Jingzhou Shen Personal Page – Vertical alive, horizontal nostalgia

Present Time-flow Prediction with ViT Variants

2022-06-22T00:00:00+00:00

Researcher, Supervised by Professor Robert Pless

•Used Resnet/ViT to predict the time for one steady camera and for the trained model;

•Visualized the attention of the model by using GradCAM and heatmap;

•Predicted the time information in the picture under a given camera with the improved ViT;

•Extracted q and k for each batch, multiplied them together and showed them in the heatmap to see how the time changes;

•Tracked how the model detected time through the ViT and visualized temporal information within the probe model;

•Made oral representation at AIPR(IEEE workshop).

here is one of our visualization sites

Present Robust Deep Graph Learning for Dynamic Graphs

2022-05-30T00:00:00+00:00

Researcher, Supervised by Professor Luosheng Dong

•Extracted the features from datasets of the brain network, DBLP, Epinion, Reddit , mdke adjacency matrix for each time step, andpreprocessed the dynamic Graph;

•Used GNNs for dynamic Graph and Graph Structure Learning to get a better and clean graph;

•Improved the performance of dynamic Graph Neural Network, its denoising part with AdaNN and the denoise method in PTDNet.

Present Multimodal Graph NN in Brain Network

2022-05-03T00:00:00+00:00

Researcher, Supervised by Professor Hongchang Gao

•Used multimodal Graph NN on the different representations of brain network based on the two brain networks datasets:DTInetwork and fMRI network;

•Combined the learned features of both sides, and used MLP to extract the features together;

•Designed a unique loss function to gather information from the above two brain networks, and iterated each feature to get a new model.

Self-Supervised Learning Movement and Its Physics Law

2022-01-28T00:00:00+00:00

Researcher, Supervised by Professor Rui Liu

Content:

•Simulated the scenes of robots with MuJuCo;

•Analyzed what parameters are needed in the model and what to be expected from the model under the condition of the fourscenarios of robots grasping the target,

•Utilized self-supervised datasets to learn the physics law through a robot (based on PINN and SymbolicGPT);

•Implentmented PINN + SymbolicGPT into a robot to evaluate its movement mode and the physics law behind it.

MOSI-Hate speech Project

2021-08-24T00:00:00+00:00

Supervised by Professor Robert Pless

Content:

•Crawled the information from social media like Reddit, Twitter, and other platforms;

•Defined several types of the images like sexuality, violence and etc.,

•Conducted image processing and effectively analyzed images with obvious negative emotions from social networks;

•Classified images by Vision Transformer and detected the emotion of the words in the images by Bert;

•Put the above information into ALBEF and CLIP to train the mode,l, and predicted the emotion of a new image.

Image Processing of Fundus Lesions Based on Deep Learning

2020-01-28T00:00:00+00:00

Researcher, Supervised by Professor Jialin Peng

Content:

• Used public dataset diaretdb as the main data source;

• Recognized four types of diseases in the fundus image, among which each one has some particular areas/features in the image；

• Analyzed the fundus image pathology by using U-Net, U2Net first to test the performance；

• Created U2-Net + PSPNET to test the performance in the datasets；

• Constructed a neural network that could analyze the pathological images of the fundus.