News

STEM Course: 2021 CVML Summary

June 21, 2021

The Computer-Vision and Machine-Learning (CVML) course, taught by Shanghai Jiao Tong University graduate students to the high-school students of SHSID, centers around the theories, implementation, and applications of machine learning. From covering the mathematical basis and implementation of deep learning algorithms to the status quo of artificial intelligence, students learned how to use machine learning to carry out research for different areas.


252A7

 

292F6


21088

 

1EC52

 

At the end of the first semester, students were tasked with designing a program to transfer artistic style. The program was to receive input of two images, a piece of artwork and a photograph, and produce a machine-generated piece of art that contained the stylistic features of the original artwork and the contents of the photograph. Students worked in groups to implement the program, create a user interface, and then collect samples to feed into the program. The students first designed a user interface using PyQt, which consisted of a substantial set of graphic user-interface widgets that students could utilize. Then, students implemented the algorithms for extracting the style features from artworks and content features from photographs taught in class. Finally, to test out the program, students collected input data including style images and content images, and produced their own machine-generated artworks.

          

   Original

61EE

                                                      

Artwork

5BEBC

                                    

    Stylished Result

1E7DF


One group managed to complete the project, and two others were able to do so with further aid from the teachers despite various technical difficulties. Although completing this project was challenging, students were passionate to see how artificial intelligence can show its charm in the field of art.


On June 11th, three groups of students presented their chosen artificial-intelligence project.


The first group presented their 2D gaze-following-detection program, which predicts where a person in an image is looking. In the first stage, the gaze direction is predicted, which generates multi-scale gaze-direction fields. These direction fields are then concatenated with the image contents in the second stage and fed into a heatmap pathway for heatmap regression.


74D1

 

1CF2E

 

 

17152

 

89DD

 

The next group demonstrated their phishing-website-detection project, which uses the random-forest algorithm. The detection system analyzes URLs and outputs a number in the range of 1 to -1, 1 being a safe website while -1 being a phishing website.


Last but not least, the final group introduced their 3D human-pose estimation, which produces accurate natural-motion sequences by using a video pose and shape-estimation method known as Video Inference for Body Pose and Shape Estimation (VIBE). In summary, this technique exploits a large-scale motion-capture dataset to train a motion discriminator using an adversarial approach.


1345C


 

 

 

 

These projects sparked students' further interest in artificial intelligence as they realized its broad applications. This project also gave them more experience and intimacy with the process of developing machine-learning projects, from the understanding and implementation of theory to more practical aspects, such as designing the user interface, as well as how to collaborate as a team to work efficiently.


Moreover, the students also undertook a thorough exploration of the theory behind machine learning. They began with learning the mathematical background knowledge, such as logic, linear algebra, and optimization, and then moved on to algorithms within different categories of machine learning. They investigated how supervised learning algorithms can be used to classify or predict new data based on known data, how unsupervised learning algorithms can yield information of unseen data without being supplied with known data, and how reinforcement learning algorithms can adapt to responses from the environment as they perform certain actions selected based on calculation. They then dived into deep learning, exploring the various types, architectures, and techniques of neural networks as well as its applications.


After this course, students not only gained a solid theoretical foundation in artificial intelligence, but also applied their knowledge in action by implementing various projects to realize its wide applications. With this experience, the students are confident to continue on their journey in the rich field of artificial intelligence.


(Written by Jasper Hu, Richard Chu    Pictures and Videos by Jasper Hu, Anya Yan, Yeji Ju, William Xu   Supervised by Tianzhou He   Reviewed by Qian Zuo)