This issue is the 355th online talk of the techbeat artificial intelligence community.
At 8:00 p.m. Beijing time on November 4 (Thursday), Zhang Huan, a postdoctoral researcher at Carnegie Mellon University, will start broadcasting in the techbeat artificial intelligence community on time!
The theme he shared with you is: "Building More Powerful and Larger Neural Network Verification Algorithms", which will be
It mainly introduces neural network verification algorithms, which are used to strictly prove the properties of neural networks such as security, robustness, fairness, and so on.
Talk information
▼
Theme: Build stronger, larger scale
Neural network validation algorithm
Guest: Huan Zhang, postdoctoral fellow at Carnegie Mellon University
Time: 20:00 Beijing time on November 4 (Thursday).
Where: Techbeat Artificial Intelligence Community
http://www.techbeat.net/
Long press to identify the QR code, one click to complete the appointment!
Talk Outline
Deep neural networks have become an important part of modern artificial intelligence and have been successful in many applications. However, when using deep neural networks in some systems that require high safety and robustness, such as autonomous driving, medical systems, and safety monitoring systems, we need neural networks with predictable, reliable, and robust performance, which is a very challenging problem. For example, some existing studies have found that neural networks in many applications lack robustness and have adversarial samples that can make neural networks error in some very simple tasks.
This talk mainly introduces the neural network verification algorithm, which is used to strictly prove the properties of neural networks such as security, robustness, fairness and so on. Since the neural network verification algorithm can give theoretical guarantees for the performance of neural networks, it has become the key to applying neural network models in some key areas.
1. Motivation and description of the underlying problems of neural network validation
2. Neural network validation algorithm crown based on efficient backward bound propagation
3. Generalize the crown to any neural network structure and auto_lirpa validation algorithm library
4. The alpha-crown algorithm uses gradient rise to optimize limit propagation on the GPU
5. Use the branch demarcation method to further enhance the verification effect, and use the beta-crown algorithm that combines articulation propagation and branch delimitation to win the neural network verification competition
Talk references
This is the information that will be mentioned in this sharing, it is recommended to preview it in advance!
github code link:
https://github.com/huanzhang12/alpha-beta-crown
paper list:
neurips 2021 (if you only have time to read a paper, look at this one):
https://arxiv.org/pdf/2103.06624.pdf
neurips 2018:
https://arxiv.org/pdf/1811.00866.pdf
iclr 2021:
https://arxiv.org/pdf/2011.13824.pdf
Getting Started Article: There were two previous articles posted on the Heart of the Machine:
September 11th: The Chinese student team won the international neural network verification competition: the overall score was the first, and the five major items were the first
September 16: Dry Goods |What is Neural Network Validation? An article to read the neural network verification competition winning algorithm α, β-crown
Talk and ask questions
Ask questions in the following two ways to get WeChat cash red envelope rewards!
Ways (1)
Participate in the interaction in the [Communication Area] under the talk interface! Leave your call🤟 and questions 🙋, discuss with more friends, and be answered by the speaker directly flop!
Every time you contribute, we will give you the corresponding i-bean points, and there will be surprise rewards!
Ways (2)
Ask your question directly in the message area of this article, or scan the QR code below to ask a question!
Come and ask questions!
Talk and guest introduction
Zhang Huan
Postdoctoral Fellow at Carnegie Mellon University
Dr. Huan Zhang graduated from UCLA in 2020 and is currently a postdoctoral fellow at Carnegie Mellon University (CMU). Zhang Huan is one of the early researchers in the field of machine learning robustness and security, proposing pioneering verification algorithms for machine learning models such as neural networks and decision trees, and applying these algorithms to tasks such as building more secure and robust image classification, natural language processing (nlp), reinforcement learning (RL), and publishing dozens of papers in first-class conferences such as neurips, icml, and iclr. The tools developed by the team led by Zhang Huan α, β-crown won the first place in the overall score in the 2021 International Neural Network Verification Competition (vnn-comp) and the first place in 5 individual items. In addition, Zhang Huan received the IBM PhD Scholarship in 2018-2020 and the Advml Rising Star Award in 2021 by the MIT-IBM Watson AI Lab.
Personal Homepage:
https://www.huan-zhang.com/
If you also want to be a speaker
Self-nomination / referral
Single-person talk | Team session | Recording or live | Communicate behind closed doors
Choose from a variety of options!
There are also rewards for the success of recommended speakers
About the techbeat AI community
Techbeat (www.techbeat.net) is affiliated with Shomen Venture Capital, a growing community that gathers the global Chinese AI elite.
We hope to create more professional services and experiences for AI talents, accelerate and accompany their learning and growth.
Look forward to this place can become a highland for you to learn the cutting-edge knowledge of AI, share the fertile ground of your latest work, and upgrade and fight monsters on the road to AI advancement!
More details>> techbeat, a learning and growth community that gathers the global Chinese AI elite