TBSI

Student Posters

Yang Bai, Yong Jiang
Tsinghua-Berkeley Shenzhen Institute, Tsinghua University
Title: Improving Query Efficiency of Black-box Adversarial Attack
Abstract: Deep neural networks (DNNs) have demonstrated excellent performance on various tasks, however they are under the risk of adversarial examples that can be easily generated when the target model is accessible to an attacker (white-box setting). As plenty of machine learning models have been deployed via online services that only provide query outputs from inaccessible models (e.g. Google Cloud Vision API2), black-box adversarial attacks (inaccessible target model) are of critical security concerns in practice rather than white-box ones. However, existing query-based black-box adversarial attacks often require excessive model queries to maintain a high attack success rate. Therefore, in order to improve query efficiency, we explore the distribution of adversarial examples around benign inputs with the help of image structure information characterized by a Neural Process, and propose a Neural Process based black-box adversarial attack (NP-Attack) in this paper. Extensive experiments show that NP-Attack could greatly decrease the query counts under the black-box setting.
download

Yang Bai, Wai Kin Chan, Ercan E. Kuruoğlu
Tsinghua-Berkeley Shenzhen Institute
Title: Complexity reduction neural networks using simulated annealing algorithm
Abstract: Artificial neural networks have proved to be very successful in multiple applications including image classification, voice analysis, and language translation. Deep learning architectures such as AlexNet, VGG16 and GoogleNet have showed outstanding accuracy on image classification task. However, this performance is achieved by millions of parameters. To make networks more light weight, real-time and accessible to individual users, we need to reduce the parameters, i.e. links between nodes which would decrease the accuracy. The goal of our research is to show that significant part of performance loss can be recovered by optimizing partially connected network architectures using Simulated Annealing algorithm.
download

Yicong Li, Calvin Chan, Yang Li, Ercan E. Kuruoğlu
Tsinghua-Berkeley Shenzhen Institute
Title: Multi-stage Modeling of COVID-19 Pandemic
Abstract: In this work, we model the COVID-19 pandemic in a multi-stage manner to evaluate the effectiveness of policies and measures taken by Italy, United Kingdom, Turkey, United States, and hopefully provide some guidance for future policy making process. We first use Unscented Kalman Filter (UKF) to find specific time points at which the transmission rate changes sharply. Then, these time points are utilized as splitting points for multi-stage SIR and SEIR modeling, whose results are evaluated by R2 score. Last, the splitting points are compared with actual time points of policies, which can further reflect the effectiveness of policies.
download

Naiqi Li, Yong Jiang, Shu-Tao Xia, Ercan E. Kuruoğlu
Tsinghua-Berkeley Shenzhen Institute, Tsinghua University
Title: GPGene: Inferring Gene Regulatory Networks with Stochastic Variational Gaussian Processes
Abstract: Understanding how genes regulate each other is a practical question that has profound impact in drug design, personalized medicine, etc. In this paper we propose GPGene, a novel method for inferring gene regulatory networks from high throughput DNA expression data. The effectiveness of GPGene is demonstrated by thorough experimental study. For the static Gene Regulatory Network inference our method outperforms GENIE3. For the dynamic GRN inference we show that our method not only can capture the dynamic change of the network, but also be used to predict future expression levels.
download

Shisong Lin, Mengchao Bai, Feng Liu, Linlin Shen, Yicong Zhou
Shenzhen University, Macau University

Title:

Abstract: We present an efficient orthogonalization-guided feature fusion network, namely OGF2Net, to fuse the features extracted from 2D and 3D faces for facial expression recognition. While 2D texture maps are fed into a 2D feature extraction pipeline (FE2DNet), the 3D attribute maps are input of the 3D feature extraction pipeline (FE3DNet). To reduce the redundancies among features extracted from 2D and 3D streams, we design an orthogonal loss-guided feature fusion network to orthogonalize the features before fusing them. Experimental results show that the proposed method significantly outperforms the state-of-the-art algorithms on both the BU-3DFE and Bosphorus databases.
download

Wenshuang Liu, Wenting Chen, Yuanlue Zhu, Linlin Shen
Shenzhen University, Guangdong Key Laboratory of Intelligent Information Processing
Title: SATGAN: Augmenting Age Biased Dataset for Cross-Age Face Recognition
Abstract: We propose a Stable Age Translation GAN (SATGAN) to generate fake face images at different ages to augment age biased face datasets for Cross-Age Face Recognition (CAFR) . The proposed SATGAN consists of both generator and discriminator. As a part of the generator, a novel Mask Attention Module (MAM) is introduced to make the generator focus on the face area. In addition, the generator employs a Uniform Distribution Discriminator (UDD) to supervise the learning of latent feature map and enforce the uniform distribution. Besides, the discriminator employs a Feature Separation Module (FSM) to disentangle identity information from the age information.
download

Weizeng Lu, Xi Jia, Weicheng Xie, Linlin Shen, Yicong Zhou, Jinming Duan
Shenzhen University, University of Birmingham, University of Macau
Title: Geometry Constrained Weakly Supervised Object Localization
Abstract: We propose a geometry constrained network, termed GC-Net, for weakly supervised object localization. GC-Net consists of three modules: a detector, a generator and a classifier. The detector predicts the object location represented by a set of coefficients, which is constrained by the mask produced by the generator. The classifier takes the resulting masked images as input and performs two complementary classification tasks (object and background). To make the mask more accurate, we propose a novel multi-task loss function that takes into account area of the mask, the categorical cross entropy and the negative entropy. Extensive experiments on the CUB-200-2011 and ILSVRC2012 datasets show that GC-Net outperforms state-of-the-art methods.
download

Shi Mao, Yinheng Zhu, Lu Fang, Ercan E. Kuruoğlu
Tsinghua-Berkeley Shenzhen Institute
Title: Continuous Non-Parametric Potential Estimation in Conditional Random Fields
Abstract: While Conditional Random Fields(CRF) have been widely used for solving “neighboring dependency problem” in Computer Vision, the validation of Gaussian assumption of potential function is usually neglected. In this project, we first introduce a non-parametric continuous representation of potential function which can be optimized in a data-driven manner by gradient decent. We can further investigate the property of potential distribution through learned potential function between various categories.
download

Yinghe Qing, Jifeng Sun, Yinghe Qing, Ercan E. Kuruoğlu
Tsinghua-Berkeley Shenzhen Institute
Title: Bayesian Hidden Markov Models For Unsupervised Classification of Financial Time Series
Abstract: The hidden Markov model is a classical doubly random process, which has been widely used in speech recognition, image analysis and financial timing prediction. We study the application of the Hidden Markov Model with Gaussian Mixture Distribution in the stock market using Bayesian method like Gibbs sampling for parameter estimation, and compare the performance with the classic Hidden Markov Model using Forward backward algorithm and Viterbi algorithm.
download

Ruhui Shen, Yang Li
Tsinghua-Berkeley Shenzhen Institute
Title: Li-ion Battery Remaining Lifetime Prediction
Abstract: Quantitative cycle life prediction using earlier cycles of battery is an interesting and important task in lithium-ion battery systems. Different selections of features including voltage, location, capacity and current extracted from semi-supervised NASA dataset with different regression methods are used to acquire an accurate early-stage fade prediction. Improvement to the baseline method was acquired in the 1A and 4A group.
download

Yang Tan, Yang Li, Shao-Lun Huang
Tsinghua-Berkeley Shenzhen Institute
Title: Cross-domain Cross-task Transferability Estimation
Abstract: Transferability can reveal how easy it is to transfer knowledge learned from one classification task to another by observing the differences of data. Usually, there exists domain difference and task difference between source and target tasks. Previous works did not decompose the total difference or just ignored the domain difference leading to inaccurate estimation and limited application scenarios. To address this problem, we propose a pipeline adopting Wasserstein distance to evaluate domain difference and Conditional Entropy to represent task difference. Basic experiments have tentatively demonstrated the effectiveness of our proposed pipeline.
download

Mingyan Wu, Xi Zhang, Linlin Shen, Hang Yu
Shenzhen University, Guangdong Key Laboratory of Intelligent Information Processing, Shenzhen Laboratory of IC Design for Internet of Things
Title: Pose-aware Multi-feature Fusion Network for Driver Distraction Recognition
Abstract: We propose a novel multi-feature fusion network based on pose estimation, for image based distracted driving detection. Since hand is the most important part of driver to infer the distracted actions, our proposed method firstly detects hands using the human body posture information. In addition to the features extracted from the whole image, our network also include the important information of hand and body posture. The global feature, hand and pose features are finally fused by concatenation of feature maps. The experimental results show that our method achieves state-of-the-art performance on SZ Bus Driver dataset and AUC dataset.
download

Jingyun Yang, Yang Li
Tsinghua-Berkeley Shenzhen Institute
Title: Multi-task learning for PVL detection in MRI images
Abstract: Periventricular white matter softening (PVL) is a common neuro-degenerative disease characterized by white matter lesions near the lateral ventricle, which causes motor control problems and stunting. Studying these small brain lesions requires expensive and time-consuming annotation. In the case of scarce data and small objects, we propose a robust and self-adapting framework on the basis of 3D U-Net to detect the small lesions. The trained model achieves satisfying segmentation and classification performance. To optimize the performance of segmentation, we plan to establish a jointly training multi-task model with the auxiliary classification task to do segmentation and classification simultaneously.
download

Runpeng Yu, Ercan E. Kuruoğlu， Shao-Lun Huang
Tsinghua-Berkeley Shenzhen Institute
Title: Non-Linear Connectivity Analysis in Brain Using Maximal Correlation and Bayesian Network
Abstract: Discovering the functional connectivity in the human brain can help us understand the mechanism of it. Most of the researchers used linear correlations to detect such connectivity. However, the functional connectivity in the brain is more than linearity. In this work, we first extend the original maximal correlation to detect the non-linear connectivity in the brain. And we use the partial correlation to construct the Bayesian network to visualize it. The results show that our non-linear correlation can discover the functional connectivity that the linear correlation cannot detect.
download

Zihao Zhou, Aihua Ran, Shuxiao Chen, Xuan Zhang, Guodan Wei
Tsinghua-Berkeley Shenzhen Institute, Tsinghua Shenzhen International Graduate School
Title: Data-driven Fast Clustering of Second-life Lithium-ion Battery: Mechanism and Algorithm
Abstract: To achieve fast retired battery screening, a pulse clustering model embedded with improved bisecting K-means algorithm has been developed to effectively sort out retired batteries with life cycles ranging from new to the end-of-life states. The relevance of selected variables is rigorously validated, reaching the accuracy as high as 88% compared with the traditional full charge-discharge test. To note, the test time has largely reduced from hours to minutes.
download

Time Jul, 2020	Presentation
2:30 pm – 2:45 pm	Mingyang Wu Shenzhen University Title: Pose-aware Multi-feature Fusion Network for Driver Distraction Recognition.
2:40 pm – 3:00 pm	Weizeng Lu Shenzhen University Title: Geometry Constrained Weakly Supervised Object Localization
3:00 pm – 3:15 pm	Naiqi Li Tsinghua-Berkeley Shenzhen Institute Title: GPGene: Inferring Gene Regulatory Networks with Stochastic Variational Gaussian Processes
3:15 pm – 3:30 pm	Shi Mao Tsinghua-Berkeley Shenzhen Institute Title: Continuous Non-Parametric Potential Estimation in Conditional Random Fields
3:30 pm – 3:45 pm	Yang Tan Tsinghua-Berkeley Shenzhen Institute Title: Cross-domain Cross-task Transferability Estimation
3:45 pm – 4:00 pm	Yang Bai Tsinghua-Berkeley Shenzhen Institute Title: Improving Query Efficiency of Black-box Adversarial Attack

Student Posters

Oral Presentation Program

Mingyang Wu

Weizeng Lu

Naiqi Li

Shi Mao

Yang Tan

Yang Bai