Hire Rujie Z. - AI Researcher

I proposed a novel low-complexity framework, 'Selection,' for instance segmentation. Unlike traditional methods such as Faster R-CNN, YOLO, HTC, and DETR, Selection utilizes edge detection to directly segment objects, eliminating the need for anchor generation, reducing hyperparameter tuning, and avoiding duplicate bounding box regression, among other improvements. I designed a parallel clustering algorithm, 'Pyramid,' embedded within the Selection framework. It eliminates the need to compute pairwise distances between data points and integrates concepts of centroids, density, and hierarchy, enhancing its ability to manage small and overlapping objects. Pyramid outperforms K-means, GMM, DBSCAN, Agglomerative, OPTICS, and HDBSCAN in both Average Precision (AP) and time efficiency. The implementation was developed on the Detectron2 platform using PyTorch and C++ CUDA extensions, and it was trained on 2 NVIDIA Tesla V100s and 8 Titan XPs using 5,000 images from the COCO dataset. I trained and evaluated the Selection framework with the Pyramid clustering algorithm, achieving a 72.8% AP, which is 1.3% higher than Faster R-CNN. Additionally, Selection utilizes 36.2 million parameters, 20% fewer than Faster R-CNN. You can find my code sample here: https://github.com/RujieZhao/Selection.

I developed an online testing algorithm to evaluate the potential of different memristors for storing neural network weights. This algorithm effectively tackles the challenge of sparse, discrete conductance values during backpropagation by implementing a dual-weight updating scheme. Verified the efficacy of the online testing algorithm by conducting validations with the MNIST dataset, subsequently providing comprehensive prediction evaluations to fellow researchers specializing in memristor technology.