Xiong Yang

3Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Xiong Yang focuses on computer vision, hierarchical image parsing, and genetic risk stratification. Early research explored bottom-up and top-down inference processes for object recognition and robust texture analysis. More recent thematic evidence indicates a shift toward multimodal integration and polygenic risk modeling in clinical genetics.

Research Trajectory

During the 2009–2010 period, Yang's research primarily addressed foundational problems in computer vision, specifically focusing on object recognition, image parsing, and texture analysis. A central theme of this early work was the integration of different computational processes to improve visual inference.

In 2009, Yang led research on scheduling computational processes in hierarchical vision models. As first author of Evaluating information contributions of bottom-up and top-down processes, Yang introduced a unified Bayesian framework to quantitatively measure the distinct contributions of bottom-up and top-down computing channels. By isolating these channels—termed alpha, beta, and gamma—using a blocking control strategy, the work provided a systematic method for evaluating how different information flows contribute to object recognition tasks.

Concurrently, Yang contributed as second author to broader efforts in stochastic image grammar. The paper Image parsing with stochastic grammar: The Lotus Hill dataset and inference scheme presented a large-scale ground truth dataset developed by the Lotus Hill Institute. This work addressed the critical need for diverse, manually annotated training data in the computer vision community, while also proposing a top-down/bottom-up scheme for scheduling inference processes within stochastic image grammars.

In 2010, Yang's focus included robust feature extraction for texture modeling. As second author of A new texture descriptor using multifractal analysis in multi-orientation wavelet pyramid, Yang helped develop a novel texture descriptor that implicitly combined spatial and frequency domain information. This approach utilized a multi-oriented wavelet leader pyramid to encode multi-scale information of texture edgels. By applying multifractal analysis to capture the empirical power-law relationships found in natural textures, the resulting descriptor demonstrated robustness to scale and rotation changes.

Current Directions

While the foundational publications in the evidence focus on computer vision, the most recent thematic data from 2024 indicates a distinct focus on bioinformatics, clinical genetics, and predictive health modeling. The most recent evidence points toward the research direction of Cross-Modality Molecular Imputation for Disease Association, which involves using genetic data to predict intermediate biological layers and understand the functional mechanisms of non-coding genetic variants.

Additionally, Yang's recent profile encompasses the broader area of Genetic Risk Stratification for Early-Onset Disease, addressing the diagnostic gap for individuals with high biological predispositions for future illnesses. This aligns with themes such as Predicting Complex Phenotypes via Polygenic Risk Aggregation, which shifts focus from monogenic conditions to the cumulative influence of thousands of genetic variants. Finally, the evidence highlights work in the area of Multimodal Integration for Clinical Risk Stratification, emphasizing the combination of genetic, biochemical, and clinical data to construct comprehensive, dynamic models of human health.

Profile updated

See something wrong? Report a mistake