case descriptions and law articles, or manually defining and extracting legal attributes from case descriptions. However, (1) correlation at the article-level fails to distinguish similar charges in the same article; (2) defining legal attributes requires expertise and is hard to be comprehensive. To convict a person or a unit, the involved fact needs to satisfy the constitutive elements (CEs) of the corresponding charge, i.e., there exist matching instances of those CEs in the fact description. This knowledge is a valuable guide for the judge in making the final decisions. However, the knowledge of CEs are far from fully exploited for charge prediction in the literature. This paper proposes a novel method for charge prediction named Constitutive Elements-guided Charge Prediction (CECP). CECP mimics human's charge identification process to extract potential instances of CEs and generate predictions accordingly. It avoids laborious labeling of matching instances of CEs by a novel reinforcement learning module which progressively selects potentially matching sentences for CEs and evaluate their relevance. The final prediction is generated based on the selected sentences and their relevant CEs. Experiments on two real-world datasets show the superiority of CECP over competitive baselines.

导 师：管子玉

报告题目：

Graph Substructure Assembling Network with Soft Sequence and Context Attention

报告摘要：

There has been a surge of researchers' interest in building predictive models over graphs. However, the overwhelming complexity of graph space often makes it challenging to extract interpretable and discriminative structural features for graph classification. In this work, we propose a new graph neural network model called Substructure Assembling Network (SAN) to learn graph representations for classification. The key innovation is a unified Substructure Assembling Unit (SAU), which is a variant of Recurrent Neural Network (RNN) designed to hierarchically assemble useful pieces of graph components so as to fabricate discriminative substructures. A key challenge is that SAUs need to process the neighbors of a node sequentially while no natural order is defined therein. SAN tries to make the model insensitive to neighborhood orders by randomly shuffling neighborhood sequences in training. However, this could suffer high variance, especially when the neighborhood size is large. Hence, we further propose to equip SAN with a novel module named Soft Sequence with Context Attention (SSCA). SAN-SSCA employs the proposed context attention technique to learn the best "soft" permutation of the neighbors w.r.t. classification. It helps the model achieve higher accuracy as well as lower variance. Experiments confirm the effectiveness of SAN-SSCA.

导 师：赵伟

报告题目：

SkipNode: On Alleviating Over-smoothing for Deep Graph Convolutional Networks

报告摘要：

Over-smoothing is a challenging problem degrading the performance of deep graph convolutional networks (GCNs). However, most existing studies for alleviating the over-smoothing lack either generality or efficiency. In this paper, we analyze the three underlying reasons behind the over-smoothing problem, i.e., feature- diversity degeneration, gradient vanishing, and weight over-decaying. To address the above issues, we propose a simple yet effective plugin module, SkipNode, to alleviate over-smoothing. SkipNode randomly (or based on node degree) selects nodes to skip the convolutional operation by backtracking their input features at the end of each layer of GCNs. Analytically, 1) backtracking node information reduces the degeneration of feature diversity, 2) the skipped nodes enable gradients to be directly passed back, thus mitigating the gradient vanishing and weight over-decaying issues. To demonstrate the superiority of SkipNode, we conduct extensive experiments on nine datasets, including homophilic and heterophilic graphs, with different graph sizes on two typical tasks: node classification and link prediction. Specifically, 1) the SkipNode has strong generality of being applied to various GCN-based models on different datasets and tasks. 2) SkipNode has better effectiveness by outperforming the state-of-art anti-over-smoothing strategies, i.e., DropEdge and DropNode.

导 师：田玉敏

报告题目：

Deep Hierarchical Multimodal Metric Learning

报告摘要：

Multimodal metric learning plays an important role in many multimedia applications such as cross-modal matching and cross-modal retrieval. This process aims to transform heterogeneous data into a common subspace, where cross-modal similarity computing can be directly performed and the semantic similarities between samples are completely preserved. Typically, existing methods are designed for non-hierarchical labeled data that treat all categories equally and ignore class-wise relationships. However, in many real-world applications, there are often hierarchical labeled multimodal data such as product data collected from an online shopping platform. Present methods fail to exploit the inter-category correlations in the label hierarchy and therefore cannot achieve optimal performance on hierarchical labeled data. To address this problem, we propose a novel metric learning method for hierarchical labeled multimodal data, named deep hierarchical multimodal metric learning (DHMML). DHMML learns the multi-layer representations for each modality by establishing a layer-specific network corresponding to each layer in the label hierarchy. In particular, a multi-layer classification mechanism is introduced to enable the layer-wise representations to not only preserve the semantic similarities within each layer but also retain the inter-category correlations across different layers. In addition, an adversarial learning mechanism is proposed to bridge the cross-modality gap by producing indistinguishable features for different modalities. Through integration of the multi-layer classification and adversarial learning mechanisms, DHMML can obtain hierarchical discriminative modality-invariant representations for multimodal data.

导 师：王笛

报告题目：

Hierarchical Semantic Structure Preserving Hashing for Cross-Modal Retrieval

报告摘要：

Most of the priors supervised cross-modal hashing methods are flat methods which are designed for non-hierarchical labeled data. They treat different categories independently and ignore the inter-category correlations. In practical applications, many instances are labeled with hierarchical categories. The hierarchical label structure provides rich information among different categories. In this paper, we propose a deep cross-modal hashing method named hierarchical semantic structure preserving hashing (HSSPH), which directly exploits the label hierarchy information to learn discriminative hash codes. Specifically, HSSPH learns a set of class-wise hash codes for each layer. By augmenting class-wise codes with labels, it generates layer-wise prototype codes which reflect the semantic structure of each layer. In order to enhance the discriminative ability of hash codes, HSSPH supervises the hash codes learning with both labels and semantic structures to preserve the hierarchical semantics. Besides, efficient optimization algorithms are developed to directly learn the discrete hash codes for each instance and each class. Extensive experiments on two benchmark datasets show the superiority of HSSPH over several state-of-the-art methods.

导 师：高海昌

报告题目：

Learning Universal Adversarial Perturbations from Local and Global Perspectives

报告摘要：

Deep neural networks (DNN) have been proven to be vulnerable to adversarial attacks. The early attacks mostly involved image-specifific approaches that generated specifific adversarial perturbations for each individual image. More studies have further demonstrated that neural networks can also be fooled by image-agnostic perturbations, called “universal adversarial perturbation”. In this paper, we consider the success rate of adversarial attacks and the quasi-imperceptibility of perturbations and introduce two novel, simple but effificient universal adversarial perturbation generation methods.We first concentrate mainly on the attack success rate of universal adversarial samples and develop an optimization-based generation method (SUAN) to achieve visible local adversarial noises, while taking into account the pixel intensity and the amount of perturbation. The method achieves excellent fooling rates in both targeted and non-targeted attacks, almost all above 95%, and performs well in cross-data and cross-model settings. Then, considering the imperceptibility of the perturbation, we propose an optimization algorithm combined with deep steganography (UAP-DS) on this basis to map the local adversarial noises to the global perturbation. This method achieves equivalent or even surpasses advanced methods in non-targeted attacks and has good transferability. Finally, we verify the attack effect of the universal adversarial perturbations generated by both methods, discuss their limitations and make a trade-off between the two in specifific application scenarios.

导 师：董学文

报告题目：

Optimizing Task Location Privacy in Mobile Crowdsensing Systems

报告摘要：

The location information for tasks may expose sensitive information, which impedes the practical use of mobile crowdsensing (MCS) in the industrial Internet. In this paper, to our knowledge, we are the first to discuss the privacy protection of task locations and propose a codebook-based task allocation mechanism to protect it. Considering the cost of system utility caused by privacy protection technology, the trade-off between local privacy and system utility is formalized a multi-objective optimization problem. The optimal solution is theoretically derived, and the optimal task allocation scheme is obtained. In addition, the selected allocation codebook (SAC) method is introduced to solve the problem of high computational resource consumption in the task allocation process and protect the task location privacy to some extent. The experimental results show that the SAC method sacrifices system utility but improves the privacy protection for task locations by 60% on average.

导 师：万波

报告题目：

Micro-Expression Recognition Based on MAML Meta-Learning Algorithm

报告摘要：

The micro-expression, an accurate and effective behavioral clue, can reveal true changes in people’s emotion. It is widely used in education, interrogation, clinical diagnosis etc. Previous researches on micro-expression recognition focused on traditional image recognition and deep-learning methods, both of which rely on large amounts of training samples and have lower accuracy. Therefore, we introduce a meta-learning training method to achieve high accuracy with small samples. A three-dimensional convolutional neural network (3DCNN) is used as the representation of the learning input of the feature extractor. In the meta-training stage, the optimal parameter is obtained through double-loop iterative updating. We have conducted extensive experiments on two typical micro-expression datasets: CASME II and SMIC-HS. The results show recognition accuracy based on meta-learning can reach 92.97%, which is significantly better than traditional algorithms. In addition, the model can adapt to new micro-expression recognition tasks flexibly and quickly with small samples.

导 师：姜晓鸿

报告题目：

A New Load Balance Scheme forHeterogeneous Entities in Cloud Network Convergence

报告摘要：

For future Internet and next-generation network, the cloud networking convergence is one of the most popular research directions, and it has attracted widespread attention from academia as well as industry. Network adapting cloud and network cloudification are two dimensions in cloud network convergence that can break the closeness and independence between cloud and network. However, the techniques related to the network adapting cloud and network cloudificationunavoidablyintroduce more heterogeneous devices, services and users. That disables the existing load balance schemes which are almost proposed for data centers in cloud computing environments, where the entities are typically standard hardware and software modules. As a result, the overhead and cost of load balance shcemes would be raised significantly in the progress of cloud network convergence. Therefore, in this paper, to make the most usage of heterogeneous entities and encourage the development of future Internet as well as next-generation networking, we proposeamodel andthe requirementsof load balance for heterogeneous entities in the convergence of cloud and network, then we present a concrete load balance scheme. Finally, we discuss the abilities and applications of our proposed model and scheme.

导 师：刘锦辉

报告题目：

Modeling the Total Ionizing Dose Effect of CMOS Digital Integrated Circuit Based on the IBIS Model

报告摘要：

Simulating the total ionizing dose (TID) effect of electrical system by using the transistor-level model is time-consuming and hard to apply to digital integrated circuit (IC), which is composed of millions of transistors.We propose a TID effect modeling method for CMOS digital IC based on the input/output buffer information specification (IBIS) model. The proposed TID model is a behavior model, which consists of input buffer, output buffer and behavior description parts. The physical characteristics of the input and output buffers are described in the VHDL-AMS language by using IBIS model data, and the logical behavior are described in VHDL. The proposed behavioral model can acquire reasonable accuracy and dramatically decrease the simulation time. Finally, we design an experiment to verify the validity of the proposed modeling method, and the results show that the proposed behavioral model can perfectly indicate the performance degradation of CMOS digital integrated circuits caused by TID effects.

导 师：徐悦甡

报告题目：

Towards Machine Learning-Driven Effective Mashups Recommendation from Big Data in Mobile Networks and Internet-of-Things

报告摘要：

As various lightweight APIs are released to the development of applications in mobile networks and Internet-of-Things (IoT) freely and publicly, developers can easily compose APIs together to develop a new service, which is also known as a mashup. The emergence of mashups has greatly reduced the workload of software development and has also encouraged companies and individuals to develop APIs in mobile networks and IoT environment, which leads to the dramatic increase of the number of mashups. Such a tend also brings big data, such as massive text data of mashups and continuously-generated usage data. So how to discover the applicable mashup from big mashups data becomes a challenging problem. In this paper, we propose a mashups service recommendation solution from big data in mobile networks and IoT. The proposed framework is driven by machine learning techniques, including neural embedding, clustering and joint matrix factorization. We employ neural embedding to learn the distributed representations of mashups and propose to use cluster analysis to learn the similarity between mashups. We propose a novel joint matrix factorization (JMF) method to finish the mashup recommendation task, where we design a new objective function and an optimization algorithm.

导 师：杨鹏飞

报告题目：

JEDERL：A task scheduling optimization algorithm for heterogeneous computing platforms

报告摘要：

With the rapid development of GPU, FPGA, and other computing units, the heterogeneous computing platform is widely used in cloud computing, data center, Internet of things, and other fields because of its rich computing resources, flexible architecture, and strong parallel processing capability. Aiming at the task scheduling problem of heterogeneous computing resources and lack of global task information for heterogeneous computing platforms, the task execution model is carried out according to the attributes of tasks and computing resources. Then, we use graph neural networks to encode the scalable state information of tasks and computing resources, and the characteristic of tasks and computing resources are aggregated from three levels, which solves the problem of the uncertain number of tasks and lack of global information. To minimize the average task completion time, we design a task scheduling algorithm based on Deep Deterministic Policy Gradient（DDPG）. Experimental results show that compared with Random scheduling, First in First Out scheduling, Shortest Job First scheduling, Roulette scheduling, and existing reinforcement learning scheduling algorithm, the average task completion time of our algorithm（JEDERL, Job Embedding Device Embedding Reinforcement Learning）is reduced by 27.8%, 12%, 28.6%, 21.9%, and 5.3%, respectively and it stays stable when the number of cluster servers and tasks changes.

导 师：高海昌

报告题目：

Adversarial Attack On Face Recognition In Physical World Based On Printable Stickers

报告摘要：

The face recognition system based on DNN is vulnerable to the attack of adversarial samples. The existing adversarial samples are mainly presented in the form of stickers and glassess. As the actual application is not considered, the disturbance is easily detected by the naked eyes. This article is based on the problem, combined with the actual scene, in the digital domain simulation stickers in the physical domain of real deformation, studied the paste sticker on person face effective area, puts forward the method to generate sticker subregional, generated by gradient descent with semantic information, physical world against printable sticker, with associated with stars or team logo for the sticker, Can be applied to hats, cheeks, combined with the actual scene, stickers are not easily detected. Experiments verify that the generated stickers reduce the similarity with the original face, and can successfully attack ArcFace, the best open source face recognition model. On this basis, explore the external factors that affect the effect of sticker attack in the real world, and conduct experimental tests.

导 师：王笛

报告题目：

Cross-modal Attention Shift Network for Multimodal Sentiment Analysis

报告摘要：

Multimodal sentiment analysis (MSA) leverages multimodal signals including verbal language, facial gestures, and acoustic behaviors to identify sentiments in videos. Language modality typically outperforms nonverbal modalities in MSA. Therefore, strengthening the significance of language in MSA will be a vital way to promote recognition accuracy. Considering that the meaning of language often varies in different nonverbal contexts, combining nonverbal information with text representation is conducive to understand the exact emotion conveyed by an utterance. In this paper, we propose a cross-modal attention shift network (CASNet) model to enhance the text representation by integrating visual and acoustic information into a language model. Specifically, it embeds a cross-modal attention shift (CAS) module that adjusts text representation according to emotional cues implied in unaligned nonverbal data, into a transformer-based pre-trained language model. Moreover, a feature transformation strategy is introduced for acoustic and visual modalities to reduce the distribution differences between initial representations of verbal and nonverbal modalities, thereby facilitating the fusion of different modalities. Extensive experiments on benchmark datasets demonstrate the significant gains of CASNet over state-of-the-art methods.

导 师：鲍亮

报告题目：

On Accurate Prediction of Cloud Workloads with Adaptive Pattern Mining

报告摘要：

Resource provisioning for cloud computing requires adaptive and accurate prediction of cloud workloads. However, existing studies have faced significant challenges in predicting time-varying cloud workloads of diverse trends and patterns, and the lack of accurate prediction often results in resource waste and violation of service-level agreements (SLAs). We propose a bagging-like ensemble framework for cloud workload prediction with Adaptive Pattern Mining (APM). Within this framework, we first design a two-step method with various models to simultaneously capture the ''low frequency'' and ''high frequency'' characteristics of highly variable workloads. For a given workload, we further develop an error-based weights aggregation method to integrate the prediction results from multiple pattern-specific models into a final result to predict a future workload. We conduct experiments to demonstrate the efficacy of APM in workload prediction with various prediction lengths using two real-world workload traces from Google and Alibaba cloud data centers, which are of different types. Extensive experimental results show that APM achieves above 19.62% improvement over several classic and state-of-the-art workload prediction methods for highly variable real-world cloud workloads.

导 师：鱼滨

报告题目：

政务数据质量评估与提升的研究与实现

报告摘要：

政务数据质量评估与提升已经成为当前的研究热点，但传统的政务数据质量评估模型和算法仍存在评估角度单一、准确性不足等问题。本文围绕政务数据质量评估和提升对传统方法的不足进行改进。首先，将改进的层次分析法和修正的熵值系数法进行结合，确定了权重的计算方式；其次，基于该种组合权重进行数据质量的多级模糊综合评估；最后，分别针对重复、缺失以及异常的数据进行相应的去重、填补以及剔除处理，从而提升政务数据的质量。实验结果显示该方法准确地修正了各指标的权重比例，并证明了本文提出的方法相较于传统方式的优越性。