Less than the scene, still see the driest academic report!
Hi, everybody. Here is the column of academic report, the core reading small edition is selected from time to time and personally run to dedicate the best academic report in the field of science and technology for everyone, record the report for the students, and try every means to get a hand of PPT and live video – enough dry, enough fresh! Without more words, I hope that these excellent young scholars and experts jieqing’s academic reports can make your knowledge reading in your spare time more valuable.
Artificial intelligence forum is now a vast sea, there are hard goods, there are dry goods of the lecture but one in a hundred. Since the launch of AI Future · Youth Academic Forum on January 19, 2019, the forum has been held for 12 consecutive sessions, attracting tens of thousands of applicants from over 30 provinces across China, 13 countries at home and abroad, and more than 400 universities and research institutes. The 12th AI Future Theory · Youth Academic Forum (special session for Baidu Scholarship Doctoral students) was held in K6 Lecture Hall, Baidu Science Park, Beijing on The afternoon of January 5, 2020. 20. Robustness of Deep Learning Adversarial Robustness of Deep Learning is brought to you by Yinpeng Dong of Tsinghua University.
AI Future talk · Youth Academic Forum _ Tencent video
Dong Yinpeng is a doctoral student in the 3rd year of the Department of Computer Science, Tsinghua University. His supervisor is Professor Zhu Jun. His research interest covers machine learning and computer vision, focusing on the robustness of deep learning in adversarial environments.
Content: Aiming at the problem that existing deep learning models are easily fooled by the adversarial samples of attackers, Dr. Dong’s three research results on the robustness of deep learning in adversarial environment.
Adversarial Robustness of Deep Learning
Deep learning has made a lot of progress in the past two years, and relevant models have been applied to various systems. But at the same time, the reliability of deep learning model is also subjected to many tests. Various findings show that deep learning models are easily fooled by an attacker’s adversarial sample, in which the attacker adds tiny samples to the original sample, which causes the model to misclassify the item. Some samples look the same to the human eye, but the deep learning model makes wrong judgments, which can lead to some very real security risks. There are also examples of confrontation in the actual system. For example, adding some noise to traffic data will make the prediction of automatic driving system wrong.
Adversarial samples can be summed up as optimization problems. In order to solve such optimization problems, there are many methods to find adversarial samples or directly optimize adversarial samples. Many methods need to obtain network gradients, that is, network parameter information, which is called white box attacks, while methods that do not need network gradients are called black box attacks. Based on the migration performance of adversarial samples, that is, adversarial samples for one model can also deceive other models, which can produce adversarial samples. On the other hand, adversarial samples can be found by estimating model gradient or by random search.
Dr. Dong’s first work was the momentum iterative style book generation method. There is no trade-off between migration performance against samples and white box attack capability. Referring to the momentum algorithm in the optimization field, the momentum superposition process is recorded and used in the generation process of countermeasures samples, which not only improves the migration performance of countermeasures samples, improves the attack ability against black box models, but also can not be too sensitive to white box models.
There are several ways to improve the model’s defenses. The second work of Dr. Dong is to better attack the defenseless black box model by using image transformation and frequency domain transformation. Compared with other algorithms, the attack efficiency remains unchanged while reducing the sensitivity to the current model.
The third work is to combine the unknown network gradient attack method and network gradient estimation method to improve the black box attack more effectively.
AI Future Say * Youth Academic Forum
The first data mining session
1. Academician Guojie Li: Rational Understanding of the “head goose” role of ARTIFICIAL intelligence
2. Professor Xiong Hui of Baidu: Intelligent Talent Management with big data
3. Professor Tang Jie, Tsinghua University: Theory and Application of Network representation learning
Dr. Liu Qiang: Personalized recommendation in the era of deep learning
5. Dr. Chai Chengliang, Tsinghua University: Data Management based on human-machine collaboration
Natural Language Processing ii special session
1. Zhang Jiajun, Chinese Academy of Sciences: A Synchronous bidirectional inference Model for Natural language generation
2. Li Lei, BUPT: Analysis and Discussion on automatic text summarization
3. Baidu Sun Ke: Industrial application and Discussion of dialogue technology
4. Ali, Tan jiwei: Text abstracts based on sequence to sequence model and Taobao practice
5. Liu Yijia, Harbin Institute of Technology: See the following related word vectors through syntactic analysis
Special session of computer Vision iii
1. Peng Yuxin, Peking University: Analysis and Application of transmedia intelligence
2. Jiwen Lu, Tsinghua University: Deep reinforcement learning and Visual content Understanding
3. Li Yingchao, Baidu: Baidu augmented reality technology and application
4. Zhang Shifeng, Chinese Academy of Sciences, Comparative Exploration of Universal Object detection Algorithms based on deep learning
5. Hong Kong Chinese Li Hongyang: Latest progress in object detection
Special session on voice technology
1. Tao Jianhua, Chinese Academy of Sciences: Current Situation and Future of speech Technology
2. Ji Wu, Tsinghua University: Deep learning Processing of audio Signals
3. Mi Wang Yujun: Mi voice technology behind Xiao Ai
Kang Yongguo, Baidu: Baidu voice technology in the ERA of AI
5. Bin Liu, Chinese Academy of Sciences: Robust end-to-end speech Recognition based on Joint Antagonism Enhancement Training
Quantum Computing episode v
Zhai Hui, Tsinghua University: Quantum Mechanics with Machine Learning
2. Dawei Lu, Southern University of Science and Technology: The Collision between quantum Computing and artificial Intelligence
Li Yinan, National Center for Mathematics and Computer Science (CWI), Netherlands: Quantum computing in the Era of Big Data
Yuxiang Yang, ETH: Quantum precision Measurement
Duan Runyao, Baidu: Quantum Architecture — Opportunities and Challenges
Machine learning session 6
1. Zhang Wensheng, Chinese Academy of Sciences: Cognitive computing in the era of health and medical big data
2. Zhuang Fuzhen, Chinese Academy of Sciences: Research and Application of Machine Learning Algorithm based on knowledge Sharing
Baidu Hu Xiaoguang: Core Technology and Application practice of PaddlePaddle
4. Tsinghua University Wang Yisen: Adversarial Machine Learning: Attack and Defence
5. Shen-yi Zhao, Nanjing University: Scope-Scalable Composite Optimization for Learning
The seventh autonomous Driving special
1. Hongbin Zha, Peking University: SLAM Technology based on Data stream processing
2. Deng Zhidong, Tsinghua University: “Sense” and “knowledge” of autonomous driving – Challenges and opportunities
3. Baidu’s Zhu Fan: Autonomous Driving in the Open Era — Baidu Apollo Project
4. Beili Song Wenjie: Autonomous navigation technology of intelligent vehicle in unknown area under time and airspace
Special session of deep Learning
1. Chinese Academy of Sciences Wen Xin: Introduction to deep learning and learning resources
2. Zhineng Chen, Chinese Academy of Sciences, Computer Vision Classics: Deep learning and object Detection
3. Peng Fu, Chinese Academy of Sciences: Deep learning and machine reading
The ninth issue of personalized content recommendation special performance
1. Xin Zhao, Renmin University: Research on Serialization recommendation Technology Based on Knowledge and Inference
2. Zhao Jun, Chinese Academy of Sciences: Key Technology of Knowledge Graph and its Application in recommendation System
The tenth video understanding and recommendation special session
1. Xiaoru Yuan, Peking University: Intelligent data visual analysis
The eleventh session of Information Retrieval and Knowledge Graph
1. Jun Xu, Renmin University of China: Sorting in intelligent Search — Breakthrough probability sorting rule
2. Yingxia Shao, Beijing University of Posts and Telecommunications: Efficient embedding method of knowledge graph
3. Baidu Song Xunchao: Baidu large-scale knowledge graph construction and intelligent application
4. Zhou Jingbo, Baidu: Construction and application of POI knowledge graph
5. Baidu Zhifan Feng: Knowledge Graph based multi-mode cognitive Technology and intelligent Application
Leave a comment like follow
We share the dry goods of AI learning and development. Welcome to pay attention to the “core reading technology” of AI vertical we-media on the whole platform.
(Add wechat: DXSXBB, join readers’ circle and discuss the freshest artificial intelligence technology.)